[
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.",
        "abstract": "## BACKGROUND\nDeep convolutional neural networks (DCNNs) can classify skin diseases at a level equivalent to a dermatologist, but their performance in specific areas requires further research.\n## OBJECTIVE\nTo evaluate the performance of a trained DCNN-based algorithm in classifying benign and malignant lip diseases.\n## METHODS\nA training set of 1629 images (743 malignant, 886 benign) was used with Inception-Resnet-V2. Performance was evaluated using another set of 344 images and 281 images from other hospitals. Classifications by 44 participants (six board-certified dermatologists, 12 dermatology residents, nine medical doctors not specialized in dermatology and 17 medical students) were used for comparison.\n## RESULTS\nThe outcomes based on the area under curve, sensitivity and specificity were 0\u00b7827 [95% confidence interval (CI) 0\u00b7782-0\u00b7873], 0\u00b7755 (95% CI 0\u00b7673-0\u00b7827) and 0\u00b7803 (95% CI 0\u00b7752-0\u00b7855), respectively, for the set of 344 images; and 0\u00b7774 (95% CI 0\u00b7699-0\u00b7849), 0\u00b7702 (95% CI 0\u00b7579-0\u00b7808) and 0\u00b7759 (95% CI 0\u00b7701-0\u00b7813), respectively, for the set of 281 images. The DCNN was equivalent to the dermatologists and superior to the nondermatologists in classifying malignancy. After referencing the DCNN result, the mean \u00b1 SD Youden index increased significantly for nondermatologists, from 0\u00b7201 \u00b1 0\u00b7156 to 0\u00b7322 \u00b1 0\u00b7141 (P < 0\u00b7001).\n## CONCLUSIONS\nDCNNs can classify lip diseases at a level similar to dermatologists. This will help unskilled physicians discriminate between benign and malignant lip diseases. What's already known about this topic? Deep convolutional neural networks (DCNNs) can classify malignant and benign skin diseases at a level equivalent to dermatologists. The lips are a unique feature in terms of histology and morphology. Previous studies of DCNNs have not investigated tumours on specific locations. What does this study add? This study shows that DCNNs can distinguish rare malignant and benign lip disorders at the same rate as dermatologists. DCNNs can help nondermatologists to distinguish malignant lip diseases. What are the clinical implications of this work? DCNNs can distinguish malignant and benign skin diseases even at specific locations such as the lips, as well as board-certified dermatologists. Malignant lip diseases are rare and difficult for less trained doctors to differentiate them from benign lesions. This study shows that in dermatology, DCNN can help improve decision-making processes for rare skin diseases in specific areas of the body.\n",
        "methods": "**METHODS**\n\n_Datasets_\n\nThe dataset in this study was taken from the photo database of Seoul National University Hospital (SNUH). The dataset contains clinical photos of dermatologic patients that were taken since October 2004. Photos related to lip diseases were selected in order to build lip datasets. The diagnoses were annotated based on the clinical and histologic findings. In addition to the SNUH photo database, clinical images of lip diseases were collected from two other affiliated hospitals, Seoul National University Bundang Hospital and SMG-SNU Boramae Medical Center. The study was approved by the Institutional Review Board of SNUH (IRB No. 1811-111-987).\n\n_Classification of lip diseases_\n\nThe lip disease diagnoses were classified either malignant or benign. Malignant lip diseases included skin cancers (SCC, BCC, MM, or adenocarcinoma) and precancerous lesions (actinic cheilitis, Bowen\u2019s disease, and melanoma in situ). Benign lip diseases included lichen planus, melanotic macule, pyogenic granuloma, plasma cell cheilitis, and venous lake.\n\n_Data preparation_\n\nBlurry or images that were inadequate for clinical diagnosis were removed from the datasets. In cases where the image contained the whole face, it was cropped to show only the lip or skin lesions. In some cases, several photos were taken of same patient from multiple viewpoints. Thus, the photos were numbered based on the patient to prevent photos of the same patient from being split into both the training and testing sets.\n\nA total of 2254 images were used for this study (Table 1). This included 1973 images of 404 patients who were diagnosed with lip disorders at SNUH between October 2004 and October 2018. In total, 34 classes of lip disorders including 4 lip cancers, 3 precancerous lip disorders, and 27 benign lip diseases were included for evaluation. The details and number of photos for each diagnosis are listed in Table 1. The images of lip diseases were classified as malignant (853 images) or benign (1120 images).\n\nIn total, 1629 of the SNUH images (743 malignant and 886 benign) were selected for the training set. The remaining 344 SNUH images (110 malignant and 234 benign) were used as the testing set along with 281 images (57 malignant and 224 benign) from Seoul National University Bundang Hospital (225 images) and SMG-SNU Boramae Medical Center (56 images).\n\n_DCNN algorithm_\n\nVarious DCNN structures show good results for image classification. The Inception-Resnet V2 (Google Inc., Mountain View, CA) structure, which recorded a Top-5 error rate of 3.1% for the ILSVRC 2012 test set, was selected for this study. The structure of Inception Resnet-V2 uses factorization and residual connections. Factorization can reduce the computational power and dimensionality; hence, it can prevent overfitting, which is often problematic. Residual connections were proposed by Microsoft ResNet and they can prevent instability caused by the deeper network, which can suffer from vanishing gradients during training. The hyper-parameters were set as follows: weight decay=0.00004, batch norm decay=0.9997, batch norm epsilon=0.001, dropout keep probability=0.8, and activation function ReLu with Adam-optimizer.\n\n_Comparing the performance of the trained DCNN to human participants_\n\nWe compared the performance of 44 medical doctors and medical students (including 6 board-certified dermatologists, 12 dermatology residents, 9 specialists who did not major in dermatology, and 17 medical students) to the trained DCNN. First, we randomly selected 40 images, 20 malignant and 20 benign. These images were viewed by the participants who were then asked: \u201cDo you think that this lip lesion is malignant or benign?\u201d Once the first test was finished, the participants received the output values from the trained DCNN. They were then asked the question again. The DCNN results indicated the degree of malignancy from 0 to 100.00%.\n\n_Statistical analysis_\n\nThe performance of the trained DCNN was evaluated using the testing dataset of 344 images from SNUH and the 281 additional images. A receiver operating characteristic (ROC) curve was drawn in order to compare the performance of the DCNN and the participants. The area under the curve (AUC) was taken as a primary outcome while sensitivity and specificity were also calculated. Delong\u2019s test was used to compare the ROC curves generated from the two image sets. Youden\u2019s index (sensitivity + specificity \u2013 1) was calculated before and after the DCNN results were reviewed and the values were analysed using the Wilcoxon signed rank test. Statistical analysis was performed using R statistical software version 3.51 (R Foundation for Statistical Computing, Vienna, Austria).",
        "results": "**RESULTS**\n\n_Outcomes of trained DCNN on testing set_\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image testing set were 0.827 (95% confidence interval (CI): 0.782\u20130.873), 0.755 (95% CI: 0.673\u20130.827), and 0.803 (95% CI: 0.752\u20130.855), respectively. For the additional 281 images these values were 0.774 (95% CI: 0.699\u20130.849), 0.702 (95% CI: 0.579\u20130.808), and 0.759 (95% CI: 0.701\u20130.813), respectively (Figure 1). There was no significant difference between the two ROC curves (p=0.229). The combined performance outcome for the 625 images in the two test tests, the AUC, sensitivity, and specificity were 0.811 (95% CI: 0.773\u20130.850), 0.737 (95% CI: 0.671\u20130.802), and 0.779 (95% CI: 0.740\u20130.817), respectively.\n\n_Outcomes for the trained DCNN compared to human participants_\n\nThe trained DCNN outperformed most of the participants, including some board-certified dermatologists, in the ROC curve (Figure 2A).\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 40 images chosen from the testing set were 0.792 (95% CI: 0.651\u20130.934), 0.850 (95% CI: 0.700\u20131.000), and 0.650 (95% CI: 0.450\u20130.850), respectively.\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants are shown in Table 2.\n\n_Outcomes of human participants changed after referencing the output value of DCNN_\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants after they referenced the DCNN output are shown in Table 2. The Youden\u2019s index of the less skilled (non-dermatologist) group increased significantly once they referenced the DCNN output (from 0.201 \u00b1 0.156 to 0.322 \u00b1 0.141, _p < 0.001), as shown in Figure 2B. However, there was no significant difference in the dermatologist group (from 0.375 \u00b1 0.167 to 0.444 \u00b1 0.108, p = 0.098).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.",
        "abstract": "## BACKGROUND\nDeep convolutional neural networks (DCNNs) can classify skin diseases at a level equivalent to a dermatologist, but their performance in specific areas requires further research.\n## OBJECTIVE\nTo evaluate the performance of a trained DCNN-based algorithm in classifying benign and malignant lip diseases.\n## METHODS\nA training set of 1629 images (743 malignant, 886 benign) was used with Inception-Resnet-V2. Performance was evaluated using another set of 344 images and 281 images from other hospitals. Classifications by 44 participants (six board-certified dermatologists, 12 dermatology residents, nine medical doctors not specialized in dermatology and 17 medical students) were used for comparison.\n## RESULTS\nThe outcomes based on the area under curve, sensitivity and specificity were 0\u00b7827 [95% confidence interval (CI) 0\u00b7782-0\u00b7873], 0\u00b7755 (95% CI 0\u00b7673-0\u00b7827) and 0\u00b7803 (95% CI 0\u00b7752-0\u00b7855), respectively, for the set of 344 images; and 0\u00b7774 (95% CI 0\u00b7699-0\u00b7849), 0\u00b7702 (95% CI 0\u00b7579-0\u00b7808) and 0\u00b7759 (95% CI 0\u00b7701-0\u00b7813), respectively, for the set of 281 images. The DCNN was equivalent to the dermatologists and superior to the nondermatologists in classifying malignancy. After referencing the DCNN result, the mean \u00b1 SD Youden index increased significantly for nondermatologists, from 0\u00b7201 \u00b1 0\u00b7156 to 0\u00b7322 \u00b1 0\u00b7141 (P < 0\u00b7001).\n## CONCLUSIONS\nDCNNs can classify lip diseases at a level similar to dermatologists. This will help unskilled physicians discriminate between benign and malignant lip diseases. What's already known about this topic? Deep convolutional neural networks (DCNNs) can classify malignant and benign skin diseases at a level equivalent to dermatologists. The lips are a unique feature in terms of histology and morphology. Previous studies of DCNNs have not investigated tumours on specific locations. What does this study add? This study shows that DCNNs can distinguish rare malignant and benign lip disorders at the same rate as dermatologists. DCNNs can help nondermatologists to distinguish malignant lip diseases. What are the clinical implications of this work? DCNNs can distinguish malignant and benign skin diseases even at specific locations such as the lips, as well as board-certified dermatologists. Malignant lip diseases are rare and difficult for less trained doctors to differentiate them from benign lesions. This study shows that in dermatology, DCNN can help improve decision-making processes for rare skin diseases in specific areas of the body.\n",
        "methods": "**METHODS**\n\n_Datasets_\n\nThe dataset in this study was taken from the photo database of Seoul National University Hospital (SNUH). The dataset contains clinical photos of dermatologic patients that were taken since October 2004. Photos related to lip diseases were selected in order to build lip datasets. The diagnoses were annotated based on the clinical and histologic findings. In addition to the SNUH photo database, clinical images of lip diseases were collected from two other affiliated hospitals, Seoul National University Bundang Hospital and SMG-SNU Boramae Medical Center. The study was approved by the Institutional Review Board of SNUH (IRB No. 1811-111-987).\n\n_Classification of lip diseases_\n\nThe lip disease diagnoses were classified either malignant or benign. Malignant lip diseases included skin cancers (SCC, BCC, MM, or adenocarcinoma) and precancerous lesions (actinic cheilitis, Bowen\u2019s disease, and melanoma in situ). Benign lip diseases included lichen planus, melanotic macule, pyogenic granuloma, plasma cell cheilitis, and venous lake.\n\n_Data preparation_\n\nBlurry or images that were inadequate for clinical diagnosis were removed from the datasets. In cases where the image contained the whole face, it was cropped to show only the lip or skin lesions. In some cases, several photos were taken of same patient from multiple viewpoints. Thus, the photos were numbered based on the patient to prevent photos of the same patient from being split into both the training and testing sets.\n\nA total of 2254 images were used for this study (Table 1). This included 1973 images of 404 patients who were diagnosed with lip disorders at SNUH between October 2004 and October 2018. In total, 34 classes of lip disorders including 4 lip cancers, 3 precancerous lip disorders, and 27 benign lip diseases were included for evaluation. The details and number of photos for each diagnosis are listed in Table 1. The images of lip diseases were classified as malignant (853 images) or benign (1120 images).\n\nIn total, 1629 of the SNUH images (743 malignant and 886 benign) were selected for the training set. The remaining 344 SNUH images (110 malignant and 234 benign) were used as the testing set along with 281 images (57 malignant and 224 benign) from Seoul National University Bundang Hospital (225 images) and SMG-SNU Boramae Medical Center (56 images).\n\n_DCNN algorithm_\n\nVarious DCNN structures show good results for image classification. The Inception-Resnet V2 (Google Inc., Mountain View, CA) structure, which recorded a Top-5 error rate of 3.1% for the ILSVRC 2012 test set, was selected for this study. The structure of Inception Resnet-V2 uses factorization and residual connections. Factorization can reduce the computational power and dimensionality; hence, it can prevent overfitting, which is often problematic. Residual connections were proposed by Microsoft ResNet and they can prevent instability caused by the deeper network, which can suffer from vanishing gradients during training. The hyper-parameters were set as follows: weight decay=0.00004, batch norm decay=0.9997, batch norm epsilon=0.001, dropout keep probability=0.8, and activation function ReLu with Adam-optimizer.\n\n_Comparing the performance of the trained DCNN to human participants_\n\nWe compared the performance of 44 medical doctors and medical students (including 6 board-certified dermatologists, 12 dermatology residents, 9 specialists who did not major in dermatology, and 17 medical students) to the trained DCNN. First, we randomly selected 40 images, 20 malignant and 20 benign. These images were viewed by the participants who were then asked: \u201cDo you think that this lip lesion is malignant or benign?\u201d Once the first test was finished, the participants received the output values from the trained DCNN. They were then asked the question again. The DCNN results indicated the degree of malignancy from 0 to 100.00%.\n\n_Statistical analysis_\n\nThe performance of the trained DCNN was evaluated using the testing dataset of 344 images from SNUH and the 281 additional images. A receiver operating characteristic (ROC) curve was drawn in order to compare the performance of the DCNN and the participants. The area under the curve (AUC) was taken as a primary outcome while sensitivity and specificity were also calculated. Delong\u2019s test was used to compare the ROC curves generated from the two image sets. Youden\u2019s index (sensitivity + specificity \u2013 1) was calculated before and after the DCNN results were reviewed and the values were analysed using the Wilcoxon signed rank test. Statistical analysis was performed using R statistical software version 3.51 (R Foundation for Statistical Computing, Vienna, Austria).",
        "results": "**RESULTS**\n\n_Outcomes of trained DCNN on testing set_\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image testing set were 0.827 (95% confidence interval (CI): 0.782\u20130.873), 0.755 (95% CI: 0.673\u20130.827), and 0.803 (95% CI: 0.752\u20130.855), respectively. For the additional 281 images these values were 0.774 (95% CI: 0.699\u20130.849), 0.702 (95% CI: 0.579\u20130.808), and 0.759 (95% CI: 0.701\u20130.813), respectively (Figure 1). There was no significant difference between the two ROC curves (p=0.229). The combined performance outcome for the 625 images in the two test tests, the AUC, sensitivity, and specificity were 0.811 (95% CI: 0.773\u20130.850), 0.737 (95% CI: 0.671\u20130.802), and 0.779 (95% CI: 0.740\u20130.817), respectively.\n\n_Outcomes for the trained DCNN compared to human participants_\n\nThe trained DCNN outperformed most of the participants, including some board-certified dermatologists, in the ROC curve (Figure 2A).\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 40 images chosen from the testing set were 0.792 (95% CI: 0.651\u20130.934), 0.850 (95% CI: 0.700\u20131.000), and 0.650 (95% CI: 0.450\u20130.850), respectively.\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants are shown in Table 2.\n\n_Outcomes of human participants changed after referencing the output value of DCNN_\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants after they referenced the DCNN output are shown in Table 2. The Youden\u2019s index of the less skilled (non-dermatologist) group increased significantly once they referenced the DCNN output (from 0.201 \u00b1 0.156 to 0.322 \u00b1 0.141, _p < 0.001), as shown in Figure 2B. However, there was no significant difference in the dermatologist group (from 0.375 \u00b1 0.167 to 0.444 \u00b1 0.108, p = 0.098).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.",
        "abstract": "## BACKGROUND\nDeep convolutional neural networks (DCNNs) can classify skin diseases at a level equivalent to a dermatologist, but their performance in specific areas requires further research.\n## OBJECTIVE\nTo evaluate the performance of a trained DCNN-based algorithm in classifying benign and malignant lip diseases.\n## METHODS\nA training set of 1629 images (743 malignant, 886 benign) was used with Inception-Resnet-V2. Performance was evaluated using another set of 344 images and 281 images from other hospitals. Classifications by 44 participants (six board-certified dermatologists, 12 dermatology residents, nine medical doctors not specialized in dermatology and 17 medical students) were used for comparison.\n## RESULTS\nThe outcomes based on the area under curve, sensitivity and specificity were 0\u00b7827 [95% confidence interval (CI) 0\u00b7782-0\u00b7873], 0\u00b7755 (95% CI 0\u00b7673-0\u00b7827) and 0\u00b7803 (95% CI 0\u00b7752-0\u00b7855), respectively, for the set of 344 images; and 0\u00b7774 (95% CI 0\u00b7699-0\u00b7849), 0\u00b7702 (95% CI 0\u00b7579-0\u00b7808) and 0\u00b7759 (95% CI 0\u00b7701-0\u00b7813), respectively, for the set of 281 images. The DCNN was equivalent to the dermatologists and superior to the nondermatologists in classifying malignancy. After referencing the DCNN result, the mean \u00b1 SD Youden index increased significantly for nondermatologists, from 0\u00b7201 \u00b1 0\u00b7156 to 0\u00b7322 \u00b1 0\u00b7141 (P < 0\u00b7001).\n## CONCLUSIONS\nDCNNs can classify lip diseases at a level similar to dermatologists. This will help unskilled physicians discriminate between benign and malignant lip diseases. What's already known about this topic? Deep convolutional neural networks (DCNNs) can classify malignant and benign skin diseases at a level equivalent to dermatologists. The lips are a unique feature in terms of histology and morphology. Previous studies of DCNNs have not investigated tumours on specific locations. What does this study add? This study shows that DCNNs can distinguish rare malignant and benign lip disorders at the same rate as dermatologists. DCNNs can help nondermatologists to distinguish malignant lip diseases. What are the clinical implications of this work? DCNNs can distinguish malignant and benign skin diseases even at specific locations such as the lips, as well as board-certified dermatologists. Malignant lip diseases are rare and difficult for less trained doctors to differentiate them from benign lesions. This study shows that in dermatology, DCNN can help improve decision-making processes for rare skin diseases in specific areas of the body.\n",
        "methods": "**METHODS**\n\n_Datasets_\n\nThe dataset in this study was taken from the photo database of Seoul National University Hospital (SNUH). The dataset contains clinical photos of dermatologic patients that were taken since October 2004. Photos related to lip diseases were selected in order to build lip datasets. The diagnoses were annotated based on the clinical and histologic findings. In addition to the SNUH photo database, clinical images of lip diseases were collected from two other affiliated hospitals, Seoul National University Bundang Hospital and SMG-SNU Boramae Medical Center. The study was approved by the Institutional Review Board of SNUH (IRB No. 1811-111-987).\n\n_Classification of lip diseases_\n\nThe lip disease diagnoses were classified either malignant or benign. Malignant lip diseases included skin cancers (SCC, BCC, MM, or adenocarcinoma) and precancerous lesions (actinic cheilitis, Bowen\u2019s disease, and melanoma in situ). Benign lip diseases included lichen planus, melanotic macule, pyogenic granuloma, plasma cell cheilitis, and venous lake.\n\n_Data preparation_\n\nBlurry or images that were inadequate for clinical diagnosis were removed from the datasets. In cases where the image contained the whole face, it was cropped to show only the lip or skin lesions. In some cases, several photos were taken of same patient from multiple viewpoints. Thus, the photos were numbered based on the patient to prevent photos of the same patient from being split into both the training and testing sets.\n\nA total of 2254 images were used for this study (Table 1). This included 1973 images of 404 patients who were diagnosed with lip disorders at SNUH between October 2004 and October 2018. In total, 34 classes of lip disorders including 4 lip cancers, 3 precancerous lip disorders, and 27 benign lip diseases were included for evaluation. The details and number of photos for each diagnosis are listed in Table 1. The images of lip diseases were classified as malignant (853 images) or benign (1120 images).\n\nIn total, 1629 of the SNUH images (743 malignant and 886 benign) were selected for the training set. The remaining 344 SNUH images (110 malignant and 234 benign) were used as the testing set along with 281 images (57 malignant and 224 benign) from Seoul National University Bundang Hospital (225 images) and SMG-SNU Boramae Medical Center (56 images).\n\n_DCNN algorithm_\n\nVarious DCNN structures show good results for image classification. The Inception-Resnet V2 (Google Inc., Mountain View, CA) structure, which recorded a Top-5 error rate of 3.1% for the ILSVRC 2012 test set, was selected for this study. The structure of Inception Resnet-V2 uses factorization and residual connections. Factorization can reduce the computational power and dimensionality; hence, it can prevent overfitting, which is often problematic. Residual connections were proposed by Microsoft ResNet and they can prevent instability caused by the deeper network, which can suffer from vanishing gradients during training. The hyper-parameters were set as follows: weight decay=0.00004, batch norm decay=0.9997, batch norm epsilon=0.001, dropout keep probability=0.8, and activation function ReLu with Adam-optimizer.\n\n_Comparing the performance of the trained DCNN to human participants_\n\nWe compared the performance of 44 medical doctors and medical students (including 6 board-certified dermatologists, 12 dermatology residents, 9 specialists who did not major in dermatology, and 17 medical students) to the trained DCNN. First, we randomly selected 40 images, 20 malignant and 20 benign. These images were viewed by the participants who were then asked: \u201cDo you think that this lip lesion is malignant or benign?\u201d Once the first test was finished, the participants received the output values from the trained DCNN. They were then asked the question again. The DCNN results indicated the degree of malignancy from 0 to 100.00%.\n\n_Statistical analysis_\n\nThe performance of the trained DCNN was evaluated using the testing dataset of 344 images from SNUH and the 281 additional images. A receiver operating characteristic (ROC) curve was drawn in order to compare the performance of the DCNN and the participants. The area under the curve (AUC) was taken as a primary outcome while sensitivity and specificity were also calculated. Delong\u2019s test was used to compare the ROC curves generated from the two image sets. Youden\u2019s index (sensitivity + specificity \u2013 1) was calculated before and after the DCNN results were reviewed and the values were analysed using the Wilcoxon signed rank test. Statistical analysis was performed using R statistical software version 3.51 (R Foundation for Statistical Computing, Vienna, Austria).",
        "results": "**RESULTS**\n\n_Outcomes of trained DCNN on testing set_\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image testing set were 0.827 (95% confidence interval (CI): 0.782\u20130.873), 0.755 (95% CI: 0.673\u20130.827), and 0.803 (95% CI: 0.752\u20130.855), respectively. For the additional 281 images these values were 0.774 (95% CI: 0.699\u20130.849), 0.702 (95% CI: 0.579\u20130.808), and 0.759 (95% CI: 0.701\u20130.813), respectively (Figure 1). There was no significant difference between the two ROC curves (p=0.229). The combined performance outcome for the 625 images in the two test tests, the AUC, sensitivity, and specificity were 0.811 (95% CI: 0.773\u20130.850), 0.737 (95% CI: 0.671\u20130.802), and 0.779 (95% CI: 0.740\u20130.817), respectively.\n\n_Outcomes for the trained DCNN compared to human participants_\n\nThe trained DCNN outperformed most of the participants, including some board-certified dermatologists, in the ROC curve (Figure 2A).\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 40 images chosen from the testing set were 0.792 (95% CI: 0.651\u20130.934), 0.850 (95% CI: 0.700\u20131.000), and 0.650 (95% CI: 0.450\u20130.850), respectively.\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants are shown in Table 2.\n\n_Outcomes of human participants changed after referencing the output value of DCNN_\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants after they referenced the DCNN output are shown in Table 2. The Youden\u2019s index of the less skilled (non-dermatologist) group increased significantly once they referenced the DCNN output (from 0.201 \u00b1 0.156 to 0.322 \u00b1 0.141, _p < 0.001), as shown in Figure 2B. However, there was no significant difference in the dermatologist group (from 0.375 \u00b1 0.167 to 0.444 \u00b1 0.108, p = 0.098).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.",
        "abstract": "## BACKGROUND\nDeep convolutional neural networks (DCNNs) can classify skin diseases at a level equivalent to a dermatologist, but their performance in specific areas requires further research.\n## OBJECTIVE\nTo evaluate the performance of a trained DCNN-based algorithm in classifying benign and malignant lip diseases.\n## METHODS\nA training set of 1629 images (743 malignant, 886 benign) was used with Inception-Resnet-V2. Performance was evaluated using another set of 344 images and 281 images from other hospitals. Classifications by 44 participants (six board-certified dermatologists, 12 dermatology residents, nine medical doctors not specialized in dermatology and 17 medical students) were used for comparison.\n## RESULTS\nThe outcomes based on the area under curve, sensitivity and specificity were 0\u00b7827 [95% confidence interval (CI) 0\u00b7782-0\u00b7873], 0\u00b7755 (95% CI 0\u00b7673-0\u00b7827) and 0\u00b7803 (95% CI 0\u00b7752-0\u00b7855), respectively, for the set of 344 images; and 0\u00b7774 (95% CI 0\u00b7699-0\u00b7849), 0\u00b7702 (95% CI 0\u00b7579-0\u00b7808) and 0\u00b7759 (95% CI 0\u00b7701-0\u00b7813), respectively, for the set of 281 images. The DCNN was equivalent to the dermatologists and superior to the nondermatologists in classifying malignancy. After referencing the DCNN result, the mean \u00b1 SD Youden index increased significantly for nondermatologists, from 0\u00b7201 \u00b1 0\u00b7156 to 0\u00b7322 \u00b1 0\u00b7141 (P < 0\u00b7001).\n## CONCLUSIONS\nDCNNs can classify lip diseases at a level similar to dermatologists. This will help unskilled physicians discriminate between benign and malignant lip diseases. What's already known about this topic? Deep convolutional neural networks (DCNNs) can classify malignant and benign skin diseases at a level equivalent to dermatologists. The lips are a unique feature in terms of histology and morphology. Previous studies of DCNNs have not investigated tumours on specific locations. What does this study add? This study shows that DCNNs can distinguish rare malignant and benign lip disorders at the same rate as dermatologists. DCNNs can help nondermatologists to distinguish malignant lip diseases. What are the clinical implications of this work? DCNNs can distinguish malignant and benign skin diseases even at specific locations such as the lips, as well as board-certified dermatologists. Malignant lip diseases are rare and difficult for less trained doctors to differentiate them from benign lesions. This study shows that in dermatology, DCNN can help improve decision-making processes for rare skin diseases in specific areas of the body.\n",
        "methods": "**METHODS**\n\n_Datasets_\n\nThe dataset in this study was taken from the photo database of Seoul National University Hospital (SNUH). The dataset contains clinical photos of dermatologic patients that were taken since October 2004. Photos related to lip diseases were selected in order to build lip datasets. The diagnoses were annotated based on the clinical and histologic findings. In addition to the SNUH photo database, clinical images of lip diseases were collected from two other affiliated hospitals, Seoul National University Bundang Hospital and SMG-SNU Boramae Medical Center. The study was approved by the Institutional Review Board of SNUH (IRB No. 1811-111-987).\n\n_Classification of lip diseases_\n\nThe lip disease diagnoses were classified either malignant or benign. Malignant lip diseases included skin cancers (SCC, BCC, MM, or adenocarcinoma) and precancerous lesions (actinic cheilitis, Bowen\u2019s disease, and melanoma in situ). Benign lip diseases included lichen planus, melanotic macule, pyogenic granuloma, plasma cell cheilitis, and venous lake.\n\n_Data preparation_\n\nBlurry or images that were inadequate for clinical diagnosis were removed from the datasets. In cases where the image contained the whole face, it was cropped to show only the lip or skin lesions. In some cases, several photos were taken of same patient from multiple viewpoints. Thus, the photos were numbered based on the patient to prevent photos of the same patient from being split into both the training and testing sets.\n\nA total of 2254 images were used for this study (Table 1). This included 1973 images of 404 patients who were diagnosed with lip disorders at SNUH between October 2004 and October 2018. In total, 34 classes of lip disorders including 4 lip cancers, 3 precancerous lip disorders, and 27 benign lip diseases were included for evaluation. The details and number of photos for each diagnosis are listed in Table 1. The images of lip diseases were classified as malignant (853 images) or benign (1120 images).\n\nIn total, 1629 of the SNUH images (743 malignant and 886 benign) were selected for the training set. The remaining 344 SNUH images (110 malignant and 234 benign) were used as the testing set along with 281 images (57 malignant and 224 benign) from Seoul National University Bundang Hospital (225 images) and SMG-SNU Boramae Medical Center (56 images).\n\n_DCNN algorithm_\n\nVarious DCNN structures show good results for image classification. The Inception-Resnet V2 (Google Inc., Mountain View, CA) structure, which recorded a Top-5 error rate of 3.1% for the ILSVRC 2012 test set, was selected for this study. The structure of Inception Resnet-V2 uses factorization and residual connections. Factorization can reduce the computational power and dimensionality; hence, it can prevent overfitting, which is often problematic. Residual connections were proposed by Microsoft ResNet and they can prevent instability caused by the deeper network, which can suffer from vanishing gradients during training. The hyper-parameters were set as follows: weight decay=0.00004, batch norm decay=0.9997, batch norm epsilon=0.001, dropout keep probability=0.8, and activation function ReLu with Adam-optimizer.\n\n_Comparing the performance of the trained DCNN to human participants_\n\nWe compared the performance of 44 medical doctors and medical students (including 6 board-certified dermatologists, 12 dermatology residents, 9 specialists who did not major in dermatology, and 17 medical students) to the trained DCNN. First, we randomly selected 40 images, 20 malignant and 20 benign. These images were viewed by the participants who were then asked: \u201cDo you think that this lip lesion is malignant or benign?\u201d Once the first test was finished, the participants received the output values from the trained DCNN. They were then asked the question again. The DCNN results indicated the degree of malignancy from 0 to 100.00%.\n\n_Statistical analysis_\n\nThe performance of the trained DCNN was evaluated using the testing dataset of 344 images from SNUH and the 281 additional images. A receiver operating characteristic (ROC) curve was drawn in order to compare the performance of the DCNN and the participants. The area under the curve (AUC) was taken as a primary outcome while sensitivity and specificity were also calculated. Delong\u2019s test was used to compare the ROC curves generated from the two image sets. Youden\u2019s index (sensitivity + specificity \u2013 1) was calculated before and after the DCNN results were reviewed and the values were analysed using the Wilcoxon signed rank test. Statistical analysis was performed using R statistical software version 3.51 (R Foundation for Statistical Computing, Vienna, Austria).",
        "results": "**RESULTS**\n\n_Outcomes of trained DCNN on testing set_\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image testing set were 0.827 (95% confidence interval (CI): 0.782\u20130.873), 0.755 (95% CI: 0.673\u20130.827), and 0.803 (95% CI: 0.752\u20130.855), respectively. For the additional 281 images these values were 0.774 (95% CI: 0.699\u20130.849), 0.702 (95% CI: 0.579\u20130.808), and 0.759 (95% CI: 0.701\u20130.813), respectively (Figure 1). There was no significant difference between the two ROC curves (p=0.229). The combined performance outcome for the 625 images in the two test tests, the AUC, sensitivity, and specificity were 0.811 (95% CI: 0.773\u20130.850), 0.737 (95% CI: 0.671\u20130.802), and 0.779 (95% CI: 0.740\u20130.817), respectively.\n\n_Outcomes for the trained DCNN compared to human participants_\n\nThe trained DCNN outperformed most of the participants, including some board-certified dermatologists, in the ROC curve (Figure 2A).\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 40 images chosen from the testing set were 0.792 (95% CI: 0.651\u20130.934), 0.850 (95% CI: 0.700\u20131.000), and 0.650 (95% CI: 0.450\u20130.850), respectively.\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants are shown in Table 2.\n\n_Outcomes of human participants changed after referencing the output value of DCNN_\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants after they referenced the DCNN output are shown in Table 2. The Youden\u2019s index of the less skilled (non-dermatologist) group increased significantly once they referenced the DCNN output (from 0.201 \u00b1 0.156 to 0.322 \u00b1 0.141, _p < 0.001), as shown in Figure 2B. However, there was no significant difference in the dermatologist group (from 0.375 \u00b1 0.167 to 0.444 \u00b1 0.108, p = 0.098).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.",
        "abstract": "## BACKGROUND\nDeep convolutional neural networks (DCNNs) can classify skin diseases at a level equivalent to a dermatologist, but their performance in specific areas requires further research.\n## OBJECTIVE\nTo evaluate the performance of a trained DCNN-based algorithm in classifying benign and malignant lip diseases.\n## METHODS\nA training set of 1629 images (743 malignant, 886 benign) was used with Inception-Resnet-V2. Performance was evaluated using another set of 344 images and 281 images from other hospitals. Classifications by 44 participants (six board-certified dermatologists, 12 dermatology residents, nine medical doctors not specialized in dermatology and 17 medical students) were used for comparison.\n## RESULTS\nThe outcomes based on the area under curve, sensitivity and specificity were 0\u00b7827 [95% confidence interval (CI) 0\u00b7782-0\u00b7873], 0\u00b7755 (95% CI 0\u00b7673-0\u00b7827) and 0\u00b7803 (95% CI 0\u00b7752-0\u00b7855), respectively, for the set of 344 images; and 0\u00b7774 (95% CI 0\u00b7699-0\u00b7849), 0\u00b7702 (95% CI 0\u00b7579-0\u00b7808) and 0\u00b7759 (95% CI 0\u00b7701-0\u00b7813), respectively, for the set of 281 images. The DCNN was equivalent to the dermatologists and superior to the nondermatologists in classifying malignancy. After referencing the DCNN result, the mean \u00b1 SD Youden index increased significantly for nondermatologists, from 0\u00b7201 \u00b1 0\u00b7156 to 0\u00b7322 \u00b1 0\u00b7141 (P < 0\u00b7001).\n## CONCLUSIONS\nDCNNs can classify lip diseases at a level similar to dermatologists. This will help unskilled physicians discriminate between benign and malignant lip diseases. What's already known about this topic? Deep convolutional neural networks (DCNNs) can classify malignant and benign skin diseases at a level equivalent to dermatologists. The lips are a unique feature in terms of histology and morphology. Previous studies of DCNNs have not investigated tumours on specific locations. What does this study add? This study shows that DCNNs can distinguish rare malignant and benign lip disorders at the same rate as dermatologists. DCNNs can help nondermatologists to distinguish malignant lip diseases. What are the clinical implications of this work? DCNNs can distinguish malignant and benign skin diseases even at specific locations such as the lips, as well as board-certified dermatologists. Malignant lip diseases are rare and difficult for less trained doctors to differentiate them from benign lesions. This study shows that in dermatology, DCNN can help improve decision-making processes for rare skin diseases in specific areas of the body.\n",
        "methods": "**METHODS**\n\n_Datasets_\n\nThe dataset in this study was taken from the photo database of Seoul National University Hospital (SNUH). The dataset contains clinical photos of dermatologic patients that were taken since October 2004. Photos related to lip diseases were selected in order to build lip datasets. The diagnoses were annotated based on the clinical and histologic findings. In addition to the SNUH photo database, clinical images of lip diseases were collected from two other affiliated hospitals, Seoul National University Bundang Hospital and SMG-SNU Boramae Medical Center. The study was approved by the Institutional Review Board of SNUH (IRB No. 1811-111-987).\n\n_Classification of lip diseases_\n\nThe lip disease diagnoses were classified either malignant or benign. Malignant lip diseases included skin cancers (SCC, BCC, MM, or adenocarcinoma) and precancerous lesions (actinic cheilitis, Bowen\u2019s disease, and melanoma in situ). Benign lip diseases included lichen planus, melanotic macule, pyogenic granuloma, plasma cell cheilitis, and venous lake.\n\n_Data preparation_\n\nBlurry or images that were inadequate for clinical diagnosis were removed from the datasets. In cases where the image contained the whole face, it was cropped to show only the lip or skin lesions. In some cases, several photos were taken of same patient from multiple viewpoints. Thus, the photos were numbered based on the patient to prevent photos of the same patient from being split into both the training and testing sets.\n\nA total of 2254 images were used for this study (Table 1). This included 1973 images of 404 patients who were diagnosed with lip disorders at SNUH between October 2004 and October 2018. In total, 34 classes of lip disorders including 4 lip cancers, 3 precancerous lip disorders, and 27 benign lip diseases were included for evaluation. The details and number of photos for each diagnosis are listed in Table 1. The images of lip diseases were classified as malignant (853 images) or benign (1120 images).\n\nIn total, 1629 of the SNUH images (743 malignant and 886 benign) were selected for the training set. The remaining 344 SNUH images (110 malignant and 234 benign) were used as the testing set along with 281 images (57 malignant and 224 benign) from Seoul National University Bundang Hospital (225 images) and SMG-SNU Boramae Medical Center (56 images).\n\n_DCNN algorithm_\n\nVarious DCNN structures show good results for image classification. The Inception-Resnet V2 (Google Inc., Mountain View, CA) structure, which recorded a Top-5 error rate of 3.1% for the ILSVRC 2012 test set, was selected for this study. The structure of Inception Resnet-V2 uses factorization and residual connections. Factorization can reduce the computational power and dimensionality; hence, it can prevent overfitting, which is often problematic. Residual connections were proposed by Microsoft ResNet and they can prevent instability caused by the deeper network, which can suffer from vanishing gradients during training. The hyper-parameters were set as follows: weight decay=0.00004, batch norm decay=0.9997, batch norm epsilon=0.001, dropout keep probability=0.8, and activation function ReLu with Adam-optimizer.\n\n_Comparing the performance of the trained DCNN to human participants_\n\nWe compared the performance of 44 medical doctors and medical students (including 6 board-certified dermatologists, 12 dermatology residents, 9 specialists who did not major in dermatology, and 17 medical students) to the trained DCNN. First, we randomly selected 40 images, 20 malignant and 20 benign. These images were viewed by the participants who were then asked: \u201cDo you think that this lip lesion is malignant or benign?\u201d Once the first test was finished, the participants received the output values from the trained DCNN. They were then asked the question again. The DCNN results indicated the degree of malignancy from 0 to 100.00%.\n\n_Statistical analysis_\n\nThe performance of the trained DCNN was evaluated using the testing dataset of 344 images from SNUH and the 281 additional images. A receiver operating characteristic (ROC) curve was drawn in order to compare the performance of the DCNN and the participants. The area under the curve (AUC) was taken as a primary outcome while sensitivity and specificity were also calculated. Delong\u2019s test was used to compare the ROC curves generated from the two image sets. Youden\u2019s index (sensitivity + specificity \u2013 1) was calculated before and after the DCNN results were reviewed and the values were analysed using the Wilcoxon signed rank test. Statistical analysis was performed using R statistical software version 3.51 (R Foundation for Statistical Computing, Vienna, Austria).",
        "results": "**RESULTS**\n\n_Outcomes of trained DCNN on testing set_\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image testing set were 0.827 (95% confidence interval (CI): 0.782\u20130.873), 0.755 (95% CI: 0.673\u20130.827), and 0.803 (95% CI: 0.752\u20130.855), respectively. For the additional 281 images these values were 0.774 (95% CI: 0.699\u20130.849), 0.702 (95% CI: 0.579\u20130.808), and 0.759 (95% CI: 0.701\u20130.813), respectively (Figure 1). There was no significant difference between the two ROC curves (p=0.229). The combined performance outcome for the 625 images in the two test tests, the AUC, sensitivity, and specificity were 0.811 (95% CI: 0.773\u20130.850), 0.737 (95% CI: 0.671\u20130.802), and 0.779 (95% CI: 0.740\u20130.817), respectively.\n\n_Outcomes for the trained DCNN compared to human participants_\n\nThe trained DCNN outperformed most of the participants, including some board-certified dermatologists, in the ROC curve (Figure 2A).\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 40 images chosen from the testing set were 0.792 (95% CI: 0.651\u20130.934), 0.850 (95% CI: 0.700\u20131.000), and 0.650 (95% CI: 0.450\u20130.850), respectively.\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants are shown in Table 2.\n\n_Outcomes of human participants changed after referencing the output value of DCNN_\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants after they referenced the DCNN output are shown in Table 2. The Youden\u2019s index of the less skilled (non-dermatologist) group increased significantly once they referenced the DCNN output (from 0.201 \u00b1 0.156 to 0.322 \u00b1 0.141, _p < 0.001), as shown in Figure 2B. However, there was no significant difference in the dermatologist group (from 0.375 \u00b1 0.167 to 0.444 \u00b1 0.108, p = 0.098).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.",
        "abstract": "## BACKGROUND\nDeep convolutional neural networks (DCNNs) can classify skin diseases at a level equivalent to a dermatologist, but their performance in specific areas requires further research.\n## OBJECTIVE\nTo evaluate the performance of a trained DCNN-based algorithm in classifying benign and malignant lip diseases.\n## METHODS\nA training set of 1629 images (743 malignant, 886 benign) was used with Inception-Resnet-V2. Performance was evaluated using another set of 344 images and 281 images from other hospitals. Classifications by 44 participants (six board-certified dermatologists, 12 dermatology residents, nine medical doctors not specialized in dermatology and 17 medical students) were used for comparison.\n## RESULTS\nThe outcomes based on the area under curve, sensitivity and specificity were 0\u00b7827 [95% confidence interval (CI) 0\u00b7782-0\u00b7873], 0\u00b7755 (95% CI 0\u00b7673-0\u00b7827) and 0\u00b7803 (95% CI 0\u00b7752-0\u00b7855), respectively, for the set of 344 images; and 0\u00b7774 (95% CI 0\u00b7699-0\u00b7849), 0\u00b7702 (95% CI 0\u00b7579-0\u00b7808) and 0\u00b7759 (95% CI 0\u00b7701-0\u00b7813), respectively, for the set of 281 images. The DCNN was equivalent to the dermatologists and superior to the nondermatologists in classifying malignancy. After referencing the DCNN result, the mean \u00b1 SD Youden index increased significantly for nondermatologists, from 0\u00b7201 \u00b1 0\u00b7156 to 0\u00b7322 \u00b1 0\u00b7141 (P < 0\u00b7001).\n## CONCLUSIONS\nDCNNs can classify lip diseases at a level similar to dermatologists. This will help unskilled physicians discriminate between benign and malignant lip diseases. What's already known about this topic? Deep convolutional neural networks (DCNNs) can classify malignant and benign skin diseases at a level equivalent to dermatologists. The lips are a unique feature in terms of histology and morphology. Previous studies of DCNNs have not investigated tumours on specific locations. What does this study add? This study shows that DCNNs can distinguish rare malignant and benign lip disorders at the same rate as dermatologists. DCNNs can help nondermatologists to distinguish malignant lip diseases. What are the clinical implications of this work? DCNNs can distinguish malignant and benign skin diseases even at specific locations such as the lips, as well as board-certified dermatologists. Malignant lip diseases are rare and difficult for less trained doctors to differentiate them from benign lesions. This study shows that in dermatology, DCNN can help improve decision-making processes for rare skin diseases in specific areas of the body.\n",
        "methods": "**METHODS**\n\n_Datasets_\n\nThe dataset in this study was taken from the photo database of Seoul National University Hospital (SNUH). The dataset contains clinical photos of dermatologic patients that were taken since October 2004. Photos related to lip diseases were selected in order to build lip datasets. The diagnoses were annotated based on the clinical and histologic findings. In addition to the SNUH photo database, clinical images of lip diseases were collected from two other affiliated hospitals, Seoul National University Bundang Hospital and SMG-SNU Boramae Medical Center. The study was approved by the Institutional Review Board of SNUH (IRB No. 1811-111-987).\n\n_Classification of lip diseases_\n\nThe lip disease diagnoses were classified either malignant or benign. Malignant lip diseases included skin cancers (SCC, BCC, MM, or adenocarcinoma) and precancerous lesions (actinic cheilitis, Bowen\u2019s disease, and melanoma in situ). Benign lip diseases included lichen planus, melanotic macule, pyogenic granuloma, plasma cell cheilitis, and venous lake.\n\n_Data preparation_\n\nBlurry or images that were inadequate for clinical diagnosis were removed from the datasets. In cases where the image contained the whole face, it was cropped to show only the lip or skin lesions. In some cases, several photos were taken of same patient from multiple viewpoints. Thus, the photos were numbered based on the patient to prevent photos of the same patient from being split into both the training and testing sets.\n\nA total of 2254 images were used for this study (Table 1). This included 1973 images of 404 patients who were diagnosed with lip disorders at SNUH between October 2004 and October 2018. In total, 34 classes of lip disorders including 4 lip cancers, 3 precancerous lip disorders, and 27 benign lip diseases were included for evaluation. The details and number of photos for each diagnosis are listed in Table 1. The images of lip diseases were classified as malignant (853 images) or benign (1120 images).\n\nIn total, 1629 of the SNUH images (743 malignant and 886 benign) were selected for the training set. The remaining 344 SNUH images (110 malignant and 234 benign) were used as the testing set along with 281 images (57 malignant and 224 benign) from Seoul National University Bundang Hospital (225 images) and SMG-SNU Boramae Medical Center (56 images).\n\n_DCNN algorithm_\n\nVarious DCNN structures show good results for image classification. The Inception-Resnet V2 (Google Inc., Mountain View, CA) structure, which recorded a Top-5 error rate of 3.1% for the ILSVRC 2012 test set, was selected for this study. The structure of Inception Resnet-V2 uses factorization and residual connections. Factorization can reduce the computational power and dimensionality; hence, it can prevent overfitting, which is often problematic. Residual connections were proposed by Microsoft ResNet and they can prevent instability caused by the deeper network, which can suffer from vanishing gradients during training. The hyper-parameters were set as follows: weight decay=0.00004, batch norm decay=0.9997, batch norm epsilon=0.001, dropout keep probability=0.8, and activation function ReLu with Adam-optimizer.\n\n_Comparing the performance of the trained DCNN to human participants_\n\nWe compared the performance of 44 medical doctors and medical students (including 6 board-certified dermatologists, 12 dermatology residents, 9 specialists who did not major in dermatology, and 17 medical students) to the trained DCNN. First, we randomly selected 40 images, 20 malignant and 20 benign. These images were viewed by the participants who were then asked: \u201cDo you think that this lip lesion is malignant or benign?\u201d Once the first test was finished, the participants received the output values from the trained DCNN. They were then asked the question again. The DCNN results indicated the degree of malignancy from 0 to 100.00%.\n\n_Statistical analysis_\n\nThe performance of the trained DCNN was evaluated using the testing dataset of 344 images from SNUH and the 281 additional images. A receiver operating characteristic (ROC) curve was drawn in order to compare the performance of the DCNN and the participants. The area under the curve (AUC) was taken as a primary outcome while sensitivity and specificity were also calculated. Delong\u2019s test was used to compare the ROC curves generated from the two image sets. Youden\u2019s index (sensitivity + specificity \u2013 1) was calculated before and after the DCNN results were reviewed and the values were analysed using the Wilcoxon signed rank test. Statistical analysis was performed using R statistical software version 3.51 (R Foundation for Statistical Computing, Vienna, Austria).",
        "results": "**RESULTS**\n\n_Outcomes of trained DCNN on testing set_\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image testing set were 0.827 (95% confidence interval (CI): 0.782\u20130.873), 0.755 (95% CI: 0.673\u20130.827), and 0.803 (95% CI: 0.752\u20130.855), respectively. For the additional 281 images these values were 0.774 (95% CI: 0.699\u20130.849), 0.702 (95% CI: 0.579\u20130.808), and 0.759 (95% CI: 0.701\u20130.813), respectively (Figure 1). There was no significant difference between the two ROC curves (p=0.229). The combined performance outcome for the 625 images in the two test tests, the AUC, sensitivity, and specificity were 0.811 (95% CI: 0.773\u20130.850), 0.737 (95% CI: 0.671\u20130.802), and 0.779 (95% CI: 0.740\u20130.817), respectively.\n\n_Outcomes for the trained DCNN compared to human participants_\n\nThe trained DCNN outperformed most of the participants, including some board-certified dermatologists, in the ROC curve (Figure 2A).\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 40 images chosen from the testing set were 0.792 (95% CI: 0.651\u20130.934), 0.850 (95% CI: 0.700\u20131.000), and 0.650 (95% CI: 0.450\u20130.850), respectively.\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants are shown in Table 2.\n\n_Outcomes of human participants changed after referencing the output value of DCNN_\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants after they referenced the DCNN output are shown in Table 2. The Youden\u2019s index of the less skilled (non-dermatologist) group increased significantly once they referenced the DCNN output (from 0.201 \u00b1 0.156 to 0.322 \u00b1 0.141, _p < 0.001), as shown in Figure 2B. However, there was no significant difference in the dermatologist group (from 0.375 \u00b1 0.167 to 0.444 \u00b1 0.108, p = 0.098).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.",
        "abstract": "## BACKGROUND\nDeep convolutional neural networks (DCNNs) can classify skin diseases at a level equivalent to a dermatologist, but their performance in specific areas requires further research.\n## OBJECTIVE\nTo evaluate the performance of a trained DCNN-based algorithm in classifying benign and malignant lip diseases.\n## METHODS\nA training set of 1629 images (743 malignant, 886 benign) was used with Inception-Resnet-V2. Performance was evaluated using another set of 344 images and 281 images from other hospitals. Classifications by 44 participants (six board-certified dermatologists, 12 dermatology residents, nine medical doctors not specialized in dermatology and 17 medical students) were used for comparison.\n## RESULTS\nThe outcomes based on the area under curve, sensitivity and specificity were 0\u00b7827 [95% confidence interval (CI) 0\u00b7782-0\u00b7873], 0\u00b7755 (95% CI 0\u00b7673-0\u00b7827) and 0\u00b7803 (95% CI 0\u00b7752-0\u00b7855), respectively, for the set of 344 images; and 0\u00b7774 (95% CI 0\u00b7699-0\u00b7849), 0\u00b7702 (95% CI 0\u00b7579-0\u00b7808) and 0\u00b7759 (95% CI 0\u00b7701-0\u00b7813), respectively, for the set of 281 images. The DCNN was equivalent to the dermatologists and superior to the nondermatologists in classifying malignancy. After referencing the DCNN result, the mean \u00b1 SD Youden index increased significantly for nondermatologists, from 0\u00b7201 \u00b1 0\u00b7156 to 0\u00b7322 \u00b1 0\u00b7141 (P < 0\u00b7001).\n## CONCLUSIONS\nDCNNs can classify lip diseases at a level similar to dermatologists. This will help unskilled physicians discriminate between benign and malignant lip diseases. What's already known about this topic? Deep convolutional neural networks (DCNNs) can classify malignant and benign skin diseases at a level equivalent to dermatologists. The lips are a unique feature in terms of histology and morphology. Previous studies of DCNNs have not investigated tumours on specific locations. What does this study add? This study shows that DCNNs can distinguish rare malignant and benign lip disorders at the same rate as dermatologists. DCNNs can help nondermatologists to distinguish malignant lip diseases. What are the clinical implications of this work? DCNNs can distinguish malignant and benign skin diseases even at specific locations such as the lips, as well as board-certified dermatologists. Malignant lip diseases are rare and difficult for less trained doctors to differentiate them from benign lesions. This study shows that in dermatology, DCNN can help improve decision-making processes for rare skin diseases in specific areas of the body.\n",
        "methods": "**METHODS**\n\n_Datasets_\n\nThe dataset in this study was taken from the photo database of Seoul National University Hospital (SNUH). The dataset contains clinical photos of dermatologic patients that were taken since October 2004. Photos related to lip diseases were selected in order to build lip datasets. The diagnoses were annotated based on the clinical and histologic findings. In addition to the SNUH photo database, clinical images of lip diseases were collected from two other affiliated hospitals, Seoul National University Bundang Hospital and SMG-SNU Boramae Medical Center. The study was approved by the Institutional Review Board of SNUH (IRB No. 1811-111-987).\n\n_Classification of lip diseases_\n\nThe lip disease diagnoses were classified either malignant or benign. Malignant lip diseases included skin cancers (SCC, BCC, MM, or adenocarcinoma) and precancerous lesions (actinic cheilitis, Bowen\u2019s disease, and melanoma in situ). Benign lip diseases included lichen planus, melanotic macule, pyogenic granuloma, plasma cell cheilitis, and venous lake.\n\n_Data preparation_\n\nBlurry or images that were inadequate for clinical diagnosis were removed from the datasets. In cases where the image contained the whole face, it was cropped to show only the lip or skin lesions. In some cases, several photos were taken of same patient from multiple viewpoints. Thus, the photos were numbered based on the patient to prevent photos of the same patient from being split into both the training and testing sets.\n\nA total of 2254 images were used for this study (Table 1). This included 1973 images of 404 patients who were diagnosed with lip disorders at SNUH between October 2004 and October 2018. In total, 34 classes of lip disorders including 4 lip cancers, 3 precancerous lip disorders, and 27 benign lip diseases were included for evaluation. The details and number of photos for each diagnosis are listed in Table 1. The images of lip diseases were classified as malignant (853 images) or benign (1120 images).\n\nIn total, 1629 of the SNUH images (743 malignant and 886 benign) were selected for the training set. The remaining 344 SNUH images (110 malignant and 234 benign) were used as the testing set along with 281 images (57 malignant and 224 benign) from Seoul National University Bundang Hospital (225 images) and SMG-SNU Boramae Medical Center (56 images).\n\n_DCNN algorithm_\n\nVarious DCNN structures show good results for image classification. The Inception-Resnet V2 (Google Inc., Mountain View, CA) structure, which recorded a Top-5 error rate of 3.1% for the ILSVRC 2012 test set, was selected for this study. The structure of Inception Resnet-V2 uses factorization and residual connections. Factorization can reduce the computational power and dimensionality; hence, it can prevent overfitting, which is often problematic. Residual connections were proposed by Microsoft ResNet and they can prevent instability caused by the deeper network, which can suffer from vanishing gradients during training. The hyper-parameters were set as follows: weight decay=0.00004, batch norm decay=0.9997, batch norm epsilon=0.001, dropout keep probability=0.8, and activation function ReLu with Adam-optimizer.\n\n_Comparing the performance of the trained DCNN to human participants_\n\nWe compared the performance of 44 medical doctors and medical students (including 6 board-certified dermatologists, 12 dermatology residents, 9 specialists who did not major in dermatology, and 17 medical students) to the trained DCNN. First, we randomly selected 40 images, 20 malignant and 20 benign. These images were viewed by the participants who were then asked: \u201cDo you think that this lip lesion is malignant or benign?\u201d Once the first test was finished, the participants received the output values from the trained DCNN. They were then asked the question again. The DCNN results indicated the degree of malignancy from 0 to 100.00%.\n\n_Statistical analysis_\n\nThe performance of the trained DCNN was evaluated using the testing dataset of 344 images from SNUH and the 281 additional images. A receiver operating characteristic (ROC) curve was drawn in order to compare the performance of the DCNN and the participants. The area under the curve (AUC) was taken as a primary outcome while sensitivity and specificity were also calculated. Delong\u2019s test was used to compare the ROC curves generated from the two image sets. Youden\u2019s index (sensitivity + specificity \u2013 1) was calculated before and after the DCNN results were reviewed and the values were analysed using the Wilcoxon signed rank test. Statistical analysis was performed using R statistical software version 3.51 (R Foundation for Statistical Computing, Vienna, Austria).",
        "results": "**RESULTS**\n\n_Outcomes of trained DCNN on testing set_\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 344 image testing set were 0.827 (95% confidence interval (CI): 0.782\u20130.873), 0.755 (95% CI: 0.673\u20130.827), and 0.803 (95% CI: 0.752\u20130.855), respectively. For the additional 281 images these values were 0.774 (95% CI: 0.699\u20130.849), 0.702 (95% CI: 0.579\u20130.808), and 0.759 (95% CI: 0.701\u20130.813), respectively (Figure 1). There was no significant difference between the two ROC curves (p=0.229). The combined performance outcome for the 625 images in the two test tests, the AUC, sensitivity, and specificity were 0.811 (95% CI: 0.773\u20130.850), 0.737 (95% CI: 0.671\u20130.802), and 0.779 (95% CI: 0.740\u20130.817), respectively.\n\n_Outcomes for the trained DCNN compared to human participants_\n\nThe trained DCNN outperformed most of the participants, including some board-certified dermatologists, in the ROC curve (Figure 2A).\n\nThe performance outcomes of the trained DCNN based on the AUC, sensitivity, and specificity for the 40 images chosen from the testing set were 0.792 (95% CI: 0.651\u20130.934), 0.850 (95% CI: 0.700\u20131.000), and 0.650 (95% CI: 0.450\u20130.850), respectively.\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants are shown in Table 2.\n\n_Outcomes of human participants changed after referencing the output value of DCNN_\n\nThe sensitivity, specificity, and Youden\u2019s index for the different groups of human participants after they referenced the DCNN output are shown in Table 2. The Youden\u2019s index of the less skilled (non-dermatologist) group increased significantly once they referenced the DCNN output (from 0.201 \u00b1 0.156 to 0.322 \u00b1 0.141, _p < 0.001), as shown in Figure 2B. However, there was no significant difference in the dermatologist group (from 0.375 \u00b1 0.167 to 0.444 \u00b1 0.108, p = 0.098).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and\u00a0Predicting Treatment Options for 134 Skin\u00a0Disorders.",
        "abstract": "Although deep learning algorithms have demonstrated expert-level performance, previous efforts were mostly binary classifications of limited disorders. We trained an algorithm with 220,680 images of 174 disorders and validated it using Edinburgh (1,300 images; 10 disorders) and SNU datasets (2,201 images; 134 disorders). The algorithm could accurately predict malignancy, suggest primary treatment options, render multi-class classification among 134 disorders, and improve the performance of medical professionals. The area under the curves for malignancy detection were 0.928 \u00b1 0.002 (Edinburgh) and 0.937 \u00b1 0.004 (SNU). The area under the curves of primary treatment suggestion (SNU) were 0.828 \u00b1 0.012, 0.885 \u00b1 0.006, 0.885 \u00b1 0.006, and 0.918 \u00b1 0.006 for steroids, antibiotics, antivirals, and antifungals, respectively. For multi-class classification, the mean top-1 and top-5 accuracies were 56.7 \u00b1 1.6% and 92.0 \u00b1 1.1% (Edinburgh) and 44.8 \u00b1 1.2% and 78.1 \u00b1 0.3% (SNU), respectively. With the assistance of our algorithm, the sensitivity and specificity of 47 clinicians (21 dermatologists and 26 dermatology residents) for malignancy prediction (SNU; 240 images) were improved by 12.1% (P < 0.0001) and 1.1% (P < 0.0001), respectively. The malignancy prediction sensitivity of 23 non-medical professionals was significantly increased by 83.8% (P < 0.0001). The top-1 and top-3 accuracies of four doctors in the multi-class classification of 134 diseases (SNU; 2,201 images) were increased by 7.0% (P\u00a0= 0.045) and 10.1% (P\u00a0= 0.0020), respectively. The results suggest that our algorithm may serve as augmented intelligence that can empower medical professionals in diagnostic dermatology.\n",
        "methods": "## METHODS\n\nA total of 224,181 clinical photographs were obtained with approvals of Asan Institutional Review Board (No 2017 0087) and Seoul National University Bundang Hospital Institutional Review Board (No. B-1802-451-005). In total, 220,680 images from four datasets (ASAN, Web, MED-NODE [Giotis et al., 2015], and Normal) were used for training the CNNs (Table 1). For the validation of our CNN models, two datasets (Edinburgh and SNU) were used. The Edinburgh and SNU datasets were used to represent the dataset consisting of Caucasian and Asian populations, respectively. The former consisted of 1,300 images of 10 tumorous skin diseases and is commercially available (https://licensing.eri.ed.ac.uk/i/software/dermofit-image-library.html), whereas the latter consisted of clinical photographs obtained in the Department of Dermatology at Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital. Out of 174 skin diseases used for training the model, 134 general skin disorders were selected for validation from the SNU dataset. The excluded 40 diseases were either a sign or rare diseases such as ulcer, purpura, dilated pore, and senile gluteal dermatosis. At least 10 images per disease were obtained in the SNU dataset. A total of 10 tumorous disorders in the SNU dataset contained at least 30 biopsy-confirmed images per class. In total, 2,201 images with either pathological confirmation (1,057 images) or clinical diagnosis identified on medical charts (1,144 images) were included in the SNU dataset.\n\nWith Berkeley Vision and Learning Center (BVLC) Caffe (Jia et al., 2014), we fine-tuned the ImageNet pretrained models of SENet (He et al., 2015, Hu et al., 2018), SE-ResNet-50 (He et al., 2015, Hu et al., 2018), and visual geometry group (VGG)-19 (Simonyan and Zisserman, 2014) (Supplementary Methods). The algorithm was trained with 174 disease classes and produced 174 outputs for a given test image. To appropriately address a specific binary classification in multi-class output models, we created the clusters of outputs that shared the same traits by using the disease categorization based on the classification listed in Supplementary Table S1, Category I:\n\n- malignancy output = a sum of model outputs for five malignant tumorous disorders (melanoma, basal cell carcinoma, squamous cell carcinoma, intraepithelial carcinoma, and keratoacanthoma)\n- benign output = a sum of model outputs for 40 benign tumorous disorders (seborrheic keratosis, wart, etc.)\n- steroids output = a sum of model outputs for 59 disorders requiring a steroids treatment (atopic dermatitis, contact dermatitis, etc.)\n- antibiotics output = a sum of model outputs for 21 disorders requiring an antibiotics treatment (cellulitis, impetigo, etc.)\n- antivirals output = a sum of model outputs for five disorders requiring an antivirals treatment (herpes simplex, herpes zoster, etc.)\n- antifungals output = a sum of model outputs for eight disorders requiring an antifungals treatment (tinea pedis, tinea versicolor, etc.)\n\nThese outputs were used for the analysis of binary classifications. For example, we used steroids output as a target-class output for the binary classification of steroid prediction. In the multi-class classification, we used the individual disease class output as a target-class output. In order to draw a receiver operating characteristic curve, a threshold for the target class was varied from 0 to 1 by a very small increment and the outputs that were greater than the threshold were considered as positive predictions.\n\nFor the binary classification of malignancy and treatment determination, a total of 240 images (40 images per each of the 6 groups) randomly selected from the SNU dataset were used for the human tests. The test images were categorized into one of the following six groups: malignant nodule, benign lesions requiring antifungals, those requiring antivirals, those requiring antibiotics, or those requiring steroids treatment, and other benign lesions. A total of 21 dermatologists, 26 residents, and 23 non-medical professionals participated in the reader tests. After the initial test, the test participants were informed of the results obtained by using the algorithm and the test was repeated.\n\nFor the multi-disease classification of 134 disorders, two dermatologists and two dermatology residents participated in the reader test, in which they were asked to provide their top-1 and 3 choices among 134 diseases for the 2,201 images from the SNU dataset. The list of 134 diseases was given to the test participants. Following the completion of the entire test set, the test participants were informed of the algorithms\u2019 top-1, 2, and 3 diagnoses for each test image and they repeated the entire test set.\n\nA two-tailed paired t-test was performed to determine if there were significant differences in the test performance before and after the assistance from the algorithm (Supplementary Table S4). We used R version 3.5.3 (pROC package version 1.14) for calculating the result of t-test and the AUC of receiver operating characteristic curve. The AUCs for each disease were calculated by converting the multi-class model outputs into a binary classification: one versus rest of the 173 diseases. In calculating top accuracy, if the top prediction of algorithm did not belong to the 134 classes in the validation, then the prediction was counted as incorrect. To calculate mean accuracy, a macro-average was computed independently for each class and then the average was taken.",
        "results": "## RESULTS\n\nBinary classification e malignancy prediction and treatment suggestion  \nUsing both the SNU dataset, which consisted of 2,201 images representing 134 diseases (5 malignancies and 129 nonmalignancies), and the Edinburgh dataset, which consisted of 1,300 images representing 10 disorders (four malignancies and six non-malignancies), the ability of our algorithm for malignancy diagnosis was validated in a situation that was representative of a real clinical practice, where clinicians are required to differentiate malignancy from several types of other skin diseases. The area under the curves (AUCs) for predicting malignancy were 0.937 \u00b1 0.004 (SNU) and 0.928 \u00b1 0.002 (Edinburgh), which were determined by using the malignancy output (Table 2). The test for predicting a treatment option among four primary medications (steroids, antibiotics, antivirals, and antifungals) using the SNU dataset demonstrated the potential of our algorithm to be applied to treatment suggestion with the mean AUCs of 0.893 \u00b1 0.006 (Table 2).  \nThe performance of our model for predicting malignancy and treatment option among 134 diseases was compared with those of 47 medical professionals (21 board-certified dermatologists and 26 dermatology residents) using randomly selected 240 images from the SNU dataset. Overall, in both the malignancy and treatment predictions, the algorithm showed the similar performance as that of the dermatology residents but slightly lower than those of the dermatologists (Figure 1).  \nThe efficacy of our algorithm to augment the diagnostic accuracy of test participants was evaluated by using the 240 images from the SNU dataset. After the initial test, the test participants were informed of the result of the algorithm for each test image and their answers were modified. For the detection of malignancy, the individual performance of 47 medical professionals assessed by F1 score showed a significant improvement from 0.66 \u00b1 0.08 to 0.75 \u00b1 0.06 (P < 0.0001) with the assistance of the model (Figure 1a). The sensitivity and specificity of the malignancy diagnosis of the 47 medical professionals improved from 77.4 \u00b1 10.7% to 86.8 \u00b1 8.7% (P < 0.0001) and from 92.9 \u00b1 2.4% to 93.9 \u00b1 2.3% (P < 0.0001), respectively. Similarly, the sensitivity of the malignancy diagnosis of 23 non-medical professionals improved markedly from 47.6 \u00b1 33.1% to 87.5 \u00b1 17.2% (P < 0.0001) without a loss in specificity. Similar to the malignancy test, the F1 score of the 47 medical professionals in predicting a primary treatment method significantly improved from 0.50 \u00b1 0.08 to 0.61 \u00b1 0.05 for steroids, from 0.47 \u00b1 0.10 to 0.55 \u00b1 0.08 for antibiotics, from 0.70 \u00b1 0.09 to 0.75 \u00b1 0.08 for antifungals, and from 0.48 \u00b1 0.11 to 0.54 \u00b1 0.09 for antivirals with the assistance of the algorithm (P < 0.0001 for four treatments, Figure 1bee).  \nIn the prediction of malignancy, medical professionals modified 5.1 \u00b1 2.7% of their answers after reviewing the result of the algorithm. A total of 74.0 \u00b1 15.0% of their modified answers became correct (26.0 \u00b1 15.0% became incorrect). The non-medical professionals modified 19.6 \u00b1 10.2% of their answers. As a result, 72.0 \u00b1 16.2% of their modified answers became correct, whereas 28.0 \u00b1 16.2% of them were incorrect. For the treatment prediction test, the medical professionals modified 8.8 \u00b1 4.9%, 7.3 \u00b1 4.2%, 4.0 \u00b1 2.4%, and 4.4 \u00b1 2.4% of their answers for predicting steroids, antibiotics, antivirals, and antifungals, respectively. As a result, 70.1 \u00b1 14.9% (steroids), 64.2 \u00b1 15.0% (antibiotics), 66.5 \u00b1 20.8% (antivirals), and 67.0 \u00b1 20.6% (antifungals) of the modified answers became correct. Overall, the diagnostic accuracy of the test participants for both the malignancy and treatment prediction improved with the assistance of the algorithm.  \n\nMulti-class classification of 134 diseases  \nUsing the SNU dataset (2,201 images), the mean top-1, 3, and 5 accuracies for the classification of 134 diseases were 44.8 \u00b1 1.2%, 69.0 \u00b1 0.9%, and 78.1 \u00b1 0.3%, respectively (Table 2). The classification performance for each of the 134 diseases are listed in Supplementary Table S1. The algorithm was able to differentiate between eczematous and infectious conditions as well as classify very rare skin lesions such as lichen amyloidosis (Figure 2f). For the same SNU dataset (2,201 images), the mean top-1 and 3 accuracies of four doctors (two dermatologists and two dermatology residents) were 49.9 \u00b1 7.0% and 67.2 \u00b1 5.4%, respectively (Supplementary Table S1).  \nWe also tested whether our algorithm could be used to augment the diagnostic accuracy of the test participants in the multi-class diagnosis task of 134 disorders by using the SNU dataset (2,201 images). The mean top-1 and 3 accuracies of four doctors showed 7.0 \u00b1 4.5% (P = 0.045) and 10.1 \u00b1 2.5% (P = 0.0020) improvements, respectively, with the reference of the top-1, 2, and 3 diagnoses predicted by the algorithm (Supplementary Table S1).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and\u00a0Predicting Treatment Options for 134 Skin\u00a0Disorders.",
        "abstract": "Although deep learning algorithms have demonstrated expert-level performance, previous efforts were mostly binary classifications of limited disorders. We trained an algorithm with 220,680 images of 174 disorders and validated it using Edinburgh (1,300 images; 10 disorders) and SNU datasets (2,201 images; 134 disorders). The algorithm could accurately predict malignancy, suggest primary treatment options, render multi-class classification among 134 disorders, and improve the performance of medical professionals. The area under the curves for malignancy detection were 0.928 \u00b1 0.002 (Edinburgh) and 0.937 \u00b1 0.004 (SNU). The area under the curves of primary treatment suggestion (SNU) were 0.828 \u00b1 0.012, 0.885 \u00b1 0.006, 0.885 \u00b1 0.006, and 0.918 \u00b1 0.006 for steroids, antibiotics, antivirals, and antifungals, respectively. For multi-class classification, the mean top-1 and top-5 accuracies were 56.7 \u00b1 1.6% and 92.0 \u00b1 1.1% (Edinburgh) and 44.8 \u00b1 1.2% and 78.1 \u00b1 0.3% (SNU), respectively. With the assistance of our algorithm, the sensitivity and specificity of 47 clinicians (21 dermatologists and 26 dermatology residents) for malignancy prediction (SNU; 240 images) were improved by 12.1% (P < 0.0001) and 1.1% (P < 0.0001), respectively. The malignancy prediction sensitivity of 23 non-medical professionals was significantly increased by 83.8% (P < 0.0001). The top-1 and top-3 accuracies of four doctors in the multi-class classification of 134 diseases (SNU; 2,201 images) were increased by 7.0% (P\u00a0= 0.045) and 10.1% (P\u00a0= 0.0020), respectively. The results suggest that our algorithm may serve as augmented intelligence that can empower medical professionals in diagnostic dermatology.\n",
        "methods": "## METHODS\n\nA total of 224,181 clinical photographs were obtained with approvals of Asan Institutional Review Board (No 2017 0087) and Seoul National University Bundang Hospital Institutional Review Board (No. B-1802-451-005). In total, 220,680 images from four datasets (ASAN, Web, MED-NODE [Giotis et al., 2015], and Normal) were used for training the CNNs (Table 1). For the validation of our CNN models, two datasets (Edinburgh and SNU) were used. The Edinburgh and SNU datasets were used to represent the dataset consisting of Caucasian and Asian populations, respectively. The former consisted of 1,300 images of 10 tumorous skin diseases and is commercially available (https://licensing.eri.ed.ac.uk/i/software/dermofit-image-library.html), whereas the latter consisted of clinical photographs obtained in the Department of Dermatology at Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital. Out of 174 skin diseases used for training the model, 134 general skin disorders were selected for validation from the SNU dataset. The excluded 40 diseases were either a sign or rare diseases such as ulcer, purpura, dilated pore, and senile gluteal dermatosis. At least 10 images per disease were obtained in the SNU dataset. A total of 10 tumorous disorders in the SNU dataset contained at least 30 biopsy-confirmed images per class. In total, 2,201 images with either pathological confirmation (1,057 images) or clinical diagnosis identified on medical charts (1,144 images) were included in the SNU dataset.\n\nWith Berkeley Vision and Learning Center (BVLC) Caffe (Jia et al., 2014), we fine-tuned the ImageNet pretrained models of SENet (He et al., 2015, Hu et al., 2018), SE-ResNet-50 (He et al., 2015, Hu et al., 2018), and visual geometry group (VGG)-19 (Simonyan and Zisserman, 2014) (Supplementary Methods). The algorithm was trained with 174 disease classes and produced 174 outputs for a given test image. To appropriately address a specific binary classification in multi-class output models, we created the clusters of outputs that shared the same traits by using the disease categorization based on the classification listed in Supplementary Table S1, Category I:\n\n- malignancy output = a sum of model outputs for five malignant tumorous disorders (melanoma, basal cell carcinoma, squamous cell carcinoma, intraepithelial carcinoma, and keratoacanthoma)\n- benign output = a sum of model outputs for 40 benign tumorous disorders (seborrheic keratosis, wart, etc.)\n- steroids output = a sum of model outputs for 59 disorders requiring a steroids treatment (atopic dermatitis, contact dermatitis, etc.)\n- antibiotics output = a sum of model outputs for 21 disorders requiring an antibiotics treatment (cellulitis, impetigo, etc.)\n- antivirals output = a sum of model outputs for five disorders requiring an antivirals treatment (herpes simplex, herpes zoster, etc.)\n- antifungals output = a sum of model outputs for eight disorders requiring an antifungals treatment (tinea pedis, tinea versicolor, etc.)\n\nThese outputs were used for the analysis of binary classifications. For example, we used steroids output as a target-class output for the binary classification of steroid prediction. In the multi-class classification, we used the individual disease class output as a target-class output. In order to draw a receiver operating characteristic curve, a threshold for the target class was varied from 0 to 1 by a very small increment and the outputs that were greater than the threshold were considered as positive predictions.\n\nFor the binary classification of malignancy and treatment determination, a total of 240 images (40 images per each of the 6 groups) randomly selected from the SNU dataset were used for the human tests. The test images were categorized into one of the following six groups: malignant nodule, benign lesions requiring antifungals, those requiring antivirals, those requiring antibiotics, or those requiring steroids treatment, and other benign lesions. A total of 21 dermatologists, 26 residents, and 23 non-medical professionals participated in the reader tests. After the initial test, the test participants were informed of the results obtained by using the algorithm and the test was repeated.\n\nFor the multi-disease classification of 134 disorders, two dermatologists and two dermatology residents participated in the reader test, in which they were asked to provide their top-1 and 3 choices among 134 diseases for the 2,201 images from the SNU dataset. The list of 134 diseases was given to the test participants. Following the completion of the entire test set, the test participants were informed of the algorithms\u2019 top-1, 2, and 3 diagnoses for each test image and they repeated the entire test set.\n\nA two-tailed paired t-test was performed to determine if there were significant differences in the test performance before and after the assistance from the algorithm (Supplementary Table S4). We used R version 3.5.3 (pROC package version 1.14) for calculating the result of t-test and the AUC of receiver operating characteristic curve. The AUCs for each disease were calculated by converting the multi-class model outputs into a binary classification: one versus rest of the 173 diseases. In calculating top accuracy, if the top prediction of algorithm did not belong to the 134 classes in the validation, then the prediction was counted as incorrect. To calculate mean accuracy, a macro-average was computed independently for each class and then the average was taken.",
        "results": "## RESULTS\n\nBinary classification e malignancy prediction and treatment suggestion  \nUsing both the SNU dataset, which consisted of 2,201 images representing 134 diseases (5 malignancies and 129 nonmalignancies), and the Edinburgh dataset, which consisted of 1,300 images representing 10 disorders (four malignancies and six non-malignancies), the ability of our algorithm for malignancy diagnosis was validated in a situation that was representative of a real clinical practice, where clinicians are required to differentiate malignancy from several types of other skin diseases. The area under the curves (AUCs) for predicting malignancy were 0.937 \u00b1 0.004 (SNU) and 0.928 \u00b1 0.002 (Edinburgh), which were determined by using the malignancy output (Table 2). The test for predicting a treatment option among four primary medications (steroids, antibiotics, antivirals, and antifungals) using the SNU dataset demonstrated the potential of our algorithm to be applied to treatment suggestion with the mean AUCs of 0.893 \u00b1 0.006 (Table 2).  \nThe performance of our model for predicting malignancy and treatment option among 134 diseases was compared with those of 47 medical professionals (21 board-certified dermatologists and 26 dermatology residents) using randomly selected 240 images from the SNU dataset. Overall, in both the malignancy and treatment predictions, the algorithm showed the similar performance as that of the dermatology residents but slightly lower than those of the dermatologists (Figure 1).  \nThe efficacy of our algorithm to augment the diagnostic accuracy of test participants was evaluated by using the 240 images from the SNU dataset. After the initial test, the test participants were informed of the result of the algorithm for each test image and their answers were modified. For the detection of malignancy, the individual performance of 47 medical professionals assessed by F1 score showed a significant improvement from 0.66 \u00b1 0.08 to 0.75 \u00b1 0.06 (P < 0.0001) with the assistance of the model (Figure 1a). The sensitivity and specificity of the malignancy diagnosis of the 47 medical professionals improved from 77.4 \u00b1 10.7% to 86.8 \u00b1 8.7% (P < 0.0001) and from 92.9 \u00b1 2.4% to 93.9 \u00b1 2.3% (P < 0.0001), respectively. Similarly, the sensitivity of the malignancy diagnosis of 23 non-medical professionals improved markedly from 47.6 \u00b1 33.1% to 87.5 \u00b1 17.2% (P < 0.0001) without a loss in specificity. Similar to the malignancy test, the F1 score of the 47 medical professionals in predicting a primary treatment method significantly improved from 0.50 \u00b1 0.08 to 0.61 \u00b1 0.05 for steroids, from 0.47 \u00b1 0.10 to 0.55 \u00b1 0.08 for antibiotics, from 0.70 \u00b1 0.09 to 0.75 \u00b1 0.08 for antifungals, and from 0.48 \u00b1 0.11 to 0.54 \u00b1 0.09 for antivirals with the assistance of the algorithm (P < 0.0001 for four treatments, Figure 1bee).  \nIn the prediction of malignancy, medical professionals modified 5.1 \u00b1 2.7% of their answers after reviewing the result of the algorithm. A total of 74.0 \u00b1 15.0% of their modified answers became correct (26.0 \u00b1 15.0% became incorrect). The non-medical professionals modified 19.6 \u00b1 10.2% of their answers. As a result, 72.0 \u00b1 16.2% of their modified answers became correct, whereas 28.0 \u00b1 16.2% of them were incorrect. For the treatment prediction test, the medical professionals modified 8.8 \u00b1 4.9%, 7.3 \u00b1 4.2%, 4.0 \u00b1 2.4%, and 4.4 \u00b1 2.4% of their answers for predicting steroids, antibiotics, antivirals, and antifungals, respectively. As a result, 70.1 \u00b1 14.9% (steroids), 64.2 \u00b1 15.0% (antibiotics), 66.5 \u00b1 20.8% (antivirals), and 67.0 \u00b1 20.6% (antifungals) of the modified answers became correct. Overall, the diagnostic accuracy of the test participants for both the malignancy and treatment prediction improved with the assistance of the algorithm.  \n\nMulti-class classification of 134 diseases  \nUsing the SNU dataset (2,201 images), the mean top-1, 3, and 5 accuracies for the classification of 134 diseases were 44.8 \u00b1 1.2%, 69.0 \u00b1 0.9%, and 78.1 \u00b1 0.3%, respectively (Table 2). The classification performance for each of the 134 diseases are listed in Supplementary Table S1. The algorithm was able to differentiate between eczematous and infectious conditions as well as classify very rare skin lesions such as lichen amyloidosis (Figure 2f). For the same SNU dataset (2,201 images), the mean top-1 and 3 accuracies of four doctors (two dermatologists and two dermatology residents) were 49.9 \u00b1 7.0% and 67.2 \u00b1 5.4%, respectively (Supplementary Table S1).  \nWe also tested whether our algorithm could be used to augment the diagnostic accuracy of the test participants in the multi-class diagnosis task of 134 disorders by using the SNU dataset (2,201 images). The mean top-1 and 3 accuracies of four doctors showed 7.0 \u00b1 4.5% (P = 0.045) and 10.1 \u00b1 2.5% (P = 0.0020) improvements, respectively, with the reference of the top-1, 2, and 3 diagnoses predicted by the algorithm (Supplementary Table S1).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and\u00a0Predicting Treatment Options for 134 Skin\u00a0Disorders.",
        "abstract": "Although deep learning algorithms have demonstrated expert-level performance, previous efforts were mostly binary classifications of limited disorders. We trained an algorithm with 220,680 images of 174 disorders and validated it using Edinburgh (1,300 images; 10 disorders) and SNU datasets (2,201 images; 134 disorders). The algorithm could accurately predict malignancy, suggest primary treatment options, render multi-class classification among 134 disorders, and improve the performance of medical professionals. The area under the curves for malignancy detection were 0.928 \u00b1 0.002 (Edinburgh) and 0.937 \u00b1 0.004 (SNU). The area under the curves of primary treatment suggestion (SNU) were 0.828 \u00b1 0.012, 0.885 \u00b1 0.006, 0.885 \u00b1 0.006, and 0.918 \u00b1 0.006 for steroids, antibiotics, antivirals, and antifungals, respectively. For multi-class classification, the mean top-1 and top-5 accuracies were 56.7 \u00b1 1.6% and 92.0 \u00b1 1.1% (Edinburgh) and 44.8 \u00b1 1.2% and 78.1 \u00b1 0.3% (SNU), respectively. With the assistance of our algorithm, the sensitivity and specificity of 47 clinicians (21 dermatologists and 26 dermatology residents) for malignancy prediction (SNU; 240 images) were improved by 12.1% (P < 0.0001) and 1.1% (P < 0.0001), respectively. The malignancy prediction sensitivity of 23 non-medical professionals was significantly increased by 83.8% (P < 0.0001). The top-1 and top-3 accuracies of four doctors in the multi-class classification of 134 diseases (SNU; 2,201 images) were increased by 7.0% (P\u00a0= 0.045) and 10.1% (P\u00a0= 0.0020), respectively. The results suggest that our algorithm may serve as augmented intelligence that can empower medical professionals in diagnostic dermatology.\n",
        "methods": "## METHODS\n\nA total of 224,181 clinical photographs were obtained with approvals of Asan Institutional Review Board (No 2017 0087) and Seoul National University Bundang Hospital Institutional Review Board (No. B-1802-451-005). In total, 220,680 images from four datasets (ASAN, Web, MED-NODE [Giotis et al., 2015], and Normal) were used for training the CNNs (Table 1). For the validation of our CNN models, two datasets (Edinburgh and SNU) were used. The Edinburgh and SNU datasets were used to represent the dataset consisting of Caucasian and Asian populations, respectively. The former consisted of 1,300 images of 10 tumorous skin diseases and is commercially available (https://licensing.eri.ed.ac.uk/i/software/dermofit-image-library.html), whereas the latter consisted of clinical photographs obtained in the Department of Dermatology at Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital. Out of 174 skin diseases used for training the model, 134 general skin disorders were selected for validation from the SNU dataset. The excluded 40 diseases were either a sign or rare diseases such as ulcer, purpura, dilated pore, and senile gluteal dermatosis. At least 10 images per disease were obtained in the SNU dataset. A total of 10 tumorous disorders in the SNU dataset contained at least 30 biopsy-confirmed images per class. In total, 2,201 images with either pathological confirmation (1,057 images) or clinical diagnosis identified on medical charts (1,144 images) were included in the SNU dataset.\n\nWith Berkeley Vision and Learning Center (BVLC) Caffe (Jia et al., 2014), we fine-tuned the ImageNet pretrained models of SENet (He et al., 2015, Hu et al., 2018), SE-ResNet-50 (He et al., 2015, Hu et al., 2018), and visual geometry group (VGG)-19 (Simonyan and Zisserman, 2014) (Supplementary Methods). The algorithm was trained with 174 disease classes and produced 174 outputs for a given test image. To appropriately address a specific binary classification in multi-class output models, we created the clusters of outputs that shared the same traits by using the disease categorization based on the classification listed in Supplementary Table S1, Category I:\n\n- malignancy output = a sum of model outputs for five malignant tumorous disorders (melanoma, basal cell carcinoma, squamous cell carcinoma, intraepithelial carcinoma, and keratoacanthoma)\n- benign output = a sum of model outputs for 40 benign tumorous disorders (seborrheic keratosis, wart, etc.)\n- steroids output = a sum of model outputs for 59 disorders requiring a steroids treatment (atopic dermatitis, contact dermatitis, etc.)\n- antibiotics output = a sum of model outputs for 21 disorders requiring an antibiotics treatment (cellulitis, impetigo, etc.)\n- antivirals output = a sum of model outputs for five disorders requiring an antivirals treatment (herpes simplex, herpes zoster, etc.)\n- antifungals output = a sum of model outputs for eight disorders requiring an antifungals treatment (tinea pedis, tinea versicolor, etc.)\n\nThese outputs were used for the analysis of binary classifications. For example, we used steroids output as a target-class output for the binary classification of steroid prediction. In the multi-class classification, we used the individual disease class output as a target-class output. In order to draw a receiver operating characteristic curve, a threshold for the target class was varied from 0 to 1 by a very small increment and the outputs that were greater than the threshold were considered as positive predictions.\n\nFor the binary classification of malignancy and treatment determination, a total of 240 images (40 images per each of the 6 groups) randomly selected from the SNU dataset were used for the human tests. The test images were categorized into one of the following six groups: malignant nodule, benign lesions requiring antifungals, those requiring antivirals, those requiring antibiotics, or those requiring steroids treatment, and other benign lesions. A total of 21 dermatologists, 26 residents, and 23 non-medical professionals participated in the reader tests. After the initial test, the test participants were informed of the results obtained by using the algorithm and the test was repeated.\n\nFor the multi-disease classification of 134 disorders, two dermatologists and two dermatology residents participated in the reader test, in which they were asked to provide their top-1 and 3 choices among 134 diseases for the 2,201 images from the SNU dataset. The list of 134 diseases was given to the test participants. Following the completion of the entire test set, the test participants were informed of the algorithms\u2019 top-1, 2, and 3 diagnoses for each test image and they repeated the entire test set.\n\nA two-tailed paired t-test was performed to determine if there were significant differences in the test performance before and after the assistance from the algorithm (Supplementary Table S4). We used R version 3.5.3 (pROC package version 1.14) for calculating the result of t-test and the AUC of receiver operating characteristic curve. The AUCs for each disease were calculated by converting the multi-class model outputs into a binary classification: one versus rest of the 173 diseases. In calculating top accuracy, if the top prediction of algorithm did not belong to the 134 classes in the validation, then the prediction was counted as incorrect. To calculate mean accuracy, a macro-average was computed independently for each class and then the average was taken.",
        "results": "## RESULTS\n\nBinary classification e malignancy prediction and treatment suggestion  \nUsing both the SNU dataset, which consisted of 2,201 images representing 134 diseases (5 malignancies and 129 nonmalignancies), and the Edinburgh dataset, which consisted of 1,300 images representing 10 disorders (four malignancies and six non-malignancies), the ability of our algorithm for malignancy diagnosis was validated in a situation that was representative of a real clinical practice, where clinicians are required to differentiate malignancy from several types of other skin diseases. The area under the curves (AUCs) for predicting malignancy were 0.937 \u00b1 0.004 (SNU) and 0.928 \u00b1 0.002 (Edinburgh), which were determined by using the malignancy output (Table 2). The test for predicting a treatment option among four primary medications (steroids, antibiotics, antivirals, and antifungals) using the SNU dataset demonstrated the potential of our algorithm to be applied to treatment suggestion with the mean AUCs of 0.893 \u00b1 0.006 (Table 2).  \nThe performance of our model for predicting malignancy and treatment option among 134 diseases was compared with those of 47 medical professionals (21 board-certified dermatologists and 26 dermatology residents) using randomly selected 240 images from the SNU dataset. Overall, in both the malignancy and treatment predictions, the algorithm showed the similar performance as that of the dermatology residents but slightly lower than those of the dermatologists (Figure 1).  \nThe efficacy of our algorithm to augment the diagnostic accuracy of test participants was evaluated by using the 240 images from the SNU dataset. After the initial test, the test participants were informed of the result of the algorithm for each test image and their answers were modified. For the detection of malignancy, the individual performance of 47 medical professionals assessed by F1 score showed a significant improvement from 0.66 \u00b1 0.08 to 0.75 \u00b1 0.06 (P < 0.0001) with the assistance of the model (Figure 1a). The sensitivity and specificity of the malignancy diagnosis of the 47 medical professionals improved from 77.4 \u00b1 10.7% to 86.8 \u00b1 8.7% (P < 0.0001) and from 92.9 \u00b1 2.4% to 93.9 \u00b1 2.3% (P < 0.0001), respectively. Similarly, the sensitivity of the malignancy diagnosis of 23 non-medical professionals improved markedly from 47.6 \u00b1 33.1% to 87.5 \u00b1 17.2% (P < 0.0001) without a loss in specificity. Similar to the malignancy test, the F1 score of the 47 medical professionals in predicting a primary treatment method significantly improved from 0.50 \u00b1 0.08 to 0.61 \u00b1 0.05 for steroids, from 0.47 \u00b1 0.10 to 0.55 \u00b1 0.08 for antibiotics, from 0.70 \u00b1 0.09 to 0.75 \u00b1 0.08 for antifungals, and from 0.48 \u00b1 0.11 to 0.54 \u00b1 0.09 for antivirals with the assistance of the algorithm (P < 0.0001 for four treatments, Figure 1bee).  \nIn the prediction of malignancy, medical professionals modified 5.1 \u00b1 2.7% of their answers after reviewing the result of the algorithm. A total of 74.0 \u00b1 15.0% of their modified answers became correct (26.0 \u00b1 15.0% became incorrect). The non-medical professionals modified 19.6 \u00b1 10.2% of their answers. As a result, 72.0 \u00b1 16.2% of their modified answers became correct, whereas 28.0 \u00b1 16.2% of them were incorrect. For the treatment prediction test, the medical professionals modified 8.8 \u00b1 4.9%, 7.3 \u00b1 4.2%, 4.0 \u00b1 2.4%, and 4.4 \u00b1 2.4% of their answers for predicting steroids, antibiotics, antivirals, and antifungals, respectively. As a result, 70.1 \u00b1 14.9% (steroids), 64.2 \u00b1 15.0% (antibiotics), 66.5 \u00b1 20.8% (antivirals), and 67.0 \u00b1 20.6% (antifungals) of the modified answers became correct. Overall, the diagnostic accuracy of the test participants for both the malignancy and treatment prediction improved with the assistance of the algorithm.  \n\nMulti-class classification of 134 diseases  \nUsing the SNU dataset (2,201 images), the mean top-1, 3, and 5 accuracies for the classification of 134 diseases were 44.8 \u00b1 1.2%, 69.0 \u00b1 0.9%, and 78.1 \u00b1 0.3%, respectively (Table 2). The classification performance for each of the 134 diseases are listed in Supplementary Table S1. The algorithm was able to differentiate between eczematous and infectious conditions as well as classify very rare skin lesions such as lichen amyloidosis (Figure 2f). For the same SNU dataset (2,201 images), the mean top-1 and 3 accuracies of four doctors (two dermatologists and two dermatology residents) were 49.9 \u00b1 7.0% and 67.2 \u00b1 5.4%, respectively (Supplementary Table S1).  \nWe also tested whether our algorithm could be used to augment the diagnostic accuracy of the test participants in the multi-class diagnosis task of 134 disorders by using the SNU dataset (2,201 images). The mean top-1 and 3 accuracies of four doctors showed 7.0 \u00b1 4.5% (P = 0.045) and 10.1 \u00b1 2.5% (P = 0.0020) improvements, respectively, with the reference of the top-1, 2, and 3 diagnoses predicted by the algorithm (Supplementary Table S1).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and\u00a0Predicting Treatment Options for 134 Skin\u00a0Disorders.",
        "abstract": "Although deep learning algorithms have demonstrated expert-level performance, previous efforts were mostly binary classifications of limited disorders. We trained an algorithm with 220,680 images of 174 disorders and validated it using Edinburgh (1,300 images; 10 disorders) and SNU datasets (2,201 images; 134 disorders). The algorithm could accurately predict malignancy, suggest primary treatment options, render multi-class classification among 134 disorders, and improve the performance of medical professionals. The area under the curves for malignancy detection were 0.928 \u00b1 0.002 (Edinburgh) and 0.937 \u00b1 0.004 (SNU). The area under the curves of primary treatment suggestion (SNU) were 0.828 \u00b1 0.012, 0.885 \u00b1 0.006, 0.885 \u00b1 0.006, and 0.918 \u00b1 0.006 for steroids, antibiotics, antivirals, and antifungals, respectively. For multi-class classification, the mean top-1 and top-5 accuracies were 56.7 \u00b1 1.6% and 92.0 \u00b1 1.1% (Edinburgh) and 44.8 \u00b1 1.2% and 78.1 \u00b1 0.3% (SNU), respectively. With the assistance of our algorithm, the sensitivity and specificity of 47 clinicians (21 dermatologists and 26 dermatology residents) for malignancy prediction (SNU; 240 images) were improved by 12.1% (P < 0.0001) and 1.1% (P < 0.0001), respectively. The malignancy prediction sensitivity of 23 non-medical professionals was significantly increased by 83.8% (P < 0.0001). The top-1 and top-3 accuracies of four doctors in the multi-class classification of 134 diseases (SNU; 2,201 images) were increased by 7.0% (P\u00a0= 0.045) and 10.1% (P\u00a0= 0.0020), respectively. The results suggest that our algorithm may serve as augmented intelligence that can empower medical professionals in diagnostic dermatology.\n",
        "methods": "## METHODS\n\nA total of 224,181 clinical photographs were obtained with approvals of Asan Institutional Review Board (No 2017 0087) and Seoul National University Bundang Hospital Institutional Review Board (No. B-1802-451-005). In total, 220,680 images from four datasets (ASAN, Web, MED-NODE [Giotis et al., 2015], and Normal) were used for training the CNNs (Table 1). For the validation of our CNN models, two datasets (Edinburgh and SNU) were used. The Edinburgh and SNU datasets were used to represent the dataset consisting of Caucasian and Asian populations, respectively. The former consisted of 1,300 images of 10 tumorous skin diseases and is commercially available (https://licensing.eri.ed.ac.uk/i/software/dermofit-image-library.html), whereas the latter consisted of clinical photographs obtained in the Department of Dermatology at Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital. Out of 174 skin diseases used for training the model, 134 general skin disorders were selected for validation from the SNU dataset. The excluded 40 diseases were either a sign or rare diseases such as ulcer, purpura, dilated pore, and senile gluteal dermatosis. At least 10 images per disease were obtained in the SNU dataset. A total of 10 tumorous disorders in the SNU dataset contained at least 30 biopsy-confirmed images per class. In total, 2,201 images with either pathological confirmation (1,057 images) or clinical diagnosis identified on medical charts (1,144 images) were included in the SNU dataset.\n\nWith Berkeley Vision and Learning Center (BVLC) Caffe (Jia et al., 2014), we fine-tuned the ImageNet pretrained models of SENet (He et al., 2015, Hu et al., 2018), SE-ResNet-50 (He et al., 2015, Hu et al., 2018), and visual geometry group (VGG)-19 (Simonyan and Zisserman, 2014) (Supplementary Methods). The algorithm was trained with 174 disease classes and produced 174 outputs for a given test image. To appropriately address a specific binary classification in multi-class output models, we created the clusters of outputs that shared the same traits by using the disease categorization based on the classification listed in Supplementary Table S1, Category I:\n\n- malignancy output = a sum of model outputs for five malignant tumorous disorders (melanoma, basal cell carcinoma, squamous cell carcinoma, intraepithelial carcinoma, and keratoacanthoma)\n- benign output = a sum of model outputs for 40 benign tumorous disorders (seborrheic keratosis, wart, etc.)\n- steroids output = a sum of model outputs for 59 disorders requiring a steroids treatment (atopic dermatitis, contact dermatitis, etc.)\n- antibiotics output = a sum of model outputs for 21 disorders requiring an antibiotics treatment (cellulitis, impetigo, etc.)\n- antivirals output = a sum of model outputs for five disorders requiring an antivirals treatment (herpes simplex, herpes zoster, etc.)\n- antifungals output = a sum of model outputs for eight disorders requiring an antifungals treatment (tinea pedis, tinea versicolor, etc.)\n\nThese outputs were used for the analysis of binary classifications. For example, we used steroids output as a target-class output for the binary classification of steroid prediction. In the multi-class classification, we used the individual disease class output as a target-class output. In order to draw a receiver operating characteristic curve, a threshold for the target class was varied from 0 to 1 by a very small increment and the outputs that were greater than the threshold were considered as positive predictions.\n\nFor the binary classification of malignancy and treatment determination, a total of 240 images (40 images per each of the 6 groups) randomly selected from the SNU dataset were used for the human tests. The test images were categorized into one of the following six groups: malignant nodule, benign lesions requiring antifungals, those requiring antivirals, those requiring antibiotics, or those requiring steroids treatment, and other benign lesions. A total of 21 dermatologists, 26 residents, and 23 non-medical professionals participated in the reader tests. After the initial test, the test participants were informed of the results obtained by using the algorithm and the test was repeated.\n\nFor the multi-disease classification of 134 disorders, two dermatologists and two dermatology residents participated in the reader test, in which they were asked to provide their top-1 and 3 choices among 134 diseases for the 2,201 images from the SNU dataset. The list of 134 diseases was given to the test participants. Following the completion of the entire test set, the test participants were informed of the algorithms\u2019 top-1, 2, and 3 diagnoses for each test image and they repeated the entire test set.\n\nA two-tailed paired t-test was performed to determine if there were significant differences in the test performance before and after the assistance from the algorithm (Supplementary Table S4). We used R version 3.5.3 (pROC package version 1.14) for calculating the result of t-test and the AUC of receiver operating characteristic curve. The AUCs for each disease were calculated by converting the multi-class model outputs into a binary classification: one versus rest of the 173 diseases. In calculating top accuracy, if the top prediction of algorithm did not belong to the 134 classes in the validation, then the prediction was counted as incorrect. To calculate mean accuracy, a macro-average was computed independently for each class and then the average was taken.",
        "results": "## RESULTS\n\nBinary classification e malignancy prediction and treatment suggestion  \nUsing both the SNU dataset, which consisted of 2,201 images representing 134 diseases (5 malignancies and 129 nonmalignancies), and the Edinburgh dataset, which consisted of 1,300 images representing 10 disorders (four malignancies and six non-malignancies), the ability of our algorithm for malignancy diagnosis was validated in a situation that was representative of a real clinical practice, where clinicians are required to differentiate malignancy from several types of other skin diseases. The area under the curves (AUCs) for predicting malignancy were 0.937 \u00b1 0.004 (SNU) and 0.928 \u00b1 0.002 (Edinburgh), which were determined by using the malignancy output (Table 2). The test for predicting a treatment option among four primary medications (steroids, antibiotics, antivirals, and antifungals) using the SNU dataset demonstrated the potential of our algorithm to be applied to treatment suggestion with the mean AUCs of 0.893 \u00b1 0.006 (Table 2).  \nThe performance of our model for predicting malignancy and treatment option among 134 diseases was compared with those of 47 medical professionals (21 board-certified dermatologists and 26 dermatology residents) using randomly selected 240 images from the SNU dataset. Overall, in both the malignancy and treatment predictions, the algorithm showed the similar performance as that of the dermatology residents but slightly lower than those of the dermatologists (Figure 1).  \nThe efficacy of our algorithm to augment the diagnostic accuracy of test participants was evaluated by using the 240 images from the SNU dataset. After the initial test, the test participants were informed of the result of the algorithm for each test image and their answers were modified. For the detection of malignancy, the individual performance of 47 medical professionals assessed by F1 score showed a significant improvement from 0.66 \u00b1 0.08 to 0.75 \u00b1 0.06 (P < 0.0001) with the assistance of the model (Figure 1a). The sensitivity and specificity of the malignancy diagnosis of the 47 medical professionals improved from 77.4 \u00b1 10.7% to 86.8 \u00b1 8.7% (P < 0.0001) and from 92.9 \u00b1 2.4% to 93.9 \u00b1 2.3% (P < 0.0001), respectively. Similarly, the sensitivity of the malignancy diagnosis of 23 non-medical professionals improved markedly from 47.6 \u00b1 33.1% to 87.5 \u00b1 17.2% (P < 0.0001) without a loss in specificity. Similar to the malignancy test, the F1 score of the 47 medical professionals in predicting a primary treatment method significantly improved from 0.50 \u00b1 0.08 to 0.61 \u00b1 0.05 for steroids, from 0.47 \u00b1 0.10 to 0.55 \u00b1 0.08 for antibiotics, from 0.70 \u00b1 0.09 to 0.75 \u00b1 0.08 for antifungals, and from 0.48 \u00b1 0.11 to 0.54 \u00b1 0.09 for antivirals with the assistance of the algorithm (P < 0.0001 for four treatments, Figure 1bee).  \nIn the prediction of malignancy, medical professionals modified 5.1 \u00b1 2.7% of their answers after reviewing the result of the algorithm. A total of 74.0 \u00b1 15.0% of their modified answers became correct (26.0 \u00b1 15.0% became incorrect). The non-medical professionals modified 19.6 \u00b1 10.2% of their answers. As a result, 72.0 \u00b1 16.2% of their modified answers became correct, whereas 28.0 \u00b1 16.2% of them were incorrect. For the treatment prediction test, the medical professionals modified 8.8 \u00b1 4.9%, 7.3 \u00b1 4.2%, 4.0 \u00b1 2.4%, and 4.4 \u00b1 2.4% of their answers for predicting steroids, antibiotics, antivirals, and antifungals, respectively. As a result, 70.1 \u00b1 14.9% (steroids), 64.2 \u00b1 15.0% (antibiotics), 66.5 \u00b1 20.8% (antivirals), and 67.0 \u00b1 20.6% (antifungals) of the modified answers became correct. Overall, the diagnostic accuracy of the test participants for both the malignancy and treatment prediction improved with the assistance of the algorithm.  \n\nMulti-class classification of 134 diseases  \nUsing the SNU dataset (2,201 images), the mean top-1, 3, and 5 accuracies for the classification of 134 diseases were 44.8 \u00b1 1.2%, 69.0 \u00b1 0.9%, and 78.1 \u00b1 0.3%, respectively (Table 2). The classification performance for each of the 134 diseases are listed in Supplementary Table S1. The algorithm was able to differentiate between eczematous and infectious conditions as well as classify very rare skin lesions such as lichen amyloidosis (Figure 2f). For the same SNU dataset (2,201 images), the mean top-1 and 3 accuracies of four doctors (two dermatologists and two dermatology residents) were 49.9 \u00b1 7.0% and 67.2 \u00b1 5.4%, respectively (Supplementary Table S1).  \nWe also tested whether our algorithm could be used to augment the diagnostic accuracy of the test participants in the multi-class diagnosis task of 134 disorders by using the SNU dataset (2,201 images). The mean top-1 and 3 accuracies of four doctors showed 7.0 \u00b1 4.5% (P = 0.045) and 10.1 \u00b1 2.5% (P = 0.0020) improvements, respectively, with the reference of the top-1, 2, and 3 diagnoses predicted by the algorithm (Supplementary Table S1).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and\u00a0Predicting Treatment Options for 134 Skin\u00a0Disorders.",
        "abstract": "Although deep learning algorithms have demonstrated expert-level performance, previous efforts were mostly binary classifications of limited disorders. We trained an algorithm with 220,680 images of 174 disorders and validated it using Edinburgh (1,300 images; 10 disorders) and SNU datasets (2,201 images; 134 disorders). The algorithm could accurately predict malignancy, suggest primary treatment options, render multi-class classification among 134 disorders, and improve the performance of medical professionals. The area under the curves for malignancy detection were 0.928 \u00b1 0.002 (Edinburgh) and 0.937 \u00b1 0.004 (SNU). The area under the curves of primary treatment suggestion (SNU) were 0.828 \u00b1 0.012, 0.885 \u00b1 0.006, 0.885 \u00b1 0.006, and 0.918 \u00b1 0.006 for steroids, antibiotics, antivirals, and antifungals, respectively. For multi-class classification, the mean top-1 and top-5 accuracies were 56.7 \u00b1 1.6% and 92.0 \u00b1 1.1% (Edinburgh) and 44.8 \u00b1 1.2% and 78.1 \u00b1 0.3% (SNU), respectively. With the assistance of our algorithm, the sensitivity and specificity of 47 clinicians (21 dermatologists and 26 dermatology residents) for malignancy prediction (SNU; 240 images) were improved by 12.1% (P < 0.0001) and 1.1% (P < 0.0001), respectively. The malignancy prediction sensitivity of 23 non-medical professionals was significantly increased by 83.8% (P < 0.0001). The top-1 and top-3 accuracies of four doctors in the multi-class classification of 134 diseases (SNU; 2,201 images) were increased by 7.0% (P\u00a0= 0.045) and 10.1% (P\u00a0= 0.0020), respectively. The results suggest that our algorithm may serve as augmented intelligence that can empower medical professionals in diagnostic dermatology.\n",
        "methods": "## METHODS\n\nA total of 224,181 clinical photographs were obtained with approvals of Asan Institutional Review Board (No 2017 0087) and Seoul National University Bundang Hospital Institutional Review Board (No. B-1802-451-005). In total, 220,680 images from four datasets (ASAN, Web, MED-NODE [Giotis et al., 2015], and Normal) were used for training the CNNs (Table 1). For the validation of our CNN models, two datasets (Edinburgh and SNU) were used. The Edinburgh and SNU datasets were used to represent the dataset consisting of Caucasian and Asian populations, respectively. The former consisted of 1,300 images of 10 tumorous skin diseases and is commercially available (https://licensing.eri.ed.ac.uk/i/software/dermofit-image-library.html), whereas the latter consisted of clinical photographs obtained in the Department of Dermatology at Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital. Out of 174 skin diseases used for training the model, 134 general skin disorders were selected for validation from the SNU dataset. The excluded 40 diseases were either a sign or rare diseases such as ulcer, purpura, dilated pore, and senile gluteal dermatosis. At least 10 images per disease were obtained in the SNU dataset. A total of 10 tumorous disorders in the SNU dataset contained at least 30 biopsy-confirmed images per class. In total, 2,201 images with either pathological confirmation (1,057 images) or clinical diagnosis identified on medical charts (1,144 images) were included in the SNU dataset.\n\nWith Berkeley Vision and Learning Center (BVLC) Caffe (Jia et al., 2014), we fine-tuned the ImageNet pretrained models of SENet (He et al., 2015, Hu et al., 2018), SE-ResNet-50 (He et al., 2015, Hu et al., 2018), and visual geometry group (VGG)-19 (Simonyan and Zisserman, 2014) (Supplementary Methods). The algorithm was trained with 174 disease classes and produced 174 outputs for a given test image. To appropriately address a specific binary classification in multi-class output models, we created the clusters of outputs that shared the same traits by using the disease categorization based on the classification listed in Supplementary Table S1, Category I:\n\n- malignancy output = a sum of model outputs for five malignant tumorous disorders (melanoma, basal cell carcinoma, squamous cell carcinoma, intraepithelial carcinoma, and keratoacanthoma)\n- benign output = a sum of model outputs for 40 benign tumorous disorders (seborrheic keratosis, wart, etc.)\n- steroids output = a sum of model outputs for 59 disorders requiring a steroids treatment (atopic dermatitis, contact dermatitis, etc.)\n- antibiotics output = a sum of model outputs for 21 disorders requiring an antibiotics treatment (cellulitis, impetigo, etc.)\n- antivirals output = a sum of model outputs for five disorders requiring an antivirals treatment (herpes simplex, herpes zoster, etc.)\n- antifungals output = a sum of model outputs for eight disorders requiring an antifungals treatment (tinea pedis, tinea versicolor, etc.)\n\nThese outputs were used for the analysis of binary classifications. For example, we used steroids output as a target-class output for the binary classification of steroid prediction. In the multi-class classification, we used the individual disease class output as a target-class output. In order to draw a receiver operating characteristic curve, a threshold for the target class was varied from 0 to 1 by a very small increment and the outputs that were greater than the threshold were considered as positive predictions.\n\nFor the binary classification of malignancy and treatment determination, a total of 240 images (40 images per each of the 6 groups) randomly selected from the SNU dataset were used for the human tests. The test images were categorized into one of the following six groups: malignant nodule, benign lesions requiring antifungals, those requiring antivirals, those requiring antibiotics, or those requiring steroids treatment, and other benign lesions. A total of 21 dermatologists, 26 residents, and 23 non-medical professionals participated in the reader tests. After the initial test, the test participants were informed of the results obtained by using the algorithm and the test was repeated.\n\nFor the multi-disease classification of 134 disorders, two dermatologists and two dermatology residents participated in the reader test, in which they were asked to provide their top-1 and 3 choices among 134 diseases for the 2,201 images from the SNU dataset. The list of 134 diseases was given to the test participants. Following the completion of the entire test set, the test participants were informed of the algorithms\u2019 top-1, 2, and 3 diagnoses for each test image and they repeated the entire test set.\n\nA two-tailed paired t-test was performed to determine if there were significant differences in the test performance before and after the assistance from the algorithm (Supplementary Table S4). We used R version 3.5.3 (pROC package version 1.14) for calculating the result of t-test and the AUC of receiver operating characteristic curve. The AUCs for each disease were calculated by converting the multi-class model outputs into a binary classification: one versus rest of the 173 diseases. In calculating top accuracy, if the top prediction of algorithm did not belong to the 134 classes in the validation, then the prediction was counted as incorrect. To calculate mean accuracy, a macro-average was computed independently for each class and then the average was taken.",
        "results": "## RESULTS\n\nBinary classification e malignancy prediction and treatment suggestion  \nUsing both the SNU dataset, which consisted of 2,201 images representing 134 diseases (5 malignancies and 129 nonmalignancies), and the Edinburgh dataset, which consisted of 1,300 images representing 10 disorders (four malignancies and six non-malignancies), the ability of our algorithm for malignancy diagnosis was validated in a situation that was representative of a real clinical practice, where clinicians are required to differentiate malignancy from several types of other skin diseases. The area under the curves (AUCs) for predicting malignancy were 0.937 \u00b1 0.004 (SNU) and 0.928 \u00b1 0.002 (Edinburgh), which were determined by using the malignancy output (Table 2). The test for predicting a treatment option among four primary medications (steroids, antibiotics, antivirals, and antifungals) using the SNU dataset demonstrated the potential of our algorithm to be applied to treatment suggestion with the mean AUCs of 0.893 \u00b1 0.006 (Table 2).  \nThe performance of our model for predicting malignancy and treatment option among 134 diseases was compared with those of 47 medical professionals (21 board-certified dermatologists and 26 dermatology residents) using randomly selected 240 images from the SNU dataset. Overall, in both the malignancy and treatment predictions, the algorithm showed the similar performance as that of the dermatology residents but slightly lower than those of the dermatologists (Figure 1).  \nThe efficacy of our algorithm to augment the diagnostic accuracy of test participants was evaluated by using the 240 images from the SNU dataset. After the initial test, the test participants were informed of the result of the algorithm for each test image and their answers were modified. For the detection of malignancy, the individual performance of 47 medical professionals assessed by F1 score showed a significant improvement from 0.66 \u00b1 0.08 to 0.75 \u00b1 0.06 (P < 0.0001) with the assistance of the model (Figure 1a). The sensitivity and specificity of the malignancy diagnosis of the 47 medical professionals improved from 77.4 \u00b1 10.7% to 86.8 \u00b1 8.7% (P < 0.0001) and from 92.9 \u00b1 2.4% to 93.9 \u00b1 2.3% (P < 0.0001), respectively. Similarly, the sensitivity of the malignancy diagnosis of 23 non-medical professionals improved markedly from 47.6 \u00b1 33.1% to 87.5 \u00b1 17.2% (P < 0.0001) without a loss in specificity. Similar to the malignancy test, the F1 score of the 47 medical professionals in predicting a primary treatment method significantly improved from 0.50 \u00b1 0.08 to 0.61 \u00b1 0.05 for steroids, from 0.47 \u00b1 0.10 to 0.55 \u00b1 0.08 for antibiotics, from 0.70 \u00b1 0.09 to 0.75 \u00b1 0.08 for antifungals, and from 0.48 \u00b1 0.11 to 0.54 \u00b1 0.09 for antivirals with the assistance of the algorithm (P < 0.0001 for four treatments, Figure 1bee).  \nIn the prediction of malignancy, medical professionals modified 5.1 \u00b1 2.7% of their answers after reviewing the result of the algorithm. A total of 74.0 \u00b1 15.0% of their modified answers became correct (26.0 \u00b1 15.0% became incorrect). The non-medical professionals modified 19.6 \u00b1 10.2% of their answers. As a result, 72.0 \u00b1 16.2% of their modified answers became correct, whereas 28.0 \u00b1 16.2% of them were incorrect. For the treatment prediction test, the medical professionals modified 8.8 \u00b1 4.9%, 7.3 \u00b1 4.2%, 4.0 \u00b1 2.4%, and 4.4 \u00b1 2.4% of their answers for predicting steroids, antibiotics, antivirals, and antifungals, respectively. As a result, 70.1 \u00b1 14.9% (steroids), 64.2 \u00b1 15.0% (antibiotics), 66.5 \u00b1 20.8% (antivirals), and 67.0 \u00b1 20.6% (antifungals) of the modified answers became correct. Overall, the diagnostic accuracy of the test participants for both the malignancy and treatment prediction improved with the assistance of the algorithm.  \n\nMulti-class classification of 134 diseases  \nUsing the SNU dataset (2,201 images), the mean top-1, 3, and 5 accuracies for the classification of 134 diseases were 44.8 \u00b1 1.2%, 69.0 \u00b1 0.9%, and 78.1 \u00b1 0.3%, respectively (Table 2). The classification performance for each of the 134 diseases are listed in Supplementary Table S1. The algorithm was able to differentiate between eczematous and infectious conditions as well as classify very rare skin lesions such as lichen amyloidosis (Figure 2f). For the same SNU dataset (2,201 images), the mean top-1 and 3 accuracies of four doctors (two dermatologists and two dermatology residents) were 49.9 \u00b1 7.0% and 67.2 \u00b1 5.4%, respectively (Supplementary Table S1).  \nWe also tested whether our algorithm could be used to augment the diagnostic accuracy of the test participants in the multi-class diagnosis task of 134 disorders by using the SNU dataset (2,201 images). The mean top-1 and 3 accuracies of four doctors showed 7.0 \u00b1 4.5% (P = 0.045) and 10.1 \u00b1 2.5% (P = 0.0020) improvements, respectively, with the reference of the top-1, 2, and 3 diagnoses predicted by the algorithm (Supplementary Table S1).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and\u00a0Predicting Treatment Options for 134 Skin\u00a0Disorders.",
        "abstract": "Although deep learning algorithms have demonstrated expert-level performance, previous efforts were mostly binary classifications of limited disorders. We trained an algorithm with 220,680 images of 174 disorders and validated it using Edinburgh (1,300 images; 10 disorders) and SNU datasets (2,201 images; 134 disorders). The algorithm could accurately predict malignancy, suggest primary treatment options, render multi-class classification among 134 disorders, and improve the performance of medical professionals. The area under the curves for malignancy detection were 0.928 \u00b1 0.002 (Edinburgh) and 0.937 \u00b1 0.004 (SNU). The area under the curves of primary treatment suggestion (SNU) were 0.828 \u00b1 0.012, 0.885 \u00b1 0.006, 0.885 \u00b1 0.006, and 0.918 \u00b1 0.006 for steroids, antibiotics, antivirals, and antifungals, respectively. For multi-class classification, the mean top-1 and top-5 accuracies were 56.7 \u00b1 1.6% and 92.0 \u00b1 1.1% (Edinburgh) and 44.8 \u00b1 1.2% and 78.1 \u00b1 0.3% (SNU), respectively. With the assistance of our algorithm, the sensitivity and specificity of 47 clinicians (21 dermatologists and 26 dermatology residents) for malignancy prediction (SNU; 240 images) were improved by 12.1% (P < 0.0001) and 1.1% (P < 0.0001), respectively. The malignancy prediction sensitivity of 23 non-medical professionals was significantly increased by 83.8% (P < 0.0001). The top-1 and top-3 accuracies of four doctors in the multi-class classification of 134 diseases (SNU; 2,201 images) were increased by 7.0% (P\u00a0= 0.045) and 10.1% (P\u00a0= 0.0020), respectively. The results suggest that our algorithm may serve as augmented intelligence that can empower medical professionals in diagnostic dermatology.\n",
        "methods": "## METHODS\n\nA total of 224,181 clinical photographs were obtained with approvals of Asan Institutional Review Board (No 2017 0087) and Seoul National University Bundang Hospital Institutional Review Board (No. B-1802-451-005). In total, 220,680 images from four datasets (ASAN, Web, MED-NODE [Giotis et al., 2015], and Normal) were used for training the CNNs (Table 1). For the validation of our CNN models, two datasets (Edinburgh and SNU) were used. The Edinburgh and SNU datasets were used to represent the dataset consisting of Caucasian and Asian populations, respectively. The former consisted of 1,300 images of 10 tumorous skin diseases and is commercially available (https://licensing.eri.ed.ac.uk/i/software/dermofit-image-library.html), whereas the latter consisted of clinical photographs obtained in the Department of Dermatology at Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital. Out of 174 skin diseases used for training the model, 134 general skin disorders were selected for validation from the SNU dataset. The excluded 40 diseases were either a sign or rare diseases such as ulcer, purpura, dilated pore, and senile gluteal dermatosis. At least 10 images per disease were obtained in the SNU dataset. A total of 10 tumorous disorders in the SNU dataset contained at least 30 biopsy-confirmed images per class. In total, 2,201 images with either pathological confirmation (1,057 images) or clinical diagnosis identified on medical charts (1,144 images) were included in the SNU dataset.\n\nWith Berkeley Vision and Learning Center (BVLC) Caffe (Jia et al., 2014), we fine-tuned the ImageNet pretrained models of SENet (He et al., 2015, Hu et al., 2018), SE-ResNet-50 (He et al., 2015, Hu et al., 2018), and visual geometry group (VGG)-19 (Simonyan and Zisserman, 2014) (Supplementary Methods). The algorithm was trained with 174 disease classes and produced 174 outputs for a given test image. To appropriately address a specific binary classification in multi-class output models, we created the clusters of outputs that shared the same traits by using the disease categorization based on the classification listed in Supplementary Table S1, Category I:\n\n- malignancy output = a sum of model outputs for five malignant tumorous disorders (melanoma, basal cell carcinoma, squamous cell carcinoma, intraepithelial carcinoma, and keratoacanthoma)\n- benign output = a sum of model outputs for 40 benign tumorous disorders (seborrheic keratosis, wart, etc.)\n- steroids output = a sum of model outputs for 59 disorders requiring a steroids treatment (atopic dermatitis, contact dermatitis, etc.)\n- antibiotics output = a sum of model outputs for 21 disorders requiring an antibiotics treatment (cellulitis, impetigo, etc.)\n- antivirals output = a sum of model outputs for five disorders requiring an antivirals treatment (herpes simplex, herpes zoster, etc.)\n- antifungals output = a sum of model outputs for eight disorders requiring an antifungals treatment (tinea pedis, tinea versicolor, etc.)\n\nThese outputs were used for the analysis of binary classifications. For example, we used steroids output as a target-class output for the binary classification of steroid prediction. In the multi-class classification, we used the individual disease class output as a target-class output. In order to draw a receiver operating characteristic curve, a threshold for the target class was varied from 0 to 1 by a very small increment and the outputs that were greater than the threshold were considered as positive predictions.\n\nFor the binary classification of malignancy and treatment determination, a total of 240 images (40 images per each of the 6 groups) randomly selected from the SNU dataset were used for the human tests. The test images were categorized into one of the following six groups: malignant nodule, benign lesions requiring antifungals, those requiring antivirals, those requiring antibiotics, or those requiring steroids treatment, and other benign lesions. A total of 21 dermatologists, 26 residents, and 23 non-medical professionals participated in the reader tests. After the initial test, the test participants were informed of the results obtained by using the algorithm and the test was repeated.\n\nFor the multi-disease classification of 134 disorders, two dermatologists and two dermatology residents participated in the reader test, in which they were asked to provide their top-1 and 3 choices among 134 diseases for the 2,201 images from the SNU dataset. The list of 134 diseases was given to the test participants. Following the completion of the entire test set, the test participants were informed of the algorithms\u2019 top-1, 2, and 3 diagnoses for each test image and they repeated the entire test set.\n\nA two-tailed paired t-test was performed to determine if there were significant differences in the test performance before and after the assistance from the algorithm (Supplementary Table S4). We used R version 3.5.3 (pROC package version 1.14) for calculating the result of t-test and the AUC of receiver operating characteristic curve. The AUCs for each disease were calculated by converting the multi-class model outputs into a binary classification: one versus rest of the 173 diseases. In calculating top accuracy, if the top prediction of algorithm did not belong to the 134 classes in the validation, then the prediction was counted as incorrect. To calculate mean accuracy, a macro-average was computed independently for each class and then the average was taken.",
        "results": "## RESULTS\n\nBinary classification e malignancy prediction and treatment suggestion  \nUsing both the SNU dataset, which consisted of 2,201 images representing 134 diseases (5 malignancies and 129 nonmalignancies), and the Edinburgh dataset, which consisted of 1,300 images representing 10 disorders (four malignancies and six non-malignancies), the ability of our algorithm for malignancy diagnosis was validated in a situation that was representative of a real clinical practice, where clinicians are required to differentiate malignancy from several types of other skin diseases. The area under the curves (AUCs) for predicting malignancy were 0.937 \u00b1 0.004 (SNU) and 0.928 \u00b1 0.002 (Edinburgh), which were determined by using the malignancy output (Table 2). The test for predicting a treatment option among four primary medications (steroids, antibiotics, antivirals, and antifungals) using the SNU dataset demonstrated the potential of our algorithm to be applied to treatment suggestion with the mean AUCs of 0.893 \u00b1 0.006 (Table 2).  \nThe performance of our model for predicting malignancy and treatment option among 134 diseases was compared with those of 47 medical professionals (21 board-certified dermatologists and 26 dermatology residents) using randomly selected 240 images from the SNU dataset. Overall, in both the malignancy and treatment predictions, the algorithm showed the similar performance as that of the dermatology residents but slightly lower than those of the dermatologists (Figure 1).  \nThe efficacy of our algorithm to augment the diagnostic accuracy of test participants was evaluated by using the 240 images from the SNU dataset. After the initial test, the test participants were informed of the result of the algorithm for each test image and their answers were modified. For the detection of malignancy, the individual performance of 47 medical professionals assessed by F1 score showed a significant improvement from 0.66 \u00b1 0.08 to 0.75 \u00b1 0.06 (P < 0.0001) with the assistance of the model (Figure 1a). The sensitivity and specificity of the malignancy diagnosis of the 47 medical professionals improved from 77.4 \u00b1 10.7% to 86.8 \u00b1 8.7% (P < 0.0001) and from 92.9 \u00b1 2.4% to 93.9 \u00b1 2.3% (P < 0.0001), respectively. Similarly, the sensitivity of the malignancy diagnosis of 23 non-medical professionals improved markedly from 47.6 \u00b1 33.1% to 87.5 \u00b1 17.2% (P < 0.0001) without a loss in specificity. Similar to the malignancy test, the F1 score of the 47 medical professionals in predicting a primary treatment method significantly improved from 0.50 \u00b1 0.08 to 0.61 \u00b1 0.05 for steroids, from 0.47 \u00b1 0.10 to 0.55 \u00b1 0.08 for antibiotics, from 0.70 \u00b1 0.09 to 0.75 \u00b1 0.08 for antifungals, and from 0.48 \u00b1 0.11 to 0.54 \u00b1 0.09 for antivirals with the assistance of the algorithm (P < 0.0001 for four treatments, Figure 1bee).  \nIn the prediction of malignancy, medical professionals modified 5.1 \u00b1 2.7% of their answers after reviewing the result of the algorithm. A total of 74.0 \u00b1 15.0% of their modified answers became correct (26.0 \u00b1 15.0% became incorrect). The non-medical professionals modified 19.6 \u00b1 10.2% of their answers. As a result, 72.0 \u00b1 16.2% of their modified answers became correct, whereas 28.0 \u00b1 16.2% of them were incorrect. For the treatment prediction test, the medical professionals modified 8.8 \u00b1 4.9%, 7.3 \u00b1 4.2%, 4.0 \u00b1 2.4%, and 4.4 \u00b1 2.4% of their answers for predicting steroids, antibiotics, antivirals, and antifungals, respectively. As a result, 70.1 \u00b1 14.9% (steroids), 64.2 \u00b1 15.0% (antibiotics), 66.5 \u00b1 20.8% (antivirals), and 67.0 \u00b1 20.6% (antifungals) of the modified answers became correct. Overall, the diagnostic accuracy of the test participants for both the malignancy and treatment prediction improved with the assistance of the algorithm.  \n\nMulti-class classification of 134 diseases  \nUsing the SNU dataset (2,201 images), the mean top-1, 3, and 5 accuracies for the classification of 134 diseases were 44.8 \u00b1 1.2%, 69.0 \u00b1 0.9%, and 78.1 \u00b1 0.3%, respectively (Table 2). The classification performance for each of the 134 diseases are listed in Supplementary Table S1. The algorithm was able to differentiate between eczematous and infectious conditions as well as classify very rare skin lesions such as lichen amyloidosis (Figure 2f). For the same SNU dataset (2,201 images), the mean top-1 and 3 accuracies of four doctors (two dermatologists and two dermatology residents) were 49.9 \u00b1 7.0% and 67.2 \u00b1 5.4%, respectively (Supplementary Table S1).  \nWe also tested whether our algorithm could be used to augment the diagnostic accuracy of the test participants in the multi-class diagnosis task of 134 disorders by using the SNU dataset (2,201 images). The mean top-1 and 3 accuracies of four doctors showed 7.0 \u00b1 4.5% (P = 0.045) and 10.1 \u00b1 2.5% (P = 0.0020) improvements, respectively, with the reference of the top-1, 2, and 3 diagnoses predicted by the algorithm (Supplementary Table S1).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and\u00a0Predicting Treatment Options for 134 Skin\u00a0Disorders.",
        "abstract": "Although deep learning algorithms have demonstrated expert-level performance, previous efforts were mostly binary classifications of limited disorders. We trained an algorithm with 220,680 images of 174 disorders and validated it using Edinburgh (1,300 images; 10 disorders) and SNU datasets (2,201 images; 134 disorders). The algorithm could accurately predict malignancy, suggest primary treatment options, render multi-class classification among 134 disorders, and improve the performance of medical professionals. The area under the curves for malignancy detection were 0.928 \u00b1 0.002 (Edinburgh) and 0.937 \u00b1 0.004 (SNU). The area under the curves of primary treatment suggestion (SNU) were 0.828 \u00b1 0.012, 0.885 \u00b1 0.006, 0.885 \u00b1 0.006, and 0.918 \u00b1 0.006 for steroids, antibiotics, antivirals, and antifungals, respectively. For multi-class classification, the mean top-1 and top-5 accuracies were 56.7 \u00b1 1.6% and 92.0 \u00b1 1.1% (Edinburgh) and 44.8 \u00b1 1.2% and 78.1 \u00b1 0.3% (SNU), respectively. With the assistance of our algorithm, the sensitivity and specificity of 47 clinicians (21 dermatologists and 26 dermatology residents) for malignancy prediction (SNU; 240 images) were improved by 12.1% (P < 0.0001) and 1.1% (P < 0.0001), respectively. The malignancy prediction sensitivity of 23 non-medical professionals was significantly increased by 83.8% (P < 0.0001). The top-1 and top-3 accuracies of four doctors in the multi-class classification of 134 diseases (SNU; 2,201 images) were increased by 7.0% (P\u00a0= 0.045) and 10.1% (P\u00a0= 0.0020), respectively. The results suggest that our algorithm may serve as augmented intelligence that can empower medical professionals in diagnostic dermatology.\n",
        "methods": "## METHODS\n\nA total of 224,181 clinical photographs were obtained with approvals of Asan Institutional Review Board (No 2017 0087) and Seoul National University Bundang Hospital Institutional Review Board (No. B-1802-451-005). In total, 220,680 images from four datasets (ASAN, Web, MED-NODE [Giotis et al., 2015], and Normal) were used for training the CNNs (Table 1). For the validation of our CNN models, two datasets (Edinburgh and SNU) were used. The Edinburgh and SNU datasets were used to represent the dataset consisting of Caucasian and Asian populations, respectively. The former consisted of 1,300 images of 10 tumorous skin diseases and is commercially available (https://licensing.eri.ed.ac.uk/i/software/dermofit-image-library.html), whereas the latter consisted of clinical photographs obtained in the Department of Dermatology at Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital. Out of 174 skin diseases used for training the model, 134 general skin disorders were selected for validation from the SNU dataset. The excluded 40 diseases were either a sign or rare diseases such as ulcer, purpura, dilated pore, and senile gluteal dermatosis. At least 10 images per disease were obtained in the SNU dataset. A total of 10 tumorous disorders in the SNU dataset contained at least 30 biopsy-confirmed images per class. In total, 2,201 images with either pathological confirmation (1,057 images) or clinical diagnosis identified on medical charts (1,144 images) were included in the SNU dataset.\n\nWith Berkeley Vision and Learning Center (BVLC) Caffe (Jia et al., 2014), we fine-tuned the ImageNet pretrained models of SENet (He et al., 2015, Hu et al., 2018), SE-ResNet-50 (He et al., 2015, Hu et al., 2018), and visual geometry group (VGG)-19 (Simonyan and Zisserman, 2014) (Supplementary Methods). The algorithm was trained with 174 disease classes and produced 174 outputs for a given test image. To appropriately address a specific binary classification in multi-class output models, we created the clusters of outputs that shared the same traits by using the disease categorization based on the classification listed in Supplementary Table S1, Category I:\n\n- malignancy output = a sum of model outputs for five malignant tumorous disorders (melanoma, basal cell carcinoma, squamous cell carcinoma, intraepithelial carcinoma, and keratoacanthoma)\n- benign output = a sum of model outputs for 40 benign tumorous disorders (seborrheic keratosis, wart, etc.)\n- steroids output = a sum of model outputs for 59 disorders requiring a steroids treatment (atopic dermatitis, contact dermatitis, etc.)\n- antibiotics output = a sum of model outputs for 21 disorders requiring an antibiotics treatment (cellulitis, impetigo, etc.)\n- antivirals output = a sum of model outputs for five disorders requiring an antivirals treatment (herpes simplex, herpes zoster, etc.)\n- antifungals output = a sum of model outputs for eight disorders requiring an antifungals treatment (tinea pedis, tinea versicolor, etc.)\n\nThese outputs were used for the analysis of binary classifications. For example, we used steroids output as a target-class output for the binary classification of steroid prediction. In the multi-class classification, we used the individual disease class output as a target-class output. In order to draw a receiver operating characteristic curve, a threshold for the target class was varied from 0 to 1 by a very small increment and the outputs that were greater than the threshold were considered as positive predictions.\n\nFor the binary classification of malignancy and treatment determination, a total of 240 images (40 images per each of the 6 groups) randomly selected from the SNU dataset were used for the human tests. The test images were categorized into one of the following six groups: malignant nodule, benign lesions requiring antifungals, those requiring antivirals, those requiring antibiotics, or those requiring steroids treatment, and other benign lesions. A total of 21 dermatologists, 26 residents, and 23 non-medical professionals participated in the reader tests. After the initial test, the test participants were informed of the results obtained by using the algorithm and the test was repeated.\n\nFor the multi-disease classification of 134 disorders, two dermatologists and two dermatology residents participated in the reader test, in which they were asked to provide their top-1 and 3 choices among 134 diseases for the 2,201 images from the SNU dataset. The list of 134 diseases was given to the test participants. Following the completion of the entire test set, the test participants were informed of the algorithms\u2019 top-1, 2, and 3 diagnoses for each test image and they repeated the entire test set.\n\nA two-tailed paired t-test was performed to determine if there were significant differences in the test performance before and after the assistance from the algorithm (Supplementary Table S4). We used R version 3.5.3 (pROC package version 1.14) for calculating the result of t-test and the AUC of receiver operating characteristic curve. The AUCs for each disease were calculated by converting the multi-class model outputs into a binary classification: one versus rest of the 173 diseases. In calculating top accuracy, if the top prediction of algorithm did not belong to the 134 classes in the validation, then the prediction was counted as incorrect. To calculate mean accuracy, a macro-average was computed independently for each class and then the average was taken.",
        "results": "## RESULTS\n\nBinary classification e malignancy prediction and treatment suggestion  \nUsing both the SNU dataset, which consisted of 2,201 images representing 134 diseases (5 malignancies and 129 nonmalignancies), and the Edinburgh dataset, which consisted of 1,300 images representing 10 disorders (four malignancies and six non-malignancies), the ability of our algorithm for malignancy diagnosis was validated in a situation that was representative of a real clinical practice, where clinicians are required to differentiate malignancy from several types of other skin diseases. The area under the curves (AUCs) for predicting malignancy were 0.937 \u00b1 0.004 (SNU) and 0.928 \u00b1 0.002 (Edinburgh), which were determined by using the malignancy output (Table 2). The test for predicting a treatment option among four primary medications (steroids, antibiotics, antivirals, and antifungals) using the SNU dataset demonstrated the potential of our algorithm to be applied to treatment suggestion with the mean AUCs of 0.893 \u00b1 0.006 (Table 2).  \nThe performance of our model for predicting malignancy and treatment option among 134 diseases was compared with those of 47 medical professionals (21 board-certified dermatologists and 26 dermatology residents) using randomly selected 240 images from the SNU dataset. Overall, in both the malignancy and treatment predictions, the algorithm showed the similar performance as that of the dermatology residents but slightly lower than those of the dermatologists (Figure 1).  \nThe efficacy of our algorithm to augment the diagnostic accuracy of test participants was evaluated by using the 240 images from the SNU dataset. After the initial test, the test participants were informed of the result of the algorithm for each test image and their answers were modified. For the detection of malignancy, the individual performance of 47 medical professionals assessed by F1 score showed a significant improvement from 0.66 \u00b1 0.08 to 0.75 \u00b1 0.06 (P < 0.0001) with the assistance of the model (Figure 1a). The sensitivity and specificity of the malignancy diagnosis of the 47 medical professionals improved from 77.4 \u00b1 10.7% to 86.8 \u00b1 8.7% (P < 0.0001) and from 92.9 \u00b1 2.4% to 93.9 \u00b1 2.3% (P < 0.0001), respectively. Similarly, the sensitivity of the malignancy diagnosis of 23 non-medical professionals improved markedly from 47.6 \u00b1 33.1% to 87.5 \u00b1 17.2% (P < 0.0001) without a loss in specificity. Similar to the malignancy test, the F1 score of the 47 medical professionals in predicting a primary treatment method significantly improved from 0.50 \u00b1 0.08 to 0.61 \u00b1 0.05 for steroids, from 0.47 \u00b1 0.10 to 0.55 \u00b1 0.08 for antibiotics, from 0.70 \u00b1 0.09 to 0.75 \u00b1 0.08 for antifungals, and from 0.48 \u00b1 0.11 to 0.54 \u00b1 0.09 for antivirals with the assistance of the algorithm (P < 0.0001 for four treatments, Figure 1bee).  \nIn the prediction of malignancy, medical professionals modified 5.1 \u00b1 2.7% of their answers after reviewing the result of the algorithm. A total of 74.0 \u00b1 15.0% of their modified answers became correct (26.0 \u00b1 15.0% became incorrect). The non-medical professionals modified 19.6 \u00b1 10.2% of their answers. As a result, 72.0 \u00b1 16.2% of their modified answers became correct, whereas 28.0 \u00b1 16.2% of them were incorrect. For the treatment prediction test, the medical professionals modified 8.8 \u00b1 4.9%, 7.3 \u00b1 4.2%, 4.0 \u00b1 2.4%, and 4.4 \u00b1 2.4% of their answers for predicting steroids, antibiotics, antivirals, and antifungals, respectively. As a result, 70.1 \u00b1 14.9% (steroids), 64.2 \u00b1 15.0% (antibiotics), 66.5 \u00b1 20.8% (antivirals), and 67.0 \u00b1 20.6% (antifungals) of the modified answers became correct. Overall, the diagnostic accuracy of the test participants for both the malignancy and treatment prediction improved with the assistance of the algorithm.  \n\nMulti-class classification of 134 diseases  \nUsing the SNU dataset (2,201 images), the mean top-1, 3, and 5 accuracies for the classification of 134 diseases were 44.8 \u00b1 1.2%, 69.0 \u00b1 0.9%, and 78.1 \u00b1 0.3%, respectively (Table 2). The classification performance for each of the 134 diseases are listed in Supplementary Table S1. The algorithm was able to differentiate between eczematous and infectious conditions as well as classify very rare skin lesions such as lichen amyloidosis (Figure 2f). For the same SNU dataset (2,201 images), the mean top-1 and 3 accuracies of four doctors (two dermatologists and two dermatology residents) were 49.9 \u00b1 7.0% and 67.2 \u00b1 5.4%, respectively (Supplementary Table S1).  \nWe also tested whether our algorithm could be used to augment the diagnostic accuracy of the test participants in the multi-class diagnosis task of 134 disorders by using the SNU dataset (2,201 images). The mean top-1 and 3 accuracies of four doctors showed 7.0 \u00b1 4.5% (P = 0.045) and 10.1 \u00b1 2.5% (P = 0.0020) improvements, respectively, with the reference of the top-1, 2, and 3 diagnoses predicted by the algorithm (Supplementary Table S1).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Neoplasms: A Single-Center, Paralleled, Unmasked, Randomized Controlled Trial.",
        "abstract": "## TRIAL DESIGN\nThis was a single-center, unmasked, paralleled, randomized controlled trial.\n## METHODS\nA randomized trial was conducted in a tertiary care institute in South Korea to validate whether artificial intelligence (AI) could augment the accuracy of nonexpert physicians in the real-world settings, which included diverse out-of-distribution conditions. Consecutive patients aged >19 years, having one or more skin lesions suspicious for skin cancer detected by either the patient or physician, were randomly allocated to four nondermatology trainees and four dermatology residents. The attending dermatologists examined the randomly allocated patients with (AI-assisted group) or without (unaided group) the real-time assistance of AI algorithm (https://b2020.modelderm.com#world; convolutional neural networks; unmasked design) after simple randomization of the patients.\n## RESULTS\nUsing 576 consecutive cases (Fitzpatrick skin phototypes III or IV) with suspicious lesions out of the initial 603 recruitments, the accuracy of the AI-assisted group (n\u00a0= 295, 53.9%) was found to be significantly higher than those of the unaided group (n\u00a0= 281, 43.8%; P\u00a0= 0.019). Whereas the augmentation was more significant from 54.7% (n\u00a0= 150) to 30.7% (n\u00a0= 138; P < 0.0001) in the nondermatology trainees who had the least experience in dermatology, it was not significant in the dermatology residents. The algorithm could help trainees in the AI-assisted group include more differential diagnoses than the unaided group (2.09 vs. 1.95 diagnoses; P\u00a0= 0.0005). However, a 12.2% drop in Top-1 accuracy of the trainees was observed in cases in which all Top-3 predictions given by the algorithm were incorrect.\n## CONCLUSIONS\nThe multiclass AI algorithm augmented the diagnostic accuracy of nonexpert physicians in dermatology.\n",
        "methods": "### METHODS\n\nThis prospective study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea) (S2018-1703-0001). It was performed in the Department of Dermatology at Asan Medical Center, a tertiary care center in Korea. The study was conducted from 30 November 2020 to 9 September 2021 after online registration (KCT0005614; cris.nih.go.kr). The development of the algorithm (Model Dermatology, Build2020; https://b2020.modelderm.com#world) is described in the Supplementary Materials, and the algorithm was fixed on 19 September 2020. Along with the prediction of five differential diagnoses, the algorithm reports a malignancy score (range = 0-100). The malignancy score was defined as the sum of malignant outputs and 0.2 \u00d7 premalignant outputs as used previously (Han et al., 2020b). Using the subset of the Seoul National University dataset (240 images; https://doi.org/10.6084/m9.figshare.6454973), the high-sensitivity threshold for determining malignancy was defined as the threshold at which 90% sensitivity was obtained because the sensitivity of the attending dermatologists was at the level of 88.1% (Han et al., 2020a). The high-specificity threshold was defined as the threshold at which 80% sensitivity was obtained.\n\nIn a previous pilot study (Kim et al., 2022), the Top-1 accuracy of trainees was 47.9%. If 25% enhancement after the assistance were regarded as significant, the sample size was calculated as 548 (alpha = 0.05 and power = 0.8), and we planned to recruit 600 cases (Rosner, 2011).\n\nAll patients signed an informed consent before inclusion in the study. We included consecutive patients (aged >19 years) who had one or more skin lesions suspicious for skin cancer, detected by either the patient or physician. Exclusion criteria of patients and input data included patient refusal (10 cases), wrong recruitment (6 cases; aged \u226419 years), biopsy refusal (2 cases), and nonreal-time analysis (9 cases). Broken blindness and disclosure of the biopsy results in the referral note were also in the exclusion criteria, but there was no such case, and there was no performance error for the loss of internet connection or other technical issues. Formal pathologic diagnosis (504 cases) was used as the ground truth; however, if the pathologic report consisted of a pathologic description only (i.e., lichenoid reaction), the pathologic diagnosis was determined by clinicopathological correlation (20 cases). Clinical diagnosis of the attending dermatologists was used as the ground truth for the 52 cases where biopsy was not performed because the attending dermatologists decided not to biopsy the definitely benign cases. Ultimately, 524 biopsy-proven cases and 52 clinically-diagnosed cases were included in the final analysis among the 603 cases of the initial recruitment (Table 1). As shown in Supplementary Tables S6 and S7, a total of 53 conditions were within the trained 178 classes (in-distribution) and 29 conditions were not trained by the algorithm (out-of-distribution). There were no cases in which train-test contamination could be concerned, although there were 15 cases with photographs taken previously.\n\nA total of four attending physicians (3, 4, 6, and 22 years of experience after board certification), four first-year dermatology residents, and four nondermatology trainees (doctors in their first year after getting a medical license who rotate between various departments) participated in this study. Attending physicians routinely recorded their impressions after thorough examinations. After the simple randomization using a custom randomizer (rand function of the PHP language) by the attending dermatologists, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and recorded their diagnostic hypothesis in real time.\n\nIn the AI group, trainees captured and selected 1-3 photographs with age and sex metadata as an input data and uploaded them to http://b2020.modelderm.com#world using internet browsers. Then the after-diagnoses was recorded, referring to the five diagnoses suggested by the algorithm and the malignancy score (Supplementary Figure S2). The photographs that the trainees thought to be of adequate quality and composition by themselves were uploaded. Trainees were instructed so that only macroimages having the large lesion of interest at the center are uploaded. In the unaided group, trainees performed routine examinations and recorded the three most probable diagnoses, without the assistance of the algorithm. The use of dermoscopy was not allowed for all trainees.\n\nStatistical analysis\nIn calculating the Top accuracy, only the exact diagnosis was recorded as correct, but giving a specific subtype of the disease was also counted to be correct. For example, intradermal nevus was counted correct for the ground truth of junctional nevus. We lumped together 364 diagnoses in natural language into the 82 diagnosis codes (https://doi.org/10.6084/m9.figshare.16640257). For evaluating a malignancy prediction, the physicians diagnoses were transformed into either malignant or benign. Top accuracies, sensitivities, and specificities were compared using Pearson\u2019s chi-square test with Yates\u2019 continuity correction (AI group vs. unaided group) or McNemar test (before vs. after the assistance of the algorithm) using R version 4.1.1, and P < 0.05 was statistically significant.\n\nData availability statement\nThe raw data of this study is available at https://doi.org/10.6084/m9.figshare.16640257. The diagnosis in natural language, converted diagnosis code (ALIAS sheet), the cases determined by clinicopathological correlation, accuracy/sensitivity/specificity/positive predictive value/negative predictive value of the participants and algorithm, and demographics of subjects are listed. The algorithm is accessible at https://b2020.modelderm.com/#world or https://modelderm.com (Build2020) for academic purposes, and only the images with the permission of the original owner should be uploaded.",
        "results": "### RESULTS\nThis study is reported per the Consolidated Standards of Reporting Trials\u2014Artificial Intelligence guidelines (Liu et al., 2020). Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u2019correct\u2019.\n\nAI group versus unaided group\nTo evaluate that the two groups were truly comparable, accuracies of the attending dermatologists and trainees (before interventions) were compared. Top-1 accuracy of attending dermatologists (62.3%) and trainees (43.8%) of the unaided group were higher than those of the AI group (dermatologists \u00bc 60.3%, trainees \u00bc 41.0%), which indicated that easier cases were not disproportionately allocated to the AI group (Table 1 and Figure 1). Overall, the Top-1 accuracy of the AI group was 53.9% and that of the unaided was 43.8% (all trainees; chi-square test, P \u00bc 0.019; Figure 2 and Table 2). There were significant differences in the result depending on whether the participant was a nondermatology trainee (general practitioner [GP]) or a dermatology resident (resident). The Top-1 accuracy of the AIGP group (54.7%) was markedly higher than that of unaidedGP group (29.7%; chi-square test, P < 0.0001), whereas the Top-1 accuracy of the AIresident group and the unaidedresident group was 53.1% and 57.3%, respectively (chi-square test, P \u00bc 0.55). In the AI group, we compared the judgment before and after receiving assistance of the algorithm and there was a significant enhancement in the Top 1 accuracy of the AIGP group (before augmentation \u00bc 30.7%, after augmentation \u00bc 54.7%; McNemar test, P < 0.0001). However, in the AIresident group, the change was not significant (before augmentation \u00bc 51.7%, after augmentation \u00bc 53.1%; McNemar test, P \u00bc 0.86). As shown in Supplementary Table S1, a greater improvement in accuracy was observed for the subsequent cases, up to \u00fe25.0% and \u00fe37.5% for Top-1 and Top-3 accuracy, respectively, although the accuracy for Top-1 and Top-3 was improved by only \u00fe0.0% and \u00fe15.0%, respectively, for the first five cases. When the analysis was restricted to 266 cases that were biopsied, the Top-1/Top-3 accuracy of the standalone algorithm, trainees before augmentation, trainees after augmentation, and attending dermatologists were 51.5%/75.9%, 40.2%/49.2%, 56.0%/71.4%, and 56.0%/66.2%, respectively. The accuracies of the AI-augmented trainees were equivalent to those of the attending dermatologists. In the 258 biopsy-proven cases of the unaided group, the Top-1/Top-3 accuracy of trainees and attending dermatologists were 42.6%/57.4% and 58.9%/68.6%, respectively. Malignancy determination affects clinical decisions such as ordering a biopsy if there is any malignancy among Top-3 predictions. Based on the Top-3 predictions, the sensitivity/specificity of the AI group and the unaided were 84.6%/69.5% and 75.6%/62.7%, respectively (chi-square test, P \u00bc 0.45/0.13; Table 2). The sensitivity/specificity of the AIGP group and the unaidedGP group were 80.0%/81.5% and 56.3%/68.9%, respectively (chi-square test, P \u00bc 0.24/0.029). The sensitivity/specificity of the AIresident group and the unaidedresident group were 89.5%/57.1% and 86.2%/56.1%, respectively (chi-square test, P \u00bc 1.0/0.98).\n\nThere was a significant difference in the number of differential diagnoses between the AI group and the unaided group (2.09 vs. 1.95; Wilcoxon rank-sum test, P \u00bc 0.0005). The number of differential diagnoses of the AIGP group (2.00) was higher than that of the unaidedGP group (1.88; Wilcoxon rank-sum test, P < 0.0001). The number of differential diagnoses of the AIresident group (2.17) was also higher than that of the unaidedresident group (2.01; Wilcoxon rank-sum test, P \u00bc 0.019).\n\nBefore and after comparison: Individual analysis\nThe Top-1 and Top-3 accuracy before the use of AI was 41.3% \u00b1 15.7% and 49.2% \u00b1 13.2%, respectively (8 trainees with 295 cases; Supplementary Table S2). The Top-1 and Top-3 accuracy after the assistance of AI was increased to 53.6% \u00b1 9.3% and 69.8% \u00b1 10.6%, respectively. Individual improvement in Top-1 diagnostic accuracy for each trainee ranged from 5.4% to \u00fe41.4%, with an average (SD) of \u00fe12.2% (16.6%) which was not statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.076). On the contrary, Top-3 accuracy was improved by \u00fe0%\u2013\u00fe41.4%, with an average improvement of \u00fe20.6% (12.9%), which was statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.0027). There was individual variation in the degree of improvement: the Top-1 accuracy of one dermatology resident decreased (\u22125.3%), whereas that of one nondermatology trainee was improved by \u00fe41.4%.\n\nStandalone performance of the algorithm\nThe standalone Top-1 and Top-3 accuracy of the algorithm in the AI group was 50.5% and 74.9%, respectively. Area under the curve for determining malignancy was 0.894 (95% confidence interval \u00bc 0.838\u20130.950; DeLong method), which was equivalent to that of the attending dermatologists on the receiver operating characteristic curve (Figure 3). At the high-sensitive threshold, the sensitivity and specificity were 92.3% and 62.1%, respectively, and at the high-specificity threshold, the sensitivity and specificity were 84.6% and 78.5%, respectively. The sensitivity/specificity of the standalone algorithm derived from the Top-1 was 66.7%/90.6% and that from the Top-3 was 84.6%/61.3% (Table 2).\n\nAdverse effects of the incorrect predictions from the algorithm\nA 12.2% drop in Top-1 accuracy of the trainees was observed in cases where all Top-3 predictions from the algorithm were incorrect (Supplementary Tables S3 and S4). In out-of-distribution cases where Top-3 predictions from the algorithm were inevitably incorrect, the Top-1 accuracy of the trainees dropped by 5.6%. Both sensitivity/specificity derived from Top-3 predictions of trainees were negatively affected (before augmentation \u00bc 84.6%/60.7% and after augmentation \u00bc 61.5%/59.0%) when all Top 3 predictions from the algorithm were incorrect.\n\nThere were four cases of malignancy (three basal cell carcinomas and one Bowen\u2019s disease) for which the trainees had initially included a malignant condition in their Top-3 but ruled out malignancy in their final Top-3 recorded after the use of the algorithm. Of these four cases, two were predicted to be conditions requiring follow-up visits (dysplastic nevus and actinic keratosis).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Neoplasms: A Single-Center, Paralleled, Unmasked, Randomized Controlled Trial.",
        "abstract": "## TRIAL DESIGN\nThis was a single-center, unmasked, paralleled, randomized controlled trial.\n## METHODS\nA randomized trial was conducted in a tertiary care institute in South Korea to validate whether artificial intelligence (AI) could augment the accuracy of nonexpert physicians in the real-world settings, which included diverse out-of-distribution conditions. Consecutive patients aged >19 years, having one or more skin lesions suspicious for skin cancer detected by either the patient or physician, were randomly allocated to four nondermatology trainees and four dermatology residents. The attending dermatologists examined the randomly allocated patients with (AI-assisted group) or without (unaided group) the real-time assistance of AI algorithm (https://b2020.modelderm.com#world; convolutional neural networks; unmasked design) after simple randomization of the patients.\n## RESULTS\nUsing 576 consecutive cases (Fitzpatrick skin phototypes III or IV) with suspicious lesions out of the initial 603 recruitments, the accuracy of the AI-assisted group (n\u00a0= 295, 53.9%) was found to be significantly higher than those of the unaided group (n\u00a0= 281, 43.8%; P\u00a0= 0.019). Whereas the augmentation was more significant from 54.7% (n\u00a0= 150) to 30.7% (n\u00a0= 138; P < 0.0001) in the nondermatology trainees who had the least experience in dermatology, it was not significant in the dermatology residents. The algorithm could help trainees in the AI-assisted group include more differential diagnoses than the unaided group (2.09 vs. 1.95 diagnoses; P\u00a0= 0.0005). However, a 12.2% drop in Top-1 accuracy of the trainees was observed in cases in which all Top-3 predictions given by the algorithm were incorrect.\n## CONCLUSIONS\nThe multiclass AI algorithm augmented the diagnostic accuracy of nonexpert physicians in dermatology.\n",
        "methods": "### METHODS\n\nThis prospective study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea) (S2018-1703-0001). It was performed in the Department of Dermatology at Asan Medical Center, a tertiary care center in Korea. The study was conducted from 30 November 2020 to 9 September 2021 after online registration (KCT0005614; cris.nih.go.kr). The development of the algorithm (Model Dermatology, Build2020; https://b2020.modelderm.com#world) is described in the Supplementary Materials, and the algorithm was fixed on 19 September 2020. Along with the prediction of five differential diagnoses, the algorithm reports a malignancy score (range = 0-100). The malignancy score was defined as the sum of malignant outputs and 0.2 \u00d7 premalignant outputs as used previously (Han et al., 2020b). Using the subset of the Seoul National University dataset (240 images; https://doi.org/10.6084/m9.figshare.6454973), the high-sensitivity threshold for determining malignancy was defined as the threshold at which 90% sensitivity was obtained because the sensitivity of the attending dermatologists was at the level of 88.1% (Han et al., 2020a). The high-specificity threshold was defined as the threshold at which 80% sensitivity was obtained.\n\nIn a previous pilot study (Kim et al., 2022), the Top-1 accuracy of trainees was 47.9%. If 25% enhancement after the assistance were regarded as significant, the sample size was calculated as 548 (alpha = 0.05 and power = 0.8), and we planned to recruit 600 cases (Rosner, 2011).\n\nAll patients signed an informed consent before inclusion in the study. We included consecutive patients (aged >19 years) who had one or more skin lesions suspicious for skin cancer, detected by either the patient or physician. Exclusion criteria of patients and input data included patient refusal (10 cases), wrong recruitment (6 cases; aged \u226419 years), biopsy refusal (2 cases), and nonreal-time analysis (9 cases). Broken blindness and disclosure of the biopsy results in the referral note were also in the exclusion criteria, but there was no such case, and there was no performance error for the loss of internet connection or other technical issues. Formal pathologic diagnosis (504 cases) was used as the ground truth; however, if the pathologic report consisted of a pathologic description only (i.e., lichenoid reaction), the pathologic diagnosis was determined by clinicopathological correlation (20 cases). Clinical diagnosis of the attending dermatologists was used as the ground truth for the 52 cases where biopsy was not performed because the attending dermatologists decided not to biopsy the definitely benign cases. Ultimately, 524 biopsy-proven cases and 52 clinically-diagnosed cases were included in the final analysis among the 603 cases of the initial recruitment (Table 1). As shown in Supplementary Tables S6 and S7, a total of 53 conditions were within the trained 178 classes (in-distribution) and 29 conditions were not trained by the algorithm (out-of-distribution). There were no cases in which train-test contamination could be concerned, although there were 15 cases with photographs taken previously.\n\nA total of four attending physicians (3, 4, 6, and 22 years of experience after board certification), four first-year dermatology residents, and four nondermatology trainees (doctors in their first year after getting a medical license who rotate between various departments) participated in this study. Attending physicians routinely recorded their impressions after thorough examinations. After the simple randomization using a custom randomizer (rand function of the PHP language) by the attending dermatologists, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and recorded their diagnostic hypothesis in real time.\n\nIn the AI group, trainees captured and selected 1-3 photographs with age and sex metadata as an input data and uploaded them to http://b2020.modelderm.com#world using internet browsers. Then the after-diagnoses was recorded, referring to the five diagnoses suggested by the algorithm and the malignancy score (Supplementary Figure S2). The photographs that the trainees thought to be of adequate quality and composition by themselves were uploaded. Trainees were instructed so that only macroimages having the large lesion of interest at the center are uploaded. In the unaided group, trainees performed routine examinations and recorded the three most probable diagnoses, without the assistance of the algorithm. The use of dermoscopy was not allowed for all trainees.\n\nStatistical analysis\nIn calculating the Top accuracy, only the exact diagnosis was recorded as correct, but giving a specific subtype of the disease was also counted to be correct. For example, intradermal nevus was counted correct for the ground truth of junctional nevus. We lumped together 364 diagnoses in natural language into the 82 diagnosis codes (https://doi.org/10.6084/m9.figshare.16640257). For evaluating a malignancy prediction, the physicians diagnoses were transformed into either malignant or benign. Top accuracies, sensitivities, and specificities were compared using Pearson\u2019s chi-square test with Yates\u2019 continuity correction (AI group vs. unaided group) or McNemar test (before vs. after the assistance of the algorithm) using R version 4.1.1, and P < 0.05 was statistically significant.\n\nData availability statement\nThe raw data of this study is available at https://doi.org/10.6084/m9.figshare.16640257. The diagnosis in natural language, converted diagnosis code (ALIAS sheet), the cases determined by clinicopathological correlation, accuracy/sensitivity/specificity/positive predictive value/negative predictive value of the participants and algorithm, and demographics of subjects are listed. The algorithm is accessible at https://b2020.modelderm.com/#world or https://modelderm.com (Build2020) for academic purposes, and only the images with the permission of the original owner should be uploaded.",
        "results": "### RESULTS\nThis study is reported per the Consolidated Standards of Reporting Trials\u2014Artificial Intelligence guidelines (Liu et al., 2020). Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u2019correct\u2019.\n\nAI group versus unaided group\nTo evaluate that the two groups were truly comparable, accuracies of the attending dermatologists and trainees (before interventions) were compared. Top-1 accuracy of attending dermatologists (62.3%) and trainees (43.8%) of the unaided group were higher than those of the AI group (dermatologists \u00bc 60.3%, trainees \u00bc 41.0%), which indicated that easier cases were not disproportionately allocated to the AI group (Table 1 and Figure 1). Overall, the Top-1 accuracy of the AI group was 53.9% and that of the unaided was 43.8% (all trainees; chi-square test, P \u00bc 0.019; Figure 2 and Table 2). There were significant differences in the result depending on whether the participant was a nondermatology trainee (general practitioner [GP]) or a dermatology resident (resident). The Top-1 accuracy of the AIGP group (54.7%) was markedly higher than that of unaidedGP group (29.7%; chi-square test, P < 0.0001), whereas the Top-1 accuracy of the AIresident group and the unaidedresident group was 53.1% and 57.3%, respectively (chi-square test, P \u00bc 0.55). In the AI group, we compared the judgment before and after receiving assistance of the algorithm and there was a significant enhancement in the Top 1 accuracy of the AIGP group (before augmentation \u00bc 30.7%, after augmentation \u00bc 54.7%; McNemar test, P < 0.0001). However, in the AIresident group, the change was not significant (before augmentation \u00bc 51.7%, after augmentation \u00bc 53.1%; McNemar test, P \u00bc 0.86). As shown in Supplementary Table S1, a greater improvement in accuracy was observed for the subsequent cases, up to \u00fe25.0% and \u00fe37.5% for Top-1 and Top-3 accuracy, respectively, although the accuracy for Top-1 and Top-3 was improved by only \u00fe0.0% and \u00fe15.0%, respectively, for the first five cases. When the analysis was restricted to 266 cases that were biopsied, the Top-1/Top-3 accuracy of the standalone algorithm, trainees before augmentation, trainees after augmentation, and attending dermatologists were 51.5%/75.9%, 40.2%/49.2%, 56.0%/71.4%, and 56.0%/66.2%, respectively. The accuracies of the AI-augmented trainees were equivalent to those of the attending dermatologists. In the 258 biopsy-proven cases of the unaided group, the Top-1/Top-3 accuracy of trainees and attending dermatologists were 42.6%/57.4% and 58.9%/68.6%, respectively. Malignancy determination affects clinical decisions such as ordering a biopsy if there is any malignancy among Top-3 predictions. Based on the Top-3 predictions, the sensitivity/specificity of the AI group and the unaided were 84.6%/69.5% and 75.6%/62.7%, respectively (chi-square test, P \u00bc 0.45/0.13; Table 2). The sensitivity/specificity of the AIGP group and the unaidedGP group were 80.0%/81.5% and 56.3%/68.9%, respectively (chi-square test, P \u00bc 0.24/0.029). The sensitivity/specificity of the AIresident group and the unaidedresident group were 89.5%/57.1% and 86.2%/56.1%, respectively (chi-square test, P \u00bc 1.0/0.98).\n\nThere was a significant difference in the number of differential diagnoses between the AI group and the unaided group (2.09 vs. 1.95; Wilcoxon rank-sum test, P \u00bc 0.0005). The number of differential diagnoses of the AIGP group (2.00) was higher than that of the unaidedGP group (1.88; Wilcoxon rank-sum test, P < 0.0001). The number of differential diagnoses of the AIresident group (2.17) was also higher than that of the unaidedresident group (2.01; Wilcoxon rank-sum test, P \u00bc 0.019).\n\nBefore and after comparison: Individual analysis\nThe Top-1 and Top-3 accuracy before the use of AI was 41.3% \u00b1 15.7% and 49.2% \u00b1 13.2%, respectively (8 trainees with 295 cases; Supplementary Table S2). The Top-1 and Top-3 accuracy after the assistance of AI was increased to 53.6% \u00b1 9.3% and 69.8% \u00b1 10.6%, respectively. Individual improvement in Top-1 diagnostic accuracy for each trainee ranged from 5.4% to \u00fe41.4%, with an average (SD) of \u00fe12.2% (16.6%) which was not statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.076). On the contrary, Top-3 accuracy was improved by \u00fe0%\u2013\u00fe41.4%, with an average improvement of \u00fe20.6% (12.9%), which was statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.0027). There was individual variation in the degree of improvement: the Top-1 accuracy of one dermatology resident decreased (\u22125.3%), whereas that of one nondermatology trainee was improved by \u00fe41.4%.\n\nStandalone performance of the algorithm\nThe standalone Top-1 and Top-3 accuracy of the algorithm in the AI group was 50.5% and 74.9%, respectively. Area under the curve for determining malignancy was 0.894 (95% confidence interval \u00bc 0.838\u20130.950; DeLong method), which was equivalent to that of the attending dermatologists on the receiver operating characteristic curve (Figure 3). At the high-sensitive threshold, the sensitivity and specificity were 92.3% and 62.1%, respectively, and at the high-specificity threshold, the sensitivity and specificity were 84.6% and 78.5%, respectively. The sensitivity/specificity of the standalone algorithm derived from the Top-1 was 66.7%/90.6% and that from the Top-3 was 84.6%/61.3% (Table 2).\n\nAdverse effects of the incorrect predictions from the algorithm\nA 12.2% drop in Top-1 accuracy of the trainees was observed in cases where all Top-3 predictions from the algorithm were incorrect (Supplementary Tables S3 and S4). In out-of-distribution cases where Top-3 predictions from the algorithm were inevitably incorrect, the Top-1 accuracy of the trainees dropped by 5.6%. Both sensitivity/specificity derived from Top-3 predictions of trainees were negatively affected (before augmentation \u00bc 84.6%/60.7% and after augmentation \u00bc 61.5%/59.0%) when all Top 3 predictions from the algorithm were incorrect.\n\nThere were four cases of malignancy (three basal cell carcinomas and one Bowen\u2019s disease) for which the trainees had initially included a malignant condition in their Top-3 but ruled out malignancy in their final Top-3 recorded after the use of the algorithm. Of these four cases, two were predicted to be conditions requiring follow-up visits (dysplastic nevus and actinic keratosis).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Neoplasms: A Single-Center, Paralleled, Unmasked, Randomized Controlled Trial.",
        "abstract": "## TRIAL DESIGN\nThis was a single-center, unmasked, paralleled, randomized controlled trial.\n## METHODS\nA randomized trial was conducted in a tertiary care institute in South Korea to validate whether artificial intelligence (AI) could augment the accuracy of nonexpert physicians in the real-world settings, which included diverse out-of-distribution conditions. Consecutive patients aged >19 years, having one or more skin lesions suspicious for skin cancer detected by either the patient or physician, were randomly allocated to four nondermatology trainees and four dermatology residents. The attending dermatologists examined the randomly allocated patients with (AI-assisted group) or without (unaided group) the real-time assistance of AI algorithm (https://b2020.modelderm.com#world; convolutional neural networks; unmasked design) after simple randomization of the patients.\n## RESULTS\nUsing 576 consecutive cases (Fitzpatrick skin phototypes III or IV) with suspicious lesions out of the initial 603 recruitments, the accuracy of the AI-assisted group (n\u00a0= 295, 53.9%) was found to be significantly higher than those of the unaided group (n\u00a0= 281, 43.8%; P\u00a0= 0.019). Whereas the augmentation was more significant from 54.7% (n\u00a0= 150) to 30.7% (n\u00a0= 138; P < 0.0001) in the nondermatology trainees who had the least experience in dermatology, it was not significant in the dermatology residents. The algorithm could help trainees in the AI-assisted group include more differential diagnoses than the unaided group (2.09 vs. 1.95 diagnoses; P\u00a0= 0.0005). However, a 12.2% drop in Top-1 accuracy of the trainees was observed in cases in which all Top-3 predictions given by the algorithm were incorrect.\n## CONCLUSIONS\nThe multiclass AI algorithm augmented the diagnostic accuracy of nonexpert physicians in dermatology.\n",
        "methods": "### METHODS\n\nThis prospective study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea) (S2018-1703-0001). It was performed in the Department of Dermatology at Asan Medical Center, a tertiary care center in Korea. The study was conducted from 30 November 2020 to 9 September 2021 after online registration (KCT0005614; cris.nih.go.kr). The development of the algorithm (Model Dermatology, Build2020; https://b2020.modelderm.com#world) is described in the Supplementary Materials, and the algorithm was fixed on 19 September 2020. Along with the prediction of five differential diagnoses, the algorithm reports a malignancy score (range = 0-100). The malignancy score was defined as the sum of malignant outputs and 0.2 \u00d7 premalignant outputs as used previously (Han et al., 2020b). Using the subset of the Seoul National University dataset (240 images; https://doi.org/10.6084/m9.figshare.6454973), the high-sensitivity threshold for determining malignancy was defined as the threshold at which 90% sensitivity was obtained because the sensitivity of the attending dermatologists was at the level of 88.1% (Han et al., 2020a). The high-specificity threshold was defined as the threshold at which 80% sensitivity was obtained.\n\nIn a previous pilot study (Kim et al., 2022), the Top-1 accuracy of trainees was 47.9%. If 25% enhancement after the assistance were regarded as significant, the sample size was calculated as 548 (alpha = 0.05 and power = 0.8), and we planned to recruit 600 cases (Rosner, 2011).\n\nAll patients signed an informed consent before inclusion in the study. We included consecutive patients (aged >19 years) who had one or more skin lesions suspicious for skin cancer, detected by either the patient or physician. Exclusion criteria of patients and input data included patient refusal (10 cases), wrong recruitment (6 cases; aged \u226419 years), biopsy refusal (2 cases), and nonreal-time analysis (9 cases). Broken blindness and disclosure of the biopsy results in the referral note were also in the exclusion criteria, but there was no such case, and there was no performance error for the loss of internet connection or other technical issues. Formal pathologic diagnosis (504 cases) was used as the ground truth; however, if the pathologic report consisted of a pathologic description only (i.e., lichenoid reaction), the pathologic diagnosis was determined by clinicopathological correlation (20 cases). Clinical diagnosis of the attending dermatologists was used as the ground truth for the 52 cases where biopsy was not performed because the attending dermatologists decided not to biopsy the definitely benign cases. Ultimately, 524 biopsy-proven cases and 52 clinically-diagnosed cases were included in the final analysis among the 603 cases of the initial recruitment (Table 1). As shown in Supplementary Tables S6 and S7, a total of 53 conditions were within the trained 178 classes (in-distribution) and 29 conditions were not trained by the algorithm (out-of-distribution). There were no cases in which train-test contamination could be concerned, although there were 15 cases with photographs taken previously.\n\nA total of four attending physicians (3, 4, 6, and 22 years of experience after board certification), four first-year dermatology residents, and four nondermatology trainees (doctors in their first year after getting a medical license who rotate between various departments) participated in this study. Attending physicians routinely recorded their impressions after thorough examinations. After the simple randomization using a custom randomizer (rand function of the PHP language) by the attending dermatologists, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and recorded their diagnostic hypothesis in real time.\n\nIn the AI group, trainees captured and selected 1-3 photographs with age and sex metadata as an input data and uploaded them to http://b2020.modelderm.com#world using internet browsers. Then the after-diagnoses was recorded, referring to the five diagnoses suggested by the algorithm and the malignancy score (Supplementary Figure S2). The photographs that the trainees thought to be of adequate quality and composition by themselves were uploaded. Trainees were instructed so that only macroimages having the large lesion of interest at the center are uploaded. In the unaided group, trainees performed routine examinations and recorded the three most probable diagnoses, without the assistance of the algorithm. The use of dermoscopy was not allowed for all trainees.\n\nStatistical analysis\nIn calculating the Top accuracy, only the exact diagnosis was recorded as correct, but giving a specific subtype of the disease was also counted to be correct. For example, intradermal nevus was counted correct for the ground truth of junctional nevus. We lumped together 364 diagnoses in natural language into the 82 diagnosis codes (https://doi.org/10.6084/m9.figshare.16640257). For evaluating a malignancy prediction, the physicians diagnoses were transformed into either malignant or benign. Top accuracies, sensitivities, and specificities were compared using Pearson\u2019s chi-square test with Yates\u2019 continuity correction (AI group vs. unaided group) or McNemar test (before vs. after the assistance of the algorithm) using R version 4.1.1, and P < 0.05 was statistically significant.\n\nData availability statement\nThe raw data of this study is available at https://doi.org/10.6084/m9.figshare.16640257. The diagnosis in natural language, converted diagnosis code (ALIAS sheet), the cases determined by clinicopathological correlation, accuracy/sensitivity/specificity/positive predictive value/negative predictive value of the participants and algorithm, and demographics of subjects are listed. The algorithm is accessible at https://b2020.modelderm.com/#world or https://modelderm.com (Build2020) for academic purposes, and only the images with the permission of the original owner should be uploaded.",
        "results": "### RESULTS\nThis study is reported per the Consolidated Standards of Reporting Trials\u2014Artificial Intelligence guidelines (Liu et al., 2020). Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u2019correct\u2019.\n\nAI group versus unaided group\nTo evaluate that the two groups were truly comparable, accuracies of the attending dermatologists and trainees (before interventions) were compared. Top-1 accuracy of attending dermatologists (62.3%) and trainees (43.8%) of the unaided group were higher than those of the AI group (dermatologists \u00bc 60.3%, trainees \u00bc 41.0%), which indicated that easier cases were not disproportionately allocated to the AI group (Table 1 and Figure 1). Overall, the Top-1 accuracy of the AI group was 53.9% and that of the unaided was 43.8% (all trainees; chi-square test, P \u00bc 0.019; Figure 2 and Table 2). There were significant differences in the result depending on whether the participant was a nondermatology trainee (general practitioner [GP]) or a dermatology resident (resident). The Top-1 accuracy of the AIGP group (54.7%) was markedly higher than that of unaidedGP group (29.7%; chi-square test, P < 0.0001), whereas the Top-1 accuracy of the AIresident group and the unaidedresident group was 53.1% and 57.3%, respectively (chi-square test, P \u00bc 0.55). In the AI group, we compared the judgment before and after receiving assistance of the algorithm and there was a significant enhancement in the Top 1 accuracy of the AIGP group (before augmentation \u00bc 30.7%, after augmentation \u00bc 54.7%; McNemar test, P < 0.0001). However, in the AIresident group, the change was not significant (before augmentation \u00bc 51.7%, after augmentation \u00bc 53.1%; McNemar test, P \u00bc 0.86). As shown in Supplementary Table S1, a greater improvement in accuracy was observed for the subsequent cases, up to \u00fe25.0% and \u00fe37.5% for Top-1 and Top-3 accuracy, respectively, although the accuracy for Top-1 and Top-3 was improved by only \u00fe0.0% and \u00fe15.0%, respectively, for the first five cases. When the analysis was restricted to 266 cases that were biopsied, the Top-1/Top-3 accuracy of the standalone algorithm, trainees before augmentation, trainees after augmentation, and attending dermatologists were 51.5%/75.9%, 40.2%/49.2%, 56.0%/71.4%, and 56.0%/66.2%, respectively. The accuracies of the AI-augmented trainees were equivalent to those of the attending dermatologists. In the 258 biopsy-proven cases of the unaided group, the Top-1/Top-3 accuracy of trainees and attending dermatologists were 42.6%/57.4% and 58.9%/68.6%, respectively. Malignancy determination affects clinical decisions such as ordering a biopsy if there is any malignancy among Top-3 predictions. Based on the Top-3 predictions, the sensitivity/specificity of the AI group and the unaided were 84.6%/69.5% and 75.6%/62.7%, respectively (chi-square test, P \u00bc 0.45/0.13; Table 2). The sensitivity/specificity of the AIGP group and the unaidedGP group were 80.0%/81.5% and 56.3%/68.9%, respectively (chi-square test, P \u00bc 0.24/0.029). The sensitivity/specificity of the AIresident group and the unaidedresident group were 89.5%/57.1% and 86.2%/56.1%, respectively (chi-square test, P \u00bc 1.0/0.98).\n\nThere was a significant difference in the number of differential diagnoses between the AI group and the unaided group (2.09 vs. 1.95; Wilcoxon rank-sum test, P \u00bc 0.0005). The number of differential diagnoses of the AIGP group (2.00) was higher than that of the unaidedGP group (1.88; Wilcoxon rank-sum test, P < 0.0001). The number of differential diagnoses of the AIresident group (2.17) was also higher than that of the unaidedresident group (2.01; Wilcoxon rank-sum test, P \u00bc 0.019).\n\nBefore and after comparison: Individual analysis\nThe Top-1 and Top-3 accuracy before the use of AI was 41.3% \u00b1 15.7% and 49.2% \u00b1 13.2%, respectively (8 trainees with 295 cases; Supplementary Table S2). The Top-1 and Top-3 accuracy after the assistance of AI was increased to 53.6% \u00b1 9.3% and 69.8% \u00b1 10.6%, respectively. Individual improvement in Top-1 diagnostic accuracy for each trainee ranged from 5.4% to \u00fe41.4%, with an average (SD) of \u00fe12.2% (16.6%) which was not statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.076). On the contrary, Top-3 accuracy was improved by \u00fe0%\u2013\u00fe41.4%, with an average improvement of \u00fe20.6% (12.9%), which was statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.0027). There was individual variation in the degree of improvement: the Top-1 accuracy of one dermatology resident decreased (\u22125.3%), whereas that of one nondermatology trainee was improved by \u00fe41.4%.\n\nStandalone performance of the algorithm\nThe standalone Top-1 and Top-3 accuracy of the algorithm in the AI group was 50.5% and 74.9%, respectively. Area under the curve for determining malignancy was 0.894 (95% confidence interval \u00bc 0.838\u20130.950; DeLong method), which was equivalent to that of the attending dermatologists on the receiver operating characteristic curve (Figure 3). At the high-sensitive threshold, the sensitivity and specificity were 92.3% and 62.1%, respectively, and at the high-specificity threshold, the sensitivity and specificity were 84.6% and 78.5%, respectively. The sensitivity/specificity of the standalone algorithm derived from the Top-1 was 66.7%/90.6% and that from the Top-3 was 84.6%/61.3% (Table 2).\n\nAdverse effects of the incorrect predictions from the algorithm\nA 12.2% drop in Top-1 accuracy of the trainees was observed in cases where all Top-3 predictions from the algorithm were incorrect (Supplementary Tables S3 and S4). In out-of-distribution cases where Top-3 predictions from the algorithm were inevitably incorrect, the Top-1 accuracy of the trainees dropped by 5.6%. Both sensitivity/specificity derived from Top-3 predictions of trainees were negatively affected (before augmentation \u00bc 84.6%/60.7% and after augmentation \u00bc 61.5%/59.0%) when all Top 3 predictions from the algorithm were incorrect.\n\nThere were four cases of malignancy (three basal cell carcinomas and one Bowen\u2019s disease) for which the trainees had initially included a malignant condition in their Top-3 but ruled out malignancy in their final Top-3 recorded after the use of the algorithm. Of these four cases, two were predicted to be conditions requiring follow-up visits (dysplastic nevus and actinic keratosis).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Neoplasms: A Single-Center, Paralleled, Unmasked, Randomized Controlled Trial.",
        "abstract": "## TRIAL DESIGN\nThis was a single-center, unmasked, paralleled, randomized controlled trial.\n## METHODS\nA randomized trial was conducted in a tertiary care institute in South Korea to validate whether artificial intelligence (AI) could augment the accuracy of nonexpert physicians in the real-world settings, which included diverse out-of-distribution conditions. Consecutive patients aged >19 years, having one or more skin lesions suspicious for skin cancer detected by either the patient or physician, were randomly allocated to four nondermatology trainees and four dermatology residents. The attending dermatologists examined the randomly allocated patients with (AI-assisted group) or without (unaided group) the real-time assistance of AI algorithm (https://b2020.modelderm.com#world; convolutional neural networks; unmasked design) after simple randomization of the patients.\n## RESULTS\nUsing 576 consecutive cases (Fitzpatrick skin phototypes III or IV) with suspicious lesions out of the initial 603 recruitments, the accuracy of the AI-assisted group (n\u00a0= 295, 53.9%) was found to be significantly higher than those of the unaided group (n\u00a0= 281, 43.8%; P\u00a0= 0.019). Whereas the augmentation was more significant from 54.7% (n\u00a0= 150) to 30.7% (n\u00a0= 138; P < 0.0001) in the nondermatology trainees who had the least experience in dermatology, it was not significant in the dermatology residents. The algorithm could help trainees in the AI-assisted group include more differential diagnoses than the unaided group (2.09 vs. 1.95 diagnoses; P\u00a0= 0.0005). However, a 12.2% drop in Top-1 accuracy of the trainees was observed in cases in which all Top-3 predictions given by the algorithm were incorrect.\n## CONCLUSIONS\nThe multiclass AI algorithm augmented the diagnostic accuracy of nonexpert physicians in dermatology.\n",
        "methods": "### METHODS\n\nThis prospective study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea) (S2018-1703-0001). It was performed in the Department of Dermatology at Asan Medical Center, a tertiary care center in Korea. The study was conducted from 30 November 2020 to 9 September 2021 after online registration (KCT0005614; cris.nih.go.kr). The development of the algorithm (Model Dermatology, Build2020; https://b2020.modelderm.com#world) is described in the Supplementary Materials, and the algorithm was fixed on 19 September 2020. Along with the prediction of five differential diagnoses, the algorithm reports a malignancy score (range = 0-100). The malignancy score was defined as the sum of malignant outputs and 0.2 \u00d7 premalignant outputs as used previously (Han et al., 2020b). Using the subset of the Seoul National University dataset (240 images; https://doi.org/10.6084/m9.figshare.6454973), the high-sensitivity threshold for determining malignancy was defined as the threshold at which 90% sensitivity was obtained because the sensitivity of the attending dermatologists was at the level of 88.1% (Han et al., 2020a). The high-specificity threshold was defined as the threshold at which 80% sensitivity was obtained.\n\nIn a previous pilot study (Kim et al., 2022), the Top-1 accuracy of trainees was 47.9%. If 25% enhancement after the assistance were regarded as significant, the sample size was calculated as 548 (alpha = 0.05 and power = 0.8), and we planned to recruit 600 cases (Rosner, 2011).\n\nAll patients signed an informed consent before inclusion in the study. We included consecutive patients (aged >19 years) who had one or more skin lesions suspicious for skin cancer, detected by either the patient or physician. Exclusion criteria of patients and input data included patient refusal (10 cases), wrong recruitment (6 cases; aged \u226419 years), biopsy refusal (2 cases), and nonreal-time analysis (9 cases). Broken blindness and disclosure of the biopsy results in the referral note were also in the exclusion criteria, but there was no such case, and there was no performance error for the loss of internet connection or other technical issues. Formal pathologic diagnosis (504 cases) was used as the ground truth; however, if the pathologic report consisted of a pathologic description only (i.e., lichenoid reaction), the pathologic diagnosis was determined by clinicopathological correlation (20 cases). Clinical diagnosis of the attending dermatologists was used as the ground truth for the 52 cases where biopsy was not performed because the attending dermatologists decided not to biopsy the definitely benign cases. Ultimately, 524 biopsy-proven cases and 52 clinically-diagnosed cases were included in the final analysis among the 603 cases of the initial recruitment (Table 1). As shown in Supplementary Tables S6 and S7, a total of 53 conditions were within the trained 178 classes (in-distribution) and 29 conditions were not trained by the algorithm (out-of-distribution). There were no cases in which train-test contamination could be concerned, although there were 15 cases with photographs taken previously.\n\nA total of four attending physicians (3, 4, 6, and 22 years of experience after board certification), four first-year dermatology residents, and four nondermatology trainees (doctors in their first year after getting a medical license who rotate between various departments) participated in this study. Attending physicians routinely recorded their impressions after thorough examinations. After the simple randomization using a custom randomizer (rand function of the PHP language) by the attending dermatologists, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and recorded their diagnostic hypothesis in real time.\n\nIn the AI group, trainees captured and selected 1-3 photographs with age and sex metadata as an input data and uploaded them to http://b2020.modelderm.com#world using internet browsers. Then the after-diagnoses was recorded, referring to the five diagnoses suggested by the algorithm and the malignancy score (Supplementary Figure S2). The photographs that the trainees thought to be of adequate quality and composition by themselves were uploaded. Trainees were instructed so that only macroimages having the large lesion of interest at the center are uploaded. In the unaided group, trainees performed routine examinations and recorded the three most probable diagnoses, without the assistance of the algorithm. The use of dermoscopy was not allowed for all trainees.\n\nStatistical analysis\nIn calculating the Top accuracy, only the exact diagnosis was recorded as correct, but giving a specific subtype of the disease was also counted to be correct. For example, intradermal nevus was counted correct for the ground truth of junctional nevus. We lumped together 364 diagnoses in natural language into the 82 diagnosis codes (https://doi.org/10.6084/m9.figshare.16640257). For evaluating a malignancy prediction, the physicians diagnoses were transformed into either malignant or benign. Top accuracies, sensitivities, and specificities were compared using Pearson\u2019s chi-square test with Yates\u2019 continuity correction (AI group vs. unaided group) or McNemar test (before vs. after the assistance of the algorithm) using R version 4.1.1, and P < 0.05 was statistically significant.\n\nData availability statement\nThe raw data of this study is available at https://doi.org/10.6084/m9.figshare.16640257. The diagnosis in natural language, converted diagnosis code (ALIAS sheet), the cases determined by clinicopathological correlation, accuracy/sensitivity/specificity/positive predictive value/negative predictive value of the participants and algorithm, and demographics of subjects are listed. The algorithm is accessible at https://b2020.modelderm.com/#world or https://modelderm.com (Build2020) for academic purposes, and only the images with the permission of the original owner should be uploaded.",
        "results": "### RESULTS\nThis study is reported per the Consolidated Standards of Reporting Trials\u2014Artificial Intelligence guidelines (Liu et al., 2020). Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u2019correct\u2019.\n\nAI group versus unaided group\nTo evaluate that the two groups were truly comparable, accuracies of the attending dermatologists and trainees (before interventions) were compared. Top-1 accuracy of attending dermatologists (62.3%) and trainees (43.8%) of the unaided group were higher than those of the AI group (dermatologists \u00bc 60.3%, trainees \u00bc 41.0%), which indicated that easier cases were not disproportionately allocated to the AI group (Table 1 and Figure 1). Overall, the Top-1 accuracy of the AI group was 53.9% and that of the unaided was 43.8% (all trainees; chi-square test, P \u00bc 0.019; Figure 2 and Table 2). There were significant differences in the result depending on whether the participant was a nondermatology trainee (general practitioner [GP]) or a dermatology resident (resident). The Top-1 accuracy of the AIGP group (54.7%) was markedly higher than that of unaidedGP group (29.7%; chi-square test, P < 0.0001), whereas the Top-1 accuracy of the AIresident group and the unaidedresident group was 53.1% and 57.3%, respectively (chi-square test, P \u00bc 0.55). In the AI group, we compared the judgment before and after receiving assistance of the algorithm and there was a significant enhancement in the Top 1 accuracy of the AIGP group (before augmentation \u00bc 30.7%, after augmentation \u00bc 54.7%; McNemar test, P < 0.0001). However, in the AIresident group, the change was not significant (before augmentation \u00bc 51.7%, after augmentation \u00bc 53.1%; McNemar test, P \u00bc 0.86). As shown in Supplementary Table S1, a greater improvement in accuracy was observed for the subsequent cases, up to \u00fe25.0% and \u00fe37.5% for Top-1 and Top-3 accuracy, respectively, although the accuracy for Top-1 and Top-3 was improved by only \u00fe0.0% and \u00fe15.0%, respectively, for the first five cases. When the analysis was restricted to 266 cases that were biopsied, the Top-1/Top-3 accuracy of the standalone algorithm, trainees before augmentation, trainees after augmentation, and attending dermatologists were 51.5%/75.9%, 40.2%/49.2%, 56.0%/71.4%, and 56.0%/66.2%, respectively. The accuracies of the AI-augmented trainees were equivalent to those of the attending dermatologists. In the 258 biopsy-proven cases of the unaided group, the Top-1/Top-3 accuracy of trainees and attending dermatologists were 42.6%/57.4% and 58.9%/68.6%, respectively. Malignancy determination affects clinical decisions such as ordering a biopsy if there is any malignancy among Top-3 predictions. Based on the Top-3 predictions, the sensitivity/specificity of the AI group and the unaided were 84.6%/69.5% and 75.6%/62.7%, respectively (chi-square test, P \u00bc 0.45/0.13; Table 2). The sensitivity/specificity of the AIGP group and the unaidedGP group were 80.0%/81.5% and 56.3%/68.9%, respectively (chi-square test, P \u00bc 0.24/0.029). The sensitivity/specificity of the AIresident group and the unaidedresident group were 89.5%/57.1% and 86.2%/56.1%, respectively (chi-square test, P \u00bc 1.0/0.98).\n\nThere was a significant difference in the number of differential diagnoses between the AI group and the unaided group (2.09 vs. 1.95; Wilcoxon rank-sum test, P \u00bc 0.0005). The number of differential diagnoses of the AIGP group (2.00) was higher than that of the unaidedGP group (1.88; Wilcoxon rank-sum test, P < 0.0001). The number of differential diagnoses of the AIresident group (2.17) was also higher than that of the unaidedresident group (2.01; Wilcoxon rank-sum test, P \u00bc 0.019).\n\nBefore and after comparison: Individual analysis\nThe Top-1 and Top-3 accuracy before the use of AI was 41.3% \u00b1 15.7% and 49.2% \u00b1 13.2%, respectively (8 trainees with 295 cases; Supplementary Table S2). The Top-1 and Top-3 accuracy after the assistance of AI was increased to 53.6% \u00b1 9.3% and 69.8% \u00b1 10.6%, respectively. Individual improvement in Top-1 diagnostic accuracy for each trainee ranged from 5.4% to \u00fe41.4%, with an average (SD) of \u00fe12.2% (16.6%) which was not statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.076). On the contrary, Top-3 accuracy was improved by \u00fe0%\u2013\u00fe41.4%, with an average improvement of \u00fe20.6% (12.9%), which was statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.0027). There was individual variation in the degree of improvement: the Top-1 accuracy of one dermatology resident decreased (\u22125.3%), whereas that of one nondermatology trainee was improved by \u00fe41.4%.\n\nStandalone performance of the algorithm\nThe standalone Top-1 and Top-3 accuracy of the algorithm in the AI group was 50.5% and 74.9%, respectively. Area under the curve for determining malignancy was 0.894 (95% confidence interval \u00bc 0.838\u20130.950; DeLong method), which was equivalent to that of the attending dermatologists on the receiver operating characteristic curve (Figure 3). At the high-sensitive threshold, the sensitivity and specificity were 92.3% and 62.1%, respectively, and at the high-specificity threshold, the sensitivity and specificity were 84.6% and 78.5%, respectively. The sensitivity/specificity of the standalone algorithm derived from the Top-1 was 66.7%/90.6% and that from the Top-3 was 84.6%/61.3% (Table 2).\n\nAdverse effects of the incorrect predictions from the algorithm\nA 12.2% drop in Top-1 accuracy of the trainees was observed in cases where all Top-3 predictions from the algorithm were incorrect (Supplementary Tables S3 and S4). In out-of-distribution cases where Top-3 predictions from the algorithm were inevitably incorrect, the Top-1 accuracy of the trainees dropped by 5.6%. Both sensitivity/specificity derived from Top-3 predictions of trainees were negatively affected (before augmentation \u00bc 84.6%/60.7% and after augmentation \u00bc 61.5%/59.0%) when all Top 3 predictions from the algorithm were incorrect.\n\nThere were four cases of malignancy (three basal cell carcinomas and one Bowen\u2019s disease) for which the trainees had initially included a malignant condition in their Top-3 but ruled out malignancy in their final Top-3 recorded after the use of the algorithm. Of these four cases, two were predicted to be conditions requiring follow-up visits (dysplastic nevus and actinic keratosis).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Neoplasms: A Single-Center, Paralleled, Unmasked, Randomized Controlled Trial.",
        "abstract": "## TRIAL DESIGN\nThis was a single-center, unmasked, paralleled, randomized controlled trial.\n## METHODS\nA randomized trial was conducted in a tertiary care institute in South Korea to validate whether artificial intelligence (AI) could augment the accuracy of nonexpert physicians in the real-world settings, which included diverse out-of-distribution conditions. Consecutive patients aged >19 years, having one or more skin lesions suspicious for skin cancer detected by either the patient or physician, were randomly allocated to four nondermatology trainees and four dermatology residents. The attending dermatologists examined the randomly allocated patients with (AI-assisted group) or without (unaided group) the real-time assistance of AI algorithm (https://b2020.modelderm.com#world; convolutional neural networks; unmasked design) after simple randomization of the patients.\n## RESULTS\nUsing 576 consecutive cases (Fitzpatrick skin phototypes III or IV) with suspicious lesions out of the initial 603 recruitments, the accuracy of the AI-assisted group (n\u00a0= 295, 53.9%) was found to be significantly higher than those of the unaided group (n\u00a0= 281, 43.8%; P\u00a0= 0.019). Whereas the augmentation was more significant from 54.7% (n\u00a0= 150) to 30.7% (n\u00a0= 138; P < 0.0001) in the nondermatology trainees who had the least experience in dermatology, it was not significant in the dermatology residents. The algorithm could help trainees in the AI-assisted group include more differential diagnoses than the unaided group (2.09 vs. 1.95 diagnoses; P\u00a0= 0.0005). However, a 12.2% drop in Top-1 accuracy of the trainees was observed in cases in which all Top-3 predictions given by the algorithm were incorrect.\n## CONCLUSIONS\nThe multiclass AI algorithm augmented the diagnostic accuracy of nonexpert physicians in dermatology.\n",
        "methods": "### METHODS\n\nThis prospective study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea) (S2018-1703-0001). It was performed in the Department of Dermatology at Asan Medical Center, a tertiary care center in Korea. The study was conducted from 30 November 2020 to 9 September 2021 after online registration (KCT0005614; cris.nih.go.kr). The development of the algorithm (Model Dermatology, Build2020; https://b2020.modelderm.com#world) is described in the Supplementary Materials, and the algorithm was fixed on 19 September 2020. Along with the prediction of five differential diagnoses, the algorithm reports a malignancy score (range = 0-100). The malignancy score was defined as the sum of malignant outputs and 0.2 \u00d7 premalignant outputs as used previously (Han et al., 2020b). Using the subset of the Seoul National University dataset (240 images; https://doi.org/10.6084/m9.figshare.6454973), the high-sensitivity threshold for determining malignancy was defined as the threshold at which 90% sensitivity was obtained because the sensitivity of the attending dermatologists was at the level of 88.1% (Han et al., 2020a). The high-specificity threshold was defined as the threshold at which 80% sensitivity was obtained.\n\nIn a previous pilot study (Kim et al., 2022), the Top-1 accuracy of trainees was 47.9%. If 25% enhancement after the assistance were regarded as significant, the sample size was calculated as 548 (alpha = 0.05 and power = 0.8), and we planned to recruit 600 cases (Rosner, 2011).\n\nAll patients signed an informed consent before inclusion in the study. We included consecutive patients (aged >19 years) who had one or more skin lesions suspicious for skin cancer, detected by either the patient or physician. Exclusion criteria of patients and input data included patient refusal (10 cases), wrong recruitment (6 cases; aged \u226419 years), biopsy refusal (2 cases), and nonreal-time analysis (9 cases). Broken blindness and disclosure of the biopsy results in the referral note were also in the exclusion criteria, but there was no such case, and there was no performance error for the loss of internet connection or other technical issues. Formal pathologic diagnosis (504 cases) was used as the ground truth; however, if the pathologic report consisted of a pathologic description only (i.e., lichenoid reaction), the pathologic diagnosis was determined by clinicopathological correlation (20 cases). Clinical diagnosis of the attending dermatologists was used as the ground truth for the 52 cases where biopsy was not performed because the attending dermatologists decided not to biopsy the definitely benign cases. Ultimately, 524 biopsy-proven cases and 52 clinically-diagnosed cases were included in the final analysis among the 603 cases of the initial recruitment (Table 1). As shown in Supplementary Tables S6 and S7, a total of 53 conditions were within the trained 178 classes (in-distribution) and 29 conditions were not trained by the algorithm (out-of-distribution). There were no cases in which train-test contamination could be concerned, although there were 15 cases with photographs taken previously.\n\nA total of four attending physicians (3, 4, 6, and 22 years of experience after board certification), four first-year dermatology residents, and four nondermatology trainees (doctors in their first year after getting a medical license who rotate between various departments) participated in this study. Attending physicians routinely recorded their impressions after thorough examinations. After the simple randomization using a custom randomizer (rand function of the PHP language) by the attending dermatologists, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and recorded their diagnostic hypothesis in real time.\n\nIn the AI group, trainees captured and selected 1-3 photographs with age and sex metadata as an input data and uploaded them to http://b2020.modelderm.com#world using internet browsers. Then the after-diagnoses was recorded, referring to the five diagnoses suggested by the algorithm and the malignancy score (Supplementary Figure S2). The photographs that the trainees thought to be of adequate quality and composition by themselves were uploaded. Trainees were instructed so that only macroimages having the large lesion of interest at the center are uploaded. In the unaided group, trainees performed routine examinations and recorded the three most probable diagnoses, without the assistance of the algorithm. The use of dermoscopy was not allowed for all trainees.\n\nStatistical analysis\nIn calculating the Top accuracy, only the exact diagnosis was recorded as correct, but giving a specific subtype of the disease was also counted to be correct. For example, intradermal nevus was counted correct for the ground truth of junctional nevus. We lumped together 364 diagnoses in natural language into the 82 diagnosis codes (https://doi.org/10.6084/m9.figshare.16640257). For evaluating a malignancy prediction, the physicians diagnoses were transformed into either malignant or benign. Top accuracies, sensitivities, and specificities were compared using Pearson\u2019s chi-square test with Yates\u2019 continuity correction (AI group vs. unaided group) or McNemar test (before vs. after the assistance of the algorithm) using R version 4.1.1, and P < 0.05 was statistically significant.\n\nData availability statement\nThe raw data of this study is available at https://doi.org/10.6084/m9.figshare.16640257. The diagnosis in natural language, converted diagnosis code (ALIAS sheet), the cases determined by clinicopathological correlation, accuracy/sensitivity/specificity/positive predictive value/negative predictive value of the participants and algorithm, and demographics of subjects are listed. The algorithm is accessible at https://b2020.modelderm.com/#world or https://modelderm.com (Build2020) for academic purposes, and only the images with the permission of the original owner should be uploaded.",
        "results": "### RESULTS\nThis study is reported per the Consolidated Standards of Reporting Trials\u2014Artificial Intelligence guidelines (Liu et al., 2020). Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u2019correct\u2019.\n\nAI group versus unaided group\nTo evaluate that the two groups were truly comparable, accuracies of the attending dermatologists and trainees (before interventions) were compared. Top-1 accuracy of attending dermatologists (62.3%) and trainees (43.8%) of the unaided group were higher than those of the AI group (dermatologists \u00bc 60.3%, trainees \u00bc 41.0%), which indicated that easier cases were not disproportionately allocated to the AI group (Table 1 and Figure 1). Overall, the Top-1 accuracy of the AI group was 53.9% and that of the unaided was 43.8% (all trainees; chi-square test, P \u00bc 0.019; Figure 2 and Table 2). There were significant differences in the result depending on whether the participant was a nondermatology trainee (general practitioner [GP]) or a dermatology resident (resident). The Top-1 accuracy of the AIGP group (54.7%) was markedly higher than that of unaidedGP group (29.7%; chi-square test, P < 0.0001), whereas the Top-1 accuracy of the AIresident group and the unaidedresident group was 53.1% and 57.3%, respectively (chi-square test, P \u00bc 0.55). In the AI group, we compared the judgment before and after receiving assistance of the algorithm and there was a significant enhancement in the Top 1 accuracy of the AIGP group (before augmentation \u00bc 30.7%, after augmentation \u00bc 54.7%; McNemar test, P < 0.0001). However, in the AIresident group, the change was not significant (before augmentation \u00bc 51.7%, after augmentation \u00bc 53.1%; McNemar test, P \u00bc 0.86). As shown in Supplementary Table S1, a greater improvement in accuracy was observed for the subsequent cases, up to \u00fe25.0% and \u00fe37.5% for Top-1 and Top-3 accuracy, respectively, although the accuracy for Top-1 and Top-3 was improved by only \u00fe0.0% and \u00fe15.0%, respectively, for the first five cases. When the analysis was restricted to 266 cases that were biopsied, the Top-1/Top-3 accuracy of the standalone algorithm, trainees before augmentation, trainees after augmentation, and attending dermatologists were 51.5%/75.9%, 40.2%/49.2%, 56.0%/71.4%, and 56.0%/66.2%, respectively. The accuracies of the AI-augmented trainees were equivalent to those of the attending dermatologists. In the 258 biopsy-proven cases of the unaided group, the Top-1/Top-3 accuracy of trainees and attending dermatologists were 42.6%/57.4% and 58.9%/68.6%, respectively. Malignancy determination affects clinical decisions such as ordering a biopsy if there is any malignancy among Top-3 predictions. Based on the Top-3 predictions, the sensitivity/specificity of the AI group and the unaided were 84.6%/69.5% and 75.6%/62.7%, respectively (chi-square test, P \u00bc 0.45/0.13; Table 2). The sensitivity/specificity of the AIGP group and the unaidedGP group were 80.0%/81.5% and 56.3%/68.9%, respectively (chi-square test, P \u00bc 0.24/0.029). The sensitivity/specificity of the AIresident group and the unaidedresident group were 89.5%/57.1% and 86.2%/56.1%, respectively (chi-square test, P \u00bc 1.0/0.98).\n\nThere was a significant difference in the number of differential diagnoses between the AI group and the unaided group (2.09 vs. 1.95; Wilcoxon rank-sum test, P \u00bc 0.0005). The number of differential diagnoses of the AIGP group (2.00) was higher than that of the unaidedGP group (1.88; Wilcoxon rank-sum test, P < 0.0001). The number of differential diagnoses of the AIresident group (2.17) was also higher than that of the unaidedresident group (2.01; Wilcoxon rank-sum test, P \u00bc 0.019).\n\nBefore and after comparison: Individual analysis\nThe Top-1 and Top-3 accuracy before the use of AI was 41.3% \u00b1 15.7% and 49.2% \u00b1 13.2%, respectively (8 trainees with 295 cases; Supplementary Table S2). The Top-1 and Top-3 accuracy after the assistance of AI was increased to 53.6% \u00b1 9.3% and 69.8% \u00b1 10.6%, respectively. Individual improvement in Top-1 diagnostic accuracy for each trainee ranged from 5.4% to \u00fe41.4%, with an average (SD) of \u00fe12.2% (16.6%) which was not statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.076). On the contrary, Top-3 accuracy was improved by \u00fe0%\u2013\u00fe41.4%, with an average improvement of \u00fe20.6% (12.9%), which was statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.0027). There was individual variation in the degree of improvement: the Top-1 accuracy of one dermatology resident decreased (\u22125.3%), whereas that of one nondermatology trainee was improved by \u00fe41.4%.\n\nStandalone performance of the algorithm\nThe standalone Top-1 and Top-3 accuracy of the algorithm in the AI group was 50.5% and 74.9%, respectively. Area under the curve for determining malignancy was 0.894 (95% confidence interval \u00bc 0.838\u20130.950; DeLong method), which was equivalent to that of the attending dermatologists on the receiver operating characteristic curve (Figure 3). At the high-sensitive threshold, the sensitivity and specificity were 92.3% and 62.1%, respectively, and at the high-specificity threshold, the sensitivity and specificity were 84.6% and 78.5%, respectively. The sensitivity/specificity of the standalone algorithm derived from the Top-1 was 66.7%/90.6% and that from the Top-3 was 84.6%/61.3% (Table 2).\n\nAdverse effects of the incorrect predictions from the algorithm\nA 12.2% drop in Top-1 accuracy of the trainees was observed in cases where all Top-3 predictions from the algorithm were incorrect (Supplementary Tables S3 and S4). In out-of-distribution cases where Top-3 predictions from the algorithm were inevitably incorrect, the Top-1 accuracy of the trainees dropped by 5.6%. Both sensitivity/specificity derived from Top-3 predictions of trainees were negatively affected (before augmentation \u00bc 84.6%/60.7% and after augmentation \u00bc 61.5%/59.0%) when all Top 3 predictions from the algorithm were incorrect.\n\nThere were four cases of malignancy (three basal cell carcinomas and one Bowen\u2019s disease) for which the trainees had initially included a malignant condition in their Top-3 but ruled out malignancy in their final Top-3 recorded after the use of the algorithm. Of these four cases, two were predicted to be conditions requiring follow-up visits (dysplastic nevus and actinic keratosis).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Neoplasms: A Single-Center, Paralleled, Unmasked, Randomized Controlled Trial.",
        "abstract": "## TRIAL DESIGN\nThis was a single-center, unmasked, paralleled, randomized controlled trial.\n## METHODS\nA randomized trial was conducted in a tertiary care institute in South Korea to validate whether artificial intelligence (AI) could augment the accuracy of nonexpert physicians in the real-world settings, which included diverse out-of-distribution conditions. Consecutive patients aged >19 years, having one or more skin lesions suspicious for skin cancer detected by either the patient or physician, were randomly allocated to four nondermatology trainees and four dermatology residents. The attending dermatologists examined the randomly allocated patients with (AI-assisted group) or without (unaided group) the real-time assistance of AI algorithm (https://b2020.modelderm.com#world; convolutional neural networks; unmasked design) after simple randomization of the patients.\n## RESULTS\nUsing 576 consecutive cases (Fitzpatrick skin phototypes III or IV) with suspicious lesions out of the initial 603 recruitments, the accuracy of the AI-assisted group (n\u00a0= 295, 53.9%) was found to be significantly higher than those of the unaided group (n\u00a0= 281, 43.8%; P\u00a0= 0.019). Whereas the augmentation was more significant from 54.7% (n\u00a0= 150) to 30.7% (n\u00a0= 138; P < 0.0001) in the nondermatology trainees who had the least experience in dermatology, it was not significant in the dermatology residents. The algorithm could help trainees in the AI-assisted group include more differential diagnoses than the unaided group (2.09 vs. 1.95 diagnoses; P\u00a0= 0.0005). However, a 12.2% drop in Top-1 accuracy of the trainees was observed in cases in which all Top-3 predictions given by the algorithm were incorrect.\n## CONCLUSIONS\nThe multiclass AI algorithm augmented the diagnostic accuracy of nonexpert physicians in dermatology.\n",
        "methods": "### METHODS\n\nThis prospective study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea) (S2018-1703-0001). It was performed in the Department of Dermatology at Asan Medical Center, a tertiary care center in Korea. The study was conducted from 30 November 2020 to 9 September 2021 after online registration (KCT0005614; cris.nih.go.kr). The development of the algorithm (Model Dermatology, Build2020; https://b2020.modelderm.com#world) is described in the Supplementary Materials, and the algorithm was fixed on 19 September 2020. Along with the prediction of five differential diagnoses, the algorithm reports a malignancy score (range = 0-100). The malignancy score was defined as the sum of malignant outputs and 0.2 \u00d7 premalignant outputs as used previously (Han et al., 2020b). Using the subset of the Seoul National University dataset (240 images; https://doi.org/10.6084/m9.figshare.6454973), the high-sensitivity threshold for determining malignancy was defined as the threshold at which 90% sensitivity was obtained because the sensitivity of the attending dermatologists was at the level of 88.1% (Han et al., 2020a). The high-specificity threshold was defined as the threshold at which 80% sensitivity was obtained.\n\nIn a previous pilot study (Kim et al., 2022), the Top-1 accuracy of trainees was 47.9%. If 25% enhancement after the assistance were regarded as significant, the sample size was calculated as 548 (alpha = 0.05 and power = 0.8), and we planned to recruit 600 cases (Rosner, 2011).\n\nAll patients signed an informed consent before inclusion in the study. We included consecutive patients (aged >19 years) who had one or more skin lesions suspicious for skin cancer, detected by either the patient or physician. Exclusion criteria of patients and input data included patient refusal (10 cases), wrong recruitment (6 cases; aged \u226419 years), biopsy refusal (2 cases), and nonreal-time analysis (9 cases). Broken blindness and disclosure of the biopsy results in the referral note were also in the exclusion criteria, but there was no such case, and there was no performance error for the loss of internet connection or other technical issues. Formal pathologic diagnosis (504 cases) was used as the ground truth; however, if the pathologic report consisted of a pathologic description only (i.e., lichenoid reaction), the pathologic diagnosis was determined by clinicopathological correlation (20 cases). Clinical diagnosis of the attending dermatologists was used as the ground truth for the 52 cases where biopsy was not performed because the attending dermatologists decided not to biopsy the definitely benign cases. Ultimately, 524 biopsy-proven cases and 52 clinically-diagnosed cases were included in the final analysis among the 603 cases of the initial recruitment (Table 1). As shown in Supplementary Tables S6 and S7, a total of 53 conditions were within the trained 178 classes (in-distribution) and 29 conditions were not trained by the algorithm (out-of-distribution). There were no cases in which train-test contamination could be concerned, although there were 15 cases with photographs taken previously.\n\nA total of four attending physicians (3, 4, 6, and 22 years of experience after board certification), four first-year dermatology residents, and four nondermatology trainees (doctors in their first year after getting a medical license who rotate between various departments) participated in this study. Attending physicians routinely recorded their impressions after thorough examinations. After the simple randomization using a custom randomizer (rand function of the PHP language) by the attending dermatologists, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and recorded their diagnostic hypothesis in real time.\n\nIn the AI group, trainees captured and selected 1-3 photographs with age and sex metadata as an input data and uploaded them to http://b2020.modelderm.com#world using internet browsers. Then the after-diagnoses was recorded, referring to the five diagnoses suggested by the algorithm and the malignancy score (Supplementary Figure S2). The photographs that the trainees thought to be of adequate quality and composition by themselves were uploaded. Trainees were instructed so that only macroimages having the large lesion of interest at the center are uploaded. In the unaided group, trainees performed routine examinations and recorded the three most probable diagnoses, without the assistance of the algorithm. The use of dermoscopy was not allowed for all trainees.\n\nStatistical analysis\nIn calculating the Top accuracy, only the exact diagnosis was recorded as correct, but giving a specific subtype of the disease was also counted to be correct. For example, intradermal nevus was counted correct for the ground truth of junctional nevus. We lumped together 364 diagnoses in natural language into the 82 diagnosis codes (https://doi.org/10.6084/m9.figshare.16640257). For evaluating a malignancy prediction, the physicians diagnoses were transformed into either malignant or benign. Top accuracies, sensitivities, and specificities were compared using Pearson\u2019s chi-square test with Yates\u2019 continuity correction (AI group vs. unaided group) or McNemar test (before vs. after the assistance of the algorithm) using R version 4.1.1, and P < 0.05 was statistically significant.\n\nData availability statement\nThe raw data of this study is available at https://doi.org/10.6084/m9.figshare.16640257. The diagnosis in natural language, converted diagnosis code (ALIAS sheet), the cases determined by clinicopathological correlation, accuracy/sensitivity/specificity/positive predictive value/negative predictive value of the participants and algorithm, and demographics of subjects are listed. The algorithm is accessible at https://b2020.modelderm.com/#world or https://modelderm.com (Build2020) for academic purposes, and only the images with the permission of the original owner should be uploaded.",
        "results": "### RESULTS\nThis study is reported per the Consolidated Standards of Reporting Trials\u2014Artificial Intelligence guidelines (Liu et al., 2020). Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u2019correct\u2019.\n\nAI group versus unaided group\nTo evaluate that the two groups were truly comparable, accuracies of the attending dermatologists and trainees (before interventions) were compared. Top-1 accuracy of attending dermatologists (62.3%) and trainees (43.8%) of the unaided group were higher than those of the AI group (dermatologists \u00bc 60.3%, trainees \u00bc 41.0%), which indicated that easier cases were not disproportionately allocated to the AI group (Table 1 and Figure 1). Overall, the Top-1 accuracy of the AI group was 53.9% and that of the unaided was 43.8% (all trainees; chi-square test, P \u00bc 0.019; Figure 2 and Table 2). There were significant differences in the result depending on whether the participant was a nondermatology trainee (general practitioner [GP]) or a dermatology resident (resident). The Top-1 accuracy of the AIGP group (54.7%) was markedly higher than that of unaidedGP group (29.7%; chi-square test, P < 0.0001), whereas the Top-1 accuracy of the AIresident group and the unaidedresident group was 53.1% and 57.3%, respectively (chi-square test, P \u00bc 0.55). In the AI group, we compared the judgment before and after receiving assistance of the algorithm and there was a significant enhancement in the Top 1 accuracy of the AIGP group (before augmentation \u00bc 30.7%, after augmentation \u00bc 54.7%; McNemar test, P < 0.0001). However, in the AIresident group, the change was not significant (before augmentation \u00bc 51.7%, after augmentation \u00bc 53.1%; McNemar test, P \u00bc 0.86). As shown in Supplementary Table S1, a greater improvement in accuracy was observed for the subsequent cases, up to \u00fe25.0% and \u00fe37.5% for Top-1 and Top-3 accuracy, respectively, although the accuracy for Top-1 and Top-3 was improved by only \u00fe0.0% and \u00fe15.0%, respectively, for the first five cases. When the analysis was restricted to 266 cases that were biopsied, the Top-1/Top-3 accuracy of the standalone algorithm, trainees before augmentation, trainees after augmentation, and attending dermatologists were 51.5%/75.9%, 40.2%/49.2%, 56.0%/71.4%, and 56.0%/66.2%, respectively. The accuracies of the AI-augmented trainees were equivalent to those of the attending dermatologists. In the 258 biopsy-proven cases of the unaided group, the Top-1/Top-3 accuracy of trainees and attending dermatologists were 42.6%/57.4% and 58.9%/68.6%, respectively. Malignancy determination affects clinical decisions such as ordering a biopsy if there is any malignancy among Top-3 predictions. Based on the Top-3 predictions, the sensitivity/specificity of the AI group and the unaided were 84.6%/69.5% and 75.6%/62.7%, respectively (chi-square test, P \u00bc 0.45/0.13; Table 2). The sensitivity/specificity of the AIGP group and the unaidedGP group were 80.0%/81.5% and 56.3%/68.9%, respectively (chi-square test, P \u00bc 0.24/0.029). The sensitivity/specificity of the AIresident group and the unaidedresident group were 89.5%/57.1% and 86.2%/56.1%, respectively (chi-square test, P \u00bc 1.0/0.98).\n\nThere was a significant difference in the number of differential diagnoses between the AI group and the unaided group (2.09 vs. 1.95; Wilcoxon rank-sum test, P \u00bc 0.0005). The number of differential diagnoses of the AIGP group (2.00) was higher than that of the unaidedGP group (1.88; Wilcoxon rank-sum test, P < 0.0001). The number of differential diagnoses of the AIresident group (2.17) was also higher than that of the unaidedresident group (2.01; Wilcoxon rank-sum test, P \u00bc 0.019).\n\nBefore and after comparison: Individual analysis\nThe Top-1 and Top-3 accuracy before the use of AI was 41.3% \u00b1 15.7% and 49.2% \u00b1 13.2%, respectively (8 trainees with 295 cases; Supplementary Table S2). The Top-1 and Top-3 accuracy after the assistance of AI was increased to 53.6% \u00b1 9.3% and 69.8% \u00b1 10.6%, respectively. Individual improvement in Top-1 diagnostic accuracy for each trainee ranged from 5.4% to \u00fe41.4%, with an average (SD) of \u00fe12.2% (16.6%) which was not statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.076). On the contrary, Top-3 accuracy was improved by \u00fe0%\u2013\u00fe41.4%, with an average improvement of \u00fe20.6% (12.9%), which was statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.0027). There was individual variation in the degree of improvement: the Top-1 accuracy of one dermatology resident decreased (\u22125.3%), whereas that of one nondermatology trainee was improved by \u00fe41.4%.\n\nStandalone performance of the algorithm\nThe standalone Top-1 and Top-3 accuracy of the algorithm in the AI group was 50.5% and 74.9%, respectively. Area under the curve for determining malignancy was 0.894 (95% confidence interval \u00bc 0.838\u20130.950; DeLong method), which was equivalent to that of the attending dermatologists on the receiver operating characteristic curve (Figure 3). At the high-sensitive threshold, the sensitivity and specificity were 92.3% and 62.1%, respectively, and at the high-specificity threshold, the sensitivity and specificity were 84.6% and 78.5%, respectively. The sensitivity/specificity of the standalone algorithm derived from the Top-1 was 66.7%/90.6% and that from the Top-3 was 84.6%/61.3% (Table 2).\n\nAdverse effects of the incorrect predictions from the algorithm\nA 12.2% drop in Top-1 accuracy of the trainees was observed in cases where all Top-3 predictions from the algorithm were incorrect (Supplementary Tables S3 and S4). In out-of-distribution cases where Top-3 predictions from the algorithm were inevitably incorrect, the Top-1 accuracy of the trainees dropped by 5.6%. Both sensitivity/specificity derived from Top-3 predictions of trainees were negatively affected (before augmentation \u00bc 84.6%/60.7% and after augmentation \u00bc 61.5%/59.0%) when all Top 3 predictions from the algorithm were incorrect.\n\nThere were four cases of malignancy (three basal cell carcinomas and one Bowen\u2019s disease) for which the trainees had initially included a malignant condition in their Top-3 but ruled out malignancy in their final Top-3 recorded after the use of the algorithm. Of these four cases, two were predicted to be conditions requiring follow-up visits (dysplastic nevus and actinic keratosis).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Unclear"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Neoplasms: A Single-Center, Paralleled, Unmasked, Randomized Controlled Trial.",
        "abstract": "## TRIAL DESIGN\nThis was a single-center, unmasked, paralleled, randomized controlled trial.\n## METHODS\nA randomized trial was conducted in a tertiary care institute in South Korea to validate whether artificial intelligence (AI) could augment the accuracy of nonexpert physicians in the real-world settings, which included diverse out-of-distribution conditions. Consecutive patients aged >19 years, having one or more skin lesions suspicious for skin cancer detected by either the patient or physician, were randomly allocated to four nondermatology trainees and four dermatology residents. The attending dermatologists examined the randomly allocated patients with (AI-assisted group) or without (unaided group) the real-time assistance of AI algorithm (https://b2020.modelderm.com#world; convolutional neural networks; unmasked design) after simple randomization of the patients.\n## RESULTS\nUsing 576 consecutive cases (Fitzpatrick skin phototypes III or IV) with suspicious lesions out of the initial 603 recruitments, the accuracy of the AI-assisted group (n\u00a0= 295, 53.9%) was found to be significantly higher than those of the unaided group (n\u00a0= 281, 43.8%; P\u00a0= 0.019). Whereas the augmentation was more significant from 54.7% (n\u00a0= 150) to 30.7% (n\u00a0= 138; P < 0.0001) in the nondermatology trainees who had the least experience in dermatology, it was not significant in the dermatology residents. The algorithm could help trainees in the AI-assisted group include more differential diagnoses than the unaided group (2.09 vs. 1.95 diagnoses; P\u00a0= 0.0005). However, a 12.2% drop in Top-1 accuracy of the trainees was observed in cases in which all Top-3 predictions given by the algorithm were incorrect.\n## CONCLUSIONS\nThe multiclass AI algorithm augmented the diagnostic accuracy of nonexpert physicians in dermatology.\n",
        "methods": "### METHODS\n\nThis prospective study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea) (S2018-1703-0001). It was performed in the Department of Dermatology at Asan Medical Center, a tertiary care center in Korea. The study was conducted from 30 November 2020 to 9 September 2021 after online registration (KCT0005614; cris.nih.go.kr). The development of the algorithm (Model Dermatology, Build2020; https://b2020.modelderm.com#world) is described in the Supplementary Materials, and the algorithm was fixed on 19 September 2020. Along with the prediction of five differential diagnoses, the algorithm reports a malignancy score (range = 0-100). The malignancy score was defined as the sum of malignant outputs and 0.2 \u00d7 premalignant outputs as used previously (Han et al., 2020b). Using the subset of the Seoul National University dataset (240 images; https://doi.org/10.6084/m9.figshare.6454973), the high-sensitivity threshold for determining malignancy was defined as the threshold at which 90% sensitivity was obtained because the sensitivity of the attending dermatologists was at the level of 88.1% (Han et al., 2020a). The high-specificity threshold was defined as the threshold at which 80% sensitivity was obtained.\n\nIn a previous pilot study (Kim et al., 2022), the Top-1 accuracy of trainees was 47.9%. If 25% enhancement after the assistance were regarded as significant, the sample size was calculated as 548 (alpha = 0.05 and power = 0.8), and we planned to recruit 600 cases (Rosner, 2011).\n\nAll patients signed an informed consent before inclusion in the study. We included consecutive patients (aged >19 years) who had one or more skin lesions suspicious for skin cancer, detected by either the patient or physician. Exclusion criteria of patients and input data included patient refusal (10 cases), wrong recruitment (6 cases; aged \u226419 years), biopsy refusal (2 cases), and nonreal-time analysis (9 cases). Broken blindness and disclosure of the biopsy results in the referral note were also in the exclusion criteria, but there was no such case, and there was no performance error for the loss of internet connection or other technical issues. Formal pathologic diagnosis (504 cases) was used as the ground truth; however, if the pathologic report consisted of a pathologic description only (i.e., lichenoid reaction), the pathologic diagnosis was determined by clinicopathological correlation (20 cases). Clinical diagnosis of the attending dermatologists was used as the ground truth for the 52 cases where biopsy was not performed because the attending dermatologists decided not to biopsy the definitely benign cases. Ultimately, 524 biopsy-proven cases and 52 clinically-diagnosed cases were included in the final analysis among the 603 cases of the initial recruitment (Table 1). As shown in Supplementary Tables S6 and S7, a total of 53 conditions were within the trained 178 classes (in-distribution) and 29 conditions were not trained by the algorithm (out-of-distribution). There were no cases in which train-test contamination could be concerned, although there were 15 cases with photographs taken previously.\n\nA total of four attending physicians (3, 4, 6, and 22 years of experience after board certification), four first-year dermatology residents, and four nondermatology trainees (doctors in their first year after getting a medical license who rotate between various departments) participated in this study. Attending physicians routinely recorded their impressions after thorough examinations. After the simple randomization using a custom randomizer (rand function of the PHP language) by the attending dermatologists, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and recorded their diagnostic hypothesis in real time.\n\nIn the AI group, trainees captured and selected 1-3 photographs with age and sex metadata as an input data and uploaded them to http://b2020.modelderm.com#world using internet browsers. Then the after-diagnoses was recorded, referring to the five diagnoses suggested by the algorithm and the malignancy score (Supplementary Figure S2). The photographs that the trainees thought to be of adequate quality and composition by themselves were uploaded. Trainees were instructed so that only macroimages having the large lesion of interest at the center are uploaded. In the unaided group, trainees performed routine examinations and recorded the three most probable diagnoses, without the assistance of the algorithm. The use of dermoscopy was not allowed for all trainees.\n\nStatistical analysis\nIn calculating the Top accuracy, only the exact diagnosis was recorded as correct, but giving a specific subtype of the disease was also counted to be correct. For example, intradermal nevus was counted correct for the ground truth of junctional nevus. We lumped together 364 diagnoses in natural language into the 82 diagnosis codes (https://doi.org/10.6084/m9.figshare.16640257). For evaluating a malignancy prediction, the physicians diagnoses were transformed into either malignant or benign. Top accuracies, sensitivities, and specificities were compared using Pearson\u2019s chi-square test with Yates\u2019 continuity correction (AI group vs. unaided group) or McNemar test (before vs. after the assistance of the algorithm) using R version 4.1.1, and P < 0.05 was statistically significant.\n\nData availability statement\nThe raw data of this study is available at https://doi.org/10.6084/m9.figshare.16640257. The diagnosis in natural language, converted diagnosis code (ALIAS sheet), the cases determined by clinicopathological correlation, accuracy/sensitivity/specificity/positive predictive value/negative predictive value of the participants and algorithm, and demographics of subjects are listed. The algorithm is accessible at https://b2020.modelderm.com/#world or https://modelderm.com (Build2020) for academic purposes, and only the images with the permission of the original owner should be uploaded.",
        "results": "### RESULTS\nThis study is reported per the Consolidated Standards of Reporting Trials\u2014Artificial Intelligence guidelines (Liu et al., 2020). Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u2019correct\u2019.\n\nAI group versus unaided group\nTo evaluate that the two groups were truly comparable, accuracies of the attending dermatologists and trainees (before interventions) were compared. Top-1 accuracy of attending dermatologists (62.3%) and trainees (43.8%) of the unaided group were higher than those of the AI group (dermatologists \u00bc 60.3%, trainees \u00bc 41.0%), which indicated that easier cases were not disproportionately allocated to the AI group (Table 1 and Figure 1). Overall, the Top-1 accuracy of the AI group was 53.9% and that of the unaided was 43.8% (all trainees; chi-square test, P \u00bc 0.019; Figure 2 and Table 2). There were significant differences in the result depending on whether the participant was a nondermatology trainee (general practitioner [GP]) or a dermatology resident (resident). The Top-1 accuracy of the AIGP group (54.7%) was markedly higher than that of unaidedGP group (29.7%; chi-square test, P < 0.0001), whereas the Top-1 accuracy of the AIresident group and the unaidedresident group was 53.1% and 57.3%, respectively (chi-square test, P \u00bc 0.55). In the AI group, we compared the judgment before and after receiving assistance of the algorithm and there was a significant enhancement in the Top 1 accuracy of the AIGP group (before augmentation \u00bc 30.7%, after augmentation \u00bc 54.7%; McNemar test, P < 0.0001). However, in the AIresident group, the change was not significant (before augmentation \u00bc 51.7%, after augmentation \u00bc 53.1%; McNemar test, P \u00bc 0.86). As shown in Supplementary Table S1, a greater improvement in accuracy was observed for the subsequent cases, up to \u00fe25.0% and \u00fe37.5% for Top-1 and Top-3 accuracy, respectively, although the accuracy for Top-1 and Top-3 was improved by only \u00fe0.0% and \u00fe15.0%, respectively, for the first five cases. When the analysis was restricted to 266 cases that were biopsied, the Top-1/Top-3 accuracy of the standalone algorithm, trainees before augmentation, trainees after augmentation, and attending dermatologists were 51.5%/75.9%, 40.2%/49.2%, 56.0%/71.4%, and 56.0%/66.2%, respectively. The accuracies of the AI-augmented trainees were equivalent to those of the attending dermatologists. In the 258 biopsy-proven cases of the unaided group, the Top-1/Top-3 accuracy of trainees and attending dermatologists were 42.6%/57.4% and 58.9%/68.6%, respectively. Malignancy determination affects clinical decisions such as ordering a biopsy if there is any malignancy among Top-3 predictions. Based on the Top-3 predictions, the sensitivity/specificity of the AI group and the unaided were 84.6%/69.5% and 75.6%/62.7%, respectively (chi-square test, P \u00bc 0.45/0.13; Table 2). The sensitivity/specificity of the AIGP group and the unaidedGP group were 80.0%/81.5% and 56.3%/68.9%, respectively (chi-square test, P \u00bc 0.24/0.029). The sensitivity/specificity of the AIresident group and the unaidedresident group were 89.5%/57.1% and 86.2%/56.1%, respectively (chi-square test, P \u00bc 1.0/0.98).\n\nThere was a significant difference in the number of differential diagnoses between the AI group and the unaided group (2.09 vs. 1.95; Wilcoxon rank-sum test, P \u00bc 0.0005). The number of differential diagnoses of the AIGP group (2.00) was higher than that of the unaidedGP group (1.88; Wilcoxon rank-sum test, P < 0.0001). The number of differential diagnoses of the AIresident group (2.17) was also higher than that of the unaidedresident group (2.01; Wilcoxon rank-sum test, P \u00bc 0.019).\n\nBefore and after comparison: Individual analysis\nThe Top-1 and Top-3 accuracy before the use of AI was 41.3% \u00b1 15.7% and 49.2% \u00b1 13.2%, respectively (8 trainees with 295 cases; Supplementary Table S2). The Top-1 and Top-3 accuracy after the assistance of AI was increased to 53.6% \u00b1 9.3% and 69.8% \u00b1 10.6%, respectively. Individual improvement in Top-1 diagnostic accuracy for each trainee ranged from 5.4% to \u00fe41.4%, with an average (SD) of \u00fe12.2% (16.6%) which was not statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.076). On the contrary, Top-3 accuracy was improved by \u00fe0%\u2013\u00fe41.4%, with an average improvement of \u00fe20.6% (12.9%), which was statistically significant (paired t-test after Shapiro-Wilk normality test, P \u00bc 0.0027). There was individual variation in the degree of improvement: the Top-1 accuracy of one dermatology resident decreased (\u22125.3%), whereas that of one nondermatology trainee was improved by \u00fe41.4%.\n\nStandalone performance of the algorithm\nThe standalone Top-1 and Top-3 accuracy of the algorithm in the AI group was 50.5% and 74.9%, respectively. Area under the curve for determining malignancy was 0.894 (95% confidence interval \u00bc 0.838\u20130.950; DeLong method), which was equivalent to that of the attending dermatologists on the receiver operating characteristic curve (Figure 3). At the high-sensitive threshold, the sensitivity and specificity were 92.3% and 62.1%, respectively, and at the high-specificity threshold, the sensitivity and specificity were 84.6% and 78.5%, respectively. The sensitivity/specificity of the standalone algorithm derived from the Top-1 was 66.7%/90.6% and that from the Top-3 was 84.6%/61.3% (Table 2).\n\nAdverse effects of the incorrect predictions from the algorithm\nA 12.2% drop in Top-1 accuracy of the trainees was observed in cases where all Top-3 predictions from the algorithm were incorrect (Supplementary Tables S3 and S4). In out-of-distribution cases where Top-3 predictions from the algorithm were inevitably incorrect, the Top-1 accuracy of the trainees dropped by 5.6%. Both sensitivity/specificity derived from Top-3 predictions of trainees were negatively affected (before augmentation \u00bc 84.6%/60.7% and after augmentation \u00bc 61.5%/59.0%) when all Top 3 predictions from the algorithm were incorrect.\n\nThere were four cases of malignancy (three basal cell carcinomas and one Bowen\u2019s disease) for which the trainees had initially included a malignant condition in their Top-3 but ruled out malignancy in their final Top-3 recorded after the use of the algorithm. Of these four cases, two were predicted to be conditions requiring follow-up visits (dysplastic nevus and actinic keratosis).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Over-Detection of Melanoma-Suspect Lesions by a CE-Certified Smartphone App: Performance in Comparison to Dermatologists, 2D and 3D Convolutional Neural Networks in a Prospective Data Set of 1204 Pigmented Skin Lesions Involving Patients' Perception.",
        "abstract": "The exponential increase in algorithm-based mobile health (mHealth) applications (apps) for melanoma screening is a reaction to a growing market. However, the performance of available apps remains to be investigated. In this prospective study, we investigated the diagnostic accuracy of a class 1 CE-certified smartphone app in melanoma risk stratification and its patient and dermatologist satisfaction. Pigmented skin lesions \u2265 3 mm and any suspicious smaller lesions were assessed by the smartphone app SkinVision\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Study Design and Participants_\n\nWe performed this prospective, single-center, comparative observational cohort study at the Department of Dermatology at the University Hospital Basel in Switzerland between January 2021 to June 2021. Seven dermatologists (beginners: <2 years professional experience, n = 4; intermediate: 2\u20135 years professional experience, n = 1; experts: >5 years professional experience, n = 2), as well as 114 patients were included in the study with an age of \u226518 years with a high risk for developing a primary cutaneous malignant melanoma defined as previous melanoma (including melanoma in situ), \u2265100 melanocytic nevi, \u22655 atypical nevi, strong family history for melanoma, diagnosis of dysplastic nevus syndrome, or known CDKN2A mutation. Exclusion criteria were acute psychiatric illness or acute crisis and Fitzpatrick sign type V-VI.\n\n_2.2. Procedures_\n\nIn the first step, the patients signed the informed consent after they were informed by a dermatologist about the study design and the following examinations. Subsequently, the dermatologist obtained a medical history, including a history of melanoma to enable classification into the different risk groups. The classification of all high-risk patients into further risk cohorts are defined by how often the patients require skin cancer screening. Based on this, we have classified our melanoma patients as high risk, as they come to dermatological control every 3\u201312 months according to the corresponding AJCC stage. Patients without melanoma in their own medical history are seen every 12 months. All study participants underwent a standard skin cancer screening by a dermatologist (Supplementary Figure S1). The pigmented skin lesions were assessed with dermatoscopy and classified as either benign or malignant. The dermatologist defined pigmented skin lesions that were suspected melanoma and indicated whether lesions smaller than 3 mm should be classified with the different AI modalities due to suspicion. The regular skin cancer screening was performed before any AI-based risk assessments were achieved, so the dermatologists had no knowledge about AI classification of the melanocytic lesions up to this point. In the next step, all melanocytic nevi \u2265 3 mm or any smaller suspicious melanocytic lesions were assessed using an iOS-based iPhone SE smartphone equipped with a 12-megapixel camera and the app SkinVision[\u00ae] (SkinVision[\u00ae] B.V., Amsterdam, the Netherlands, App-Version 6.8.1), which is based on a machine-learning algorithm classifying pigmented lesions into low, medium, and high risk for melanoma. Only the lesions classified as high-risk\u2014indicating a recommendation to consult a dermatologist\u2014were recorded. Afterwards, patients received a whole-body screening using the 2D automated total body mapping (ATBM[\u00ae]) master of FotoFinder (FotoFinder ATBM[\u00ae] Systems GmbH, Bad Birnbach, Germany, Version 3.3.1.0, risk scoring with MoleAnalyzer Pro), and the 3D TBP Vectra[\u00ae] WB360 System (Canfield Scientific, Parsippany, NJ, USA, Version 4.7.1, risk scoring with DEXI) which was performed by medical students and a study nurse. Subsequently, all lesions over 3 mm or any smaller suspicious lesions were assessed by the 2D and 3D AI devices using their dermoscopes (dermoscope medicam 1000 (FotoFinder ATBM[\u00ae]) and VISIOMED[\u00ae] D200evo dermoscope (Vectra[\u00ae] WB360)). For dermoscopic imaged melanocytic nevi, we obtained AI-based risk scores between 0.0\u20131.0 for 2D FotoFinder ATBM[\u00ae] and 0.0\u201310.0 for 3D TBP Vectra[\u00ae] WB360. FotoFinder\u2019s ATBM[\u00ae] master creates a 2D-image of the entire surface of the patient using 20 photos taken from 8 different parts of the body, while Vectra[\u00ae] WB360 generates a 3D-image using 92 individual photos that are converted into a 3D Avatar. Finally, the dermatologists assessed the skin lesions a second time with knowledge of the AI risk assessment scores of 2D FotoFinder ATBM[\u00ae] and 3D TBP Vectra[\u00ae] WB360. The patient was informed about the lesions with suspicion for melanoma by a dermatologist and about the 2D and 3D CNN classifications above the study-defined cut-off scores. The indication for excision of suspicious lesions was based on the dermatologists\u2019 suspicion for melanoma and/or malignancy risk assessment scores of the 2D imaging tool FotoFinder ATBM[\u00ae] and/or 3D imaging tool Vectra[\u00ae] WB360 indicating values over the study-defined cut-off score (>0.5 resp. >5.0). The patient was informed which lesions would be removed with the appropriate rationale. A maximum of two excisions were performed per visit and per patient due to ethical reasons, with an exception for further biopsies in cases of high-grade suspicion of malignancy by the dermatologist or patient request. The reference standard for a benign lesion was either the histological diagnosis (no melanoma) or the combination of the dermatologist\u2019s evaluation (benign lesion) plus the AI scores of two independent medical devices (FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360) each below the cut-off score. After all the assessments, the participating patients completed an anonymous 9-item questionnaire that surveyed their attitudes, personal preferences, and concerns about the use of AI in melanoma detection compared with skin cancer screenings by a dermatologist in addition to sociodemographic data. The dermatologists completed a 2-item questionnaire assessing their attitudes toward the smartphone app in melanoma screening, which was answered specifically for each of the 114 included patients. Neither the patients nor dermatologists were informed about the result of the app\u2019s risk assessment due to the extremely high number of false-positive findings to prevent response bias. Our survey was designed de novo after literature research. All answers were optional. The questionnaire was available in German, English and French and included binary questions (yes/no), multiple choice questions and visual analogue scales (VAS) with scores from 0\u201310.\n\n_2.3. Statistical Analysis_\n\nComparison of continuous variables was tested with the Wilcoxon rank test. Comparison of categorical variables was tested with Fishers Exact test. A p-value of less than 0.05 was deemed as significant. Receiver Operating Characteristics (ROC) analysis was used to assess the performance of the SkinVision[\u00ae] app against the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologist and the SkinVision[\u00ae] app against histology. All analyses were performed with R (version 4.1) and visualized with ggplot2.\n\n_2.4. Ethics_\n\nThe study was approved by the local ethics committee (22020-02482), registered with ClinicalTrials.gov (NCT04605822), and was conducted in compliance with the Declaration of Helsinki and Good Clinical Practice GCP-rules.",
        "results": "**3. Results**\n\n_3.1. Study Population_\n\nOverall, data from 114 patients were analyzed, including 55 patients at high-risk for developing a melanoma (mean age 55 years (age range): 22\u201385), 47% females) and 59 melanoma patients (mean age 60 years (age range: 29\u201381), 54% females) (Table 1). The family history for melanoma was positive in 19% of the melanoma patients and in 56% of the non-melanoma patients. Most of the patients used sunscreen SPF 30\u201350 (51% of melanoma patients vs. 62% of patients at high-risk for melanoma) and had previous sunburns in childhood (54% resp. 69%).\n\n**Table 1. Characteristics of the study population and their skin cancer awareness.**\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Age, n (age range)** 59 (22\u201385) 60 (29\u201381) 55 (22\u201385)\n**Sex, n (%)**\nFemale 58 (51%) 32 (54%) 26 (47%)\nMale 56 (49%) 27 (46%) 29 (53%)\n**Risk profile, n (%)**\nMultiple melanocytic nevi (\u2265100)\nand/or dysplastic nevi (\u22655) and/or positive\nfamily history for melanoma and/or 55 (48%) 0 (0%) 55 (100%)\ndiagnosis of dysplastic nevus syndrome\nand/or CDKN2A mutation\n\nPrevious resected melanoma in situ or\n57 (50%) 57 (97%) 0 (0%)\nprimary cutaneous melanoma\n\nMetastatic melanoma 2 (1.8%) 2 (3.4%) 0 (0%)\n**Positive family history for melanoma,**\n42 (37%) 11 (19%) 31 (56%)\n**n (%)**\n\n**Frequency of skin cancer screening,**\n**n (%)**\nSeveral times per year 40 (35%) 34 (58%) 6 (11%)\nEvery 12 months 39 (34%) 16 (27%) 23 (42%)\nEvery 1\u20132 years 8 (7%) 4 (6.8%) 4 (7.3%)\nEvery 2 years 9 (7.9%) 2 (3.4%) 7 (13%)\nLess than every 2 years 14 (12%) 3 (5.1%) 11 (20%)\nNever 4 (3.5%) 0 (0%) 4 (7.3%)\n**History of sunburns in childhood,**\n70 (61%) 32 (54%) 38 (69%)\n**n (%)**\n\n**Frequency of sunburns (Child), n (%)**\nRarely (less than once per year) 44 (63%) 20 (62%) 24 (63%)\nRegularly (once per year) 22 (31%) 10 (31%) 12 (32%)\nOften (more than once per year) 4 (5.7%) 2 (6.2%) 2 (5.3%)\n**History of sunburns in adulthood,**\n39 (34%) 18 (31%) 21 (38%)\n**n (%)**\n\n\n-----\n\n_Cancers 2022, 14, 3829_ 6 of 21\n\n**Table 1. Cont.**\n\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Frequency of sunburns (Adult), n (%)**\nRarely (less than once per year) 38 (97%) 18 (100%) 20 (95%)\nRegularly (once per year) 0 (0%) 0 (0%) 0 (0%)\nOften (more than once per year) 1 (2.6%) 0 (0%) 1 (4.8%)\n**Previous tanning in the solarium,**\n38 (33%) 13 (22%) 25 (45%)\n**n (%)**\n\n**Usage of sunscreen (SPF), n (%)**\nSPF 6\u201310 2 (1.8%) 1 (1.7%) 1 (1.8%)\nSPF 15\u201325 10 (8.8%) 3 (5.1%) 7 (13%)\nSPF 30\u201350 64 (56%) 30 (51%) 34 (62%)\nSPF 50+ 38 (33%) 25 (42%) 13 (24%)\n\n1 Median (Range); n (%).\n\nA total of 1204 pigmented skin lesions were assessed in this study (Figure 1). In 61 cases (5.1%), we performed a histopathology examination, while 1129 lesions (94%) were diagnosed to be clinically clearly benign based on the combination of a risk assessment by a dermatologist and AI-based risk scores below the cut-off (2D and 3D TBP) and had no indication for obtaining histopathology.\n\n**Figure 1. Flow chart of all included pigmented skin lesions and their histopathological outcome.**\nAI = artificial intelligence; TBP = total body photography.\n\n_3.2. Diagnostic Accuracy and Performance of the Smartphone App SkinVision[\u00ae]_\n\n3.2.1. Comparison of all Risk Assessments\n\nThe smartphone app SkinVision[\u00ae] classified 980 (81%) lesions as benign and indicated an increased risk for melanoma in 224 (19%) cases, while the dermatologists diagnosed 1195 (99.3%) lesions as benign and only nine (0.7%) as suspicious. Consequently, the CE-certified app had a 27-fold higher rate of melanoma-suspicious lesions compared to dermatologists. The AI scores of the 2D and 3D CNN devices classified most lesions as benign and 47 (3.9%) lesions (FotoFinder ATBM[\u00ae]) resp. 39 (3.2%) lesions (Vectra[\u00ae] WB360) as suspicious for melanoma (Figure 2).\n\n-----\n\n_Cancers 2022, 14, 3829_ 7 of 21\n\n**Figure 2. Comparison of risk assessments with the highest rate of suspected melanoma cases by the**\nsmartphone app (n = 1204).\n\nAmong the 224 lesions classified as suspicious by SkinVision[\u00ae], 193 were considered clinically benign by the physician and the 2D and 3D CNN AI-risk assessment scores, whereas four pigmented skin lesions were classified suspicious by the smartphone app, dermatologists, 2D, and 3D TBP devices (Table 2, Supplementary Table S1). The knowledge of AI-based risk assessment scores did not meaningfully affect dermatologists\u2019 classification of skin lesions, as they changed their decision towards the indication for excision in only three lesions that later turned out to be benign (Table 2).\n\n**Table 2. Risk assessments of 1204 pigmented skin lesions by the smartphone app SkinVision[\u00ae], 2D**\nimaging FotoFinder ATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in\ncombination with knowledge of FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360 AI-scores.\n\n**Characteristic** **N = 1204** **[1]**\n\n**Smartphone app SkinVision[\u00ae]**\n\nbenign 980 (81%)\nsuspicious 224 (19%)\n**2D Imaging FotoFinder ATBM[\u00ae]**\n\nbenign 1157 (96%)\nsuspicious 47 (3.9%)\n**3D Imaging VECTRA[\u00ae]** **WB360**\nbenign 1165 (97%)\nsuspicious 39 (3.2%)\n**Dermatologists**\nbenign 1195 (99%)\nsuspicious 9 (0.7%)\n**Dermatologists informed about risk assessment scores by**\n**FotoFinder ATBM[\u00ae]** **+ VECTRA[\u00ae]** **WB360**\nbenign 1192 (99%)\nsuspicious 12 (1.0%)\n\n1 n (%); AI = artificial intelligence.\n\n-----\n\n_Cancers 2022, 14, 3829_ 8 of 21\n\n3.2.2. Diagnostic Accuracy of the Smartphone App Based on the Combination of the\nDermatologist\u2019s Evaluation plus the AI Risk-Assessment Scores of Two Independent\nMedical Devices\n\nReceiver operating characteristic (ROC) analysis of the classification of benign and suspicious lesions of the SkinVision[\u00ae] app compared to the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologists had an area under the curve (AUC) score of 0.621, sensitivity of 0.41, and specificity of 0.83 (Figure 3). Although the specificity is reasonable, the sensitivity is low, thus suggesting that the SkinVision[\u00ae] app has poor diagnostic accuracy.\n\n**Figure 3. Receiver operating characteristic curve of the smartphone app in relation to the results**\nof the combination of risk assessments by dermatologists, FotoFinder ATBM[\u00ae], and Vectra[\u00ae] WB360\n(sensitivity: 41.3%, specificity: 82.9%); AUC = area under the curve.\n\n3.2.3. Diagnostic Accuracy of the Smartphone App Based on Histopathology\n\nAmong 61 pigmented skin lesions examined histologically, we detected six melanomas, 19 melanocytic nevi, 20 dysplastic nevi, as well as 16 otherwise classified lesions were diagnosed (Table 3, Figure 1).\n\n**Table 3. Diagnostic accuracy of the AI-based smartphone app SkinVision[\u00ae], 2D imaging FotoFinder**\nATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in combination with AI in\nmelanoma detection based on histopathology: sensitivity and specificity.\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**Smartphone app SkinVision[\u00ae]** 61\nbenign 13 (68%) 10 (50%) 1 (17%) 10 (62%)\nsuspicious 6 (32%) 10 (50%) 5 (83%) 6 (38%)\n**2D imaging FotoFinder ATBM[\u00ae]** 61\nbenign 7 (37%) 11 (55%) 1 (17%) 4 (25%)\nsuspicious 12 (63%) 9 (45%) 5 (83%) 12 (75%)\n\n-----\n\n_Cancers 2022, 14, 3829_ 9 of 21\n\n**Table 3. Cont.**\n\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**3D imaging VECTRA[\u00ae]** **WB360** 61\nbenign 18 (95%) 9 (45%) 1 (17%) 8 (50%)\nsuspicious 1 (5.3%) 11 (55%) 5 (83%) 8 (50%)\n**Dermatologists** 61\nbenign 17 (89%) 18 (90%) 1 (17%) 16 (100%)\nsuspicious 2 (11%) 2 (10%) 5 (83%) 0 (0%)\n_Beginner: <2 years\u2019 work_\n44 N = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 14 (93%) 10 (83%) 1 (20%) 13 (100%)\nsuspicious 1 (6.7%) 2 (17%) 4 (80%) 0 (0%)\n_Intermediate: 2\u20135 years\u2019 work_\n5 N = 2 N = 3 N = 0 N = 0\n_experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\n11 N = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n**Dermatologists informed about**\n61\n**AI scores** **[2]**\n\nbenign 16 (84%) 17 (85%) 1 (17%) 15 (94%)\nsuspicious 3 (16%) 3 (15%) 5 (83%) 1 (6.2%)\n_Beginner: <2 years\u2019 work_\nN = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 13 (87%) 9 (75%) 1 (20%) 12 (92%)\nsuspicious 2 (13%) 3 (25%) 4 (80%) 1 (7.7%)\n_Intermediate: 2\u20135 years\u2019_\nN = 2 N = 3 N = 0 N = 0\n_work experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\nN = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n\n1 n (%); * Other = pigmented basal cell carcinoma, histiocytoma, lentigo solaris, pigmented seborrhoic keratosis,\nfolliculitis with perifolliculitis, collisional tumor: seborrhoic keratosis and nevus, collisional tumor: actinic\nkeratosis and lentigo solaris; [2] risk assessment scores by FotoFinder ATBM[\u00ae] and VECTRA[\u00ae] WB360; AI = artificial\nintelligence.\n\nBased on histopathology, Figure 4 represents the risk assessments of the dermatologists, the mHealth app, and the combination of AI and dermatologists. Both the SkinVision[\u00ae]\n\napp and the 2D and 3D CNN devices indicated for 5 of 6 histological verified melanomas\n(83% sensitivity) an elevated score indicating suspicion of melanoma. The false-negative rate for all AI-based medical devices was 17%. The three different false-negative classified melanomas were all superficial spreading stage IA melanomas\u2014SkinVision[\u00ae]: superficial spreading melanoma, 0.9 mm Breslow thickness, AJCC stage IA (pT1bN0M0); 3D TBP Vectra[\u00ae] WB360: superficial spreading melanoma, 0.3 mm Breslow thickness, AJCC stage IA (pT1aN0M0); 2D TBP FotoFinder ATBM[\u00ae]: superficial spreading melanoma, 0.7 mm Breslow thickness, AJCC stage IA (pT1aN0M0). The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses was 60.0% of the smartphone app, 63.6% of the 3D Vectra[\u00ae] WB360 imaging device, and 40.0% of the 2D FotoFinder ATBM[\u00ae] risk assessment tool. Dermatologists correctly identified five of six melanomas (83% sensitivity). The false-negative rate was 17%. The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses among all dermatologists was 92.7%. The performance of the dermatologists strongly correlated with their professional experience. Hence, the true-positive rates for melanoma were greater for experts and dermatologists with intermediate experience (100% sensitivity) compared to beginners (80% sensitivity); whereas the true-negative rates were similar (93.3% resp. 92.5% specificity) (Table 3).\n\n**Figure 4. Histology and corresponding diagnosis of lesions assessed by dermatologists, smartphone**\napp SkinVision[\u00ae], and dermatologists and AI (n = 61).\n\nFigure 5 reveals several pigmented skin lesions that were correctly and incorrectly classified as melanomas by the smartphone app SkinVision[\u00ae].\n\n**Figure 5. Correctly and falsely classified melanomas by the smartphone app SkinVision[\u00ae]: (A,B). True-**\npositive classified melanoma; (C,D). False-positive classified melanoma; (E,F). True-negative classified\nmelanoma; (G). False-negative classified melanoma.\n\nROC analysis of the classification of melanoma of the SkinVision[\u00ae] app compared\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision[\u00ae] app had\nhigh sensitivity but mediocre specificity.\n\nROC analysis of the classification of melanoma of the SkinVision\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision\nhigh sensitivity but mediocre specificity.\n\n-----\n\n_Cancers 2022, 14, 3829_ 11 of 21\n\n**Figure 6. Receiver operating characteristic curve of the smartphone app in relation to the results of**\nthe histology (sensitivity of 83.3%, specificity 60.0%); AUC = area under the curve.\n\n_3.3. Patient Perspective on AI in Melanoma Screening_\n3.3.1. Confidence in Dermatologists vs. Smartphone App\n\nMost of the patients at high-risk for melanoma (55% (30/55)) and patients with\nmelanoma (53% (31/59)) reported being very confident about a mole examination by\na dermatologist (rating scale 10/10). In contrast, only a minority of high-risk patients (16%\n(9/55)) and melanoma patients (12% (7/59)) felt very safe when being investigated by a\nsmartphone app alone. No significant difference in ratings was identified between the two\nrisk groups (p < 0.9; p < 0.7) (Supplementary Table S2).\n\n3.3.2. Trustworthiness of the Smartphone App\n\nAll participants primarily rated physician examination (100% (55/55) among highrisk patients, 100% (59/59) among melanoma patients) as trustworthy, and the majority\ndid so for 2D TBP imaging (93% (51/55) resp. 88% (52/59)) and 3D TBP imaging (91%\n(50/55) resp. 90% (53/59)). The smartphone app was less frequently rated as trustworthy,\nwith only 36% (20/55) among patients at high-risk for melanoma and 49% (29/59) among\nmelanoma patients (Table 4). The age revealed a significant correlation with the evaluation\nof smartphones\u2019 trustworthiness (p < 0.004), with older patients (>60 years old) having\ntrusted the app three times more than younger patients (\u226460 years old) (Figure 7). Neither\nprevious melanoma vs. high-risk criteria for melanoma nor sex significantly influenced the\nevaluation of the trustworthiness of the smartphone app.\n\n-----\n\n_Cancers 2022, 14, 3829_ 12 of 21\n\n**Table 4. Assessment of trustworthiness of the AI-based smartphone app SkinVision[\u00ae], 2D imaging**\nFotoFinder ATBM[\u00ae], and 3D imaging Vectra[\u00ae] WB360 compared to dermatologists.\n\n\n**Patients at**\n**High-Risk**\n**for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Characteristic** **N**\n\n\n**Patients**\n**with**\n**Melanoma,**\n**N = 59** **[1]**\n\n\n**_p-Value_** **[2]**\n\n\n**The following examination was**\n**trustworthy: Smartphone** 114 0.3\n**app assessment**\n\nYes 29 (49%) 20 (36%)\nNo 5 (8.5%) 8 (15%)\nI don\u2019t know 23 (39%) 22 (40%)\nNo answer 2 (3.4%) 5 (9.1%)\n**Dermatologist assessment** 114\nYes 59 (100%) 55 (100%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 0 (0%) 0 (0%)\nNo answer 0 (0%) 0 (0%)\n**2D TBP assessment** 114 0.3\nYes 52 (88%) 51 (93%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 7 (12%) 3 (5.5%)\nNo answer 0 (0%) 1 (1.8%)\n**3D TBP assessment** 114 0.3\nYes 53 (90%) 50 (91%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 6 (10%) 3 (5.5%)\nNo answer 0 (0%) 2 (3.6%)\n\n1 n (%); 2 Fisher\u2019s exact test; Pearson\u2019s Chi-squared test; TBP = total body photography.\n\n**Figure 7. Odds ratio for variables influencing the trustworthiness of smartphones\u2019 risk assessment in**\nmelanoma detection.\n\n-----\n\n_Cancers 2022, 14, 3829_ 13 of 21\n\n3.3.3. Impact of AI vs. Dermatologists\u2019 Examination on Patients\u2019 Fear of Developing\nSkin Cancer\n\nMost participants indicated that an examination by a physician reduced their fear of\ndeveloping skin cancer, namely in 89% (49/55) among high-risk patients and 81% (48/59)\namong melanoma patients. Comparably, the 2D TBP imaging achieved the same effect in\n78% (43/55) resp. 76% (45/59), and the 3D TBP device in 82% (45/55) resp. 75% (44/59). In\ncontrast, the assessment with the smartphone app appeased fear of skin cancer in only 33%\n(18/55) of patients at high-risk for melanoma and in 32% (19/59) of patients with melanoma.\nIndeed, one high-risk patient (1.8%) even reported an increased fear of developing skin\ncancer by using the smartphone app (Supplementary Table S2).\n\n3.3.4. Patients\u2019 Subjective Assessment of the Accuracy of AI vs. Dermatologists\n\nPatients expected reliable results with the highest accuracy by both the assessment\nby a physician (98% (54/55) among high-risk patients, 92% (54/59) among melanoma\npatients) and by the 2D TBP imaging (82% (45/55) resp. 86% (51/59)) as well as the 3D\nTBP device (89% (49/55) resp. 88% (52/59)). Only 16% (9/55) of high-risk patients and\n31% (18/59) of melanoma patients expected reliable results from the smartphone app\n(Supplementary Table S2). No significant differences were identified between the two risk\ngroups in the evaluations.\n\n3.3.5. Patient Preference for Skin Cancer Screening\n\nBoth cohorts favored a combination of dermatologist and 3D TBP risk assessment for\nthe examination of pigmented skin lesions (64% (35/55) among patients at high-risk for\nmelanoma, 51% (30/59) among melanoma patients), while neither preferred assessment\nby a smartphone app alone (Figure 8A, Supplementary Table S2). The combination of\ndermatologist and smartphone app was favored by only 1.8% (1/55) of patients at high-risk\nfor melanoma and 3.4% (2/59) of patients with melanoma.\n\n**Figure 8. Cont.**\n\n-----\n\n_Cancers 2022, 14, 3829_ 14 of 21\n\n**Figure 8. (A) Patient preference for mole assessment (patients at high-risk for melanoma, n = 55;**\npatients with melanoma, n = 59); (B) Patient preference for AI in skin cancer screening (patients at\nhigh-risk for melanoma, n = 55; patients with melanoma, n = 59).\n\nRegarding patient preference for skin cancer screenings, almost all high-risk patients\n(98% (54/55)) and melanoma patients (95% (56/59)) indicated their belief that AI can improve\na physician\u2019s diagnostic performance. Most patients (64% (35/55) among high-risk patients,\n54% (32/59) among melanoma patients) would prefer that the physicians would always\nconsider the result of AI in their diagnosis (Figure 8B) (Supplementary Table S2).\n\n3.3.6. Dermatologists\u2019 Perspective of Smartphone Apps for Melanoma Screening\n\nAmong the seven dermatologists, they stated in only 5.3% of the skin cancer screenings\n(6/114) that the smartphone app increased diagnostic confidence and in only 8.8% of the\nassessments (10/114) they trusted the app (Supplementary Table S3).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "High"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Over-Detection of Melanoma-Suspect Lesions by a CE-Certified Smartphone App: Performance in Comparison to Dermatologists, 2D and 3D Convolutional Neural Networks in a Prospective Data Set of 1204 Pigmented Skin Lesions Involving Patients' Perception.",
        "abstract": "The exponential increase in algorithm-based mobile health (mHealth) applications (apps) for melanoma screening is a reaction to a growing market. However, the performance of available apps remains to be investigated. In this prospective study, we investigated the diagnostic accuracy of a class 1 CE-certified smartphone app in melanoma risk stratification and its patient and dermatologist satisfaction. Pigmented skin lesions \u2265 3 mm and any suspicious smaller lesions were assessed by the smartphone app SkinVision\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Study Design and Participants_\n\nWe performed this prospective, single-center, comparative observational cohort study at the Department of Dermatology at the University Hospital Basel in Switzerland between January 2021 to June 2021. Seven dermatologists (beginners: <2 years professional experience, n = 4; intermediate: 2\u20135 years professional experience, n = 1; experts: >5 years professional experience, n = 2), as well as 114 patients were included in the study with an age of \u226518 years with a high risk for developing a primary cutaneous malignant melanoma defined as previous melanoma (including melanoma in situ), \u2265100 melanocytic nevi, \u22655 atypical nevi, strong family history for melanoma, diagnosis of dysplastic nevus syndrome, or known CDKN2A mutation. Exclusion criteria were acute psychiatric illness or acute crisis and Fitzpatrick sign type V-VI.\n\n_2.2. Procedures_\n\nIn the first step, the patients signed the informed consent after they were informed by a dermatologist about the study design and the following examinations. Subsequently, the dermatologist obtained a medical history, including a history of melanoma to enable classification into the different risk groups. The classification of all high-risk patients into further risk cohorts are defined by how often the patients require skin cancer screening. Based on this, we have classified our melanoma patients as high risk, as they come to dermatological control every 3\u201312 months according to the corresponding AJCC stage. Patients without melanoma in their own medical history are seen every 12 months. All study participants underwent a standard skin cancer screening by a dermatologist (Supplementary Figure S1). The pigmented skin lesions were assessed with dermatoscopy and classified as either benign or malignant. The dermatologist defined pigmented skin lesions that were suspected melanoma and indicated whether lesions smaller than 3 mm should be classified with the different AI modalities due to suspicion. The regular skin cancer screening was performed before any AI-based risk assessments were achieved, so the dermatologists had no knowledge about AI classification of the melanocytic lesions up to this point. In the next step, all melanocytic nevi \u2265 3 mm or any smaller suspicious melanocytic lesions were assessed using an iOS-based iPhone SE smartphone equipped with a 12-megapixel camera and the app SkinVision[\u00ae] (SkinVision[\u00ae] B.V., Amsterdam, the Netherlands, App-Version 6.8.1), which is based on a machine-learning algorithm classifying pigmented lesions into low, medium, and high risk for melanoma. Only the lesions classified as high-risk\u2014indicating a recommendation to consult a dermatologist\u2014were recorded. Afterwards, patients received a whole-body screening using the 2D automated total body mapping (ATBM[\u00ae]) master of FotoFinder (FotoFinder ATBM[\u00ae] Systems GmbH, Bad Birnbach, Germany, Version 3.3.1.0, risk scoring with MoleAnalyzer Pro), and the 3D TBP Vectra[\u00ae] WB360 System (Canfield Scientific, Parsippany, NJ, USA, Version 4.7.1, risk scoring with DEXI) which was performed by medical students and a study nurse. Subsequently, all lesions over 3 mm or any smaller suspicious lesions were assessed by the 2D and 3D AI devices using their dermoscopes (dermoscope medicam 1000 (FotoFinder ATBM[\u00ae]) and VISIOMED[\u00ae] D200evo dermoscope (Vectra[\u00ae] WB360)). For dermoscopic imaged melanocytic nevi, we obtained AI-based risk scores between 0.0\u20131.0 for 2D FotoFinder ATBM[\u00ae] and 0.0\u201310.0 for 3D TBP Vectra[\u00ae] WB360. FotoFinder\u2019s ATBM[\u00ae] master creates a 2D-image of the entire surface of the patient using 20 photos taken from 8 different parts of the body, while Vectra[\u00ae] WB360 generates a 3D-image using 92 individual photos that are converted into a 3D Avatar. Finally, the dermatologists assessed the skin lesions a second time with knowledge of the AI risk assessment scores of 2D FotoFinder ATBM[\u00ae] and 3D TBP Vectra[\u00ae] WB360. The patient was informed about the lesions with suspicion for melanoma by a dermatologist and about the 2D and 3D CNN classifications above the study-defined cut-off scores. The indication for excision of suspicious lesions was based on the dermatologists\u2019 suspicion for melanoma and/or malignancy risk assessment scores of the 2D imaging tool FotoFinder ATBM[\u00ae] and/or 3D imaging tool Vectra[\u00ae] WB360 indicating values over the study-defined cut-off score (>0.5 resp. >5.0). The patient was informed which lesions would be removed with the appropriate rationale. A maximum of two excisions were performed per visit and per patient due to ethical reasons, with an exception for further biopsies in cases of high-grade suspicion of malignancy by the dermatologist or patient request. The reference standard for a benign lesion was either the histological diagnosis (no melanoma) or the combination of the dermatologist\u2019s evaluation (benign lesion) plus the AI scores of two independent medical devices (FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360) each below the cut-off score. After all the assessments, the participating patients completed an anonymous 9-item questionnaire that surveyed their attitudes, personal preferences, and concerns about the use of AI in melanoma detection compared with skin cancer screenings by a dermatologist in addition to sociodemographic data. The dermatologists completed a 2-item questionnaire assessing their attitudes toward the smartphone app in melanoma screening, which was answered specifically for each of the 114 included patients. Neither the patients nor dermatologists were informed about the result of the app\u2019s risk assessment due to the extremely high number of false-positive findings to prevent response bias. Our survey was designed de novo after literature research. All answers were optional. The questionnaire was available in German, English and French and included binary questions (yes/no), multiple choice questions and visual analogue scales (VAS) with scores from 0\u201310.\n\n_2.3. Statistical Analysis_\n\nComparison of continuous variables was tested with the Wilcoxon rank test. Comparison of categorical variables was tested with Fishers Exact test. A p-value of less than 0.05 was deemed as significant. Receiver Operating Characteristics (ROC) analysis was used to assess the performance of the SkinVision[\u00ae] app against the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologist and the SkinVision[\u00ae] app against histology. All analyses were performed with R (version 4.1) and visualized with ggplot2.\n\n_2.4. Ethics_\n\nThe study was approved by the local ethics committee (22020-02482), registered with ClinicalTrials.gov (NCT04605822), and was conducted in compliance with the Declaration of Helsinki and Good Clinical Practice GCP-rules.",
        "results": "**3. Results**\n\n_3.1. Study Population_\n\nOverall, data from 114 patients were analyzed, including 55 patients at high-risk for developing a melanoma (mean age 55 years (age range): 22\u201385), 47% females) and 59 melanoma patients (mean age 60 years (age range: 29\u201381), 54% females) (Table 1). The family history for melanoma was positive in 19% of the melanoma patients and in 56% of the non-melanoma patients. Most of the patients used sunscreen SPF 30\u201350 (51% of melanoma patients vs. 62% of patients at high-risk for melanoma) and had previous sunburns in childhood (54% resp. 69%).\n\n**Table 1. Characteristics of the study population and their skin cancer awareness.**\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Age, n (age range)** 59 (22\u201385) 60 (29\u201381) 55 (22\u201385)\n**Sex, n (%)**\nFemale 58 (51%) 32 (54%) 26 (47%)\nMale 56 (49%) 27 (46%) 29 (53%)\n**Risk profile, n (%)**\nMultiple melanocytic nevi (\u2265100)\nand/or dysplastic nevi (\u22655) and/or positive\nfamily history for melanoma and/or 55 (48%) 0 (0%) 55 (100%)\ndiagnosis of dysplastic nevus syndrome\nand/or CDKN2A mutation\n\nPrevious resected melanoma in situ or\n57 (50%) 57 (97%) 0 (0%)\nprimary cutaneous melanoma\n\nMetastatic melanoma 2 (1.8%) 2 (3.4%) 0 (0%)\n**Positive family history for melanoma,**\n42 (37%) 11 (19%) 31 (56%)\n**n (%)**\n\n**Frequency of skin cancer screening,**\n**n (%)**\nSeveral times per year 40 (35%) 34 (58%) 6 (11%)\nEvery 12 months 39 (34%) 16 (27%) 23 (42%)\nEvery 1\u20132 years 8 (7%) 4 (6.8%) 4 (7.3%)\nEvery 2 years 9 (7.9%) 2 (3.4%) 7 (13%)\nLess than every 2 years 14 (12%) 3 (5.1%) 11 (20%)\nNever 4 (3.5%) 0 (0%) 4 (7.3%)\n**History of sunburns in childhood,**\n70 (61%) 32 (54%) 38 (69%)\n**n (%)**\n\n**Frequency of sunburns (Child), n (%)**\nRarely (less than once per year) 44 (63%) 20 (62%) 24 (63%)\nRegularly (once per year) 22 (31%) 10 (31%) 12 (32%)\nOften (more than once per year) 4 (5.7%) 2 (6.2%) 2 (5.3%)\n**History of sunburns in adulthood,**\n39 (34%) 18 (31%) 21 (38%)\n**n (%)**\n\n\n-----\n\n_Cancers 2022, 14, 3829_ 6 of 21\n\n**Table 1. Cont.**\n\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Frequency of sunburns (Adult), n (%)**\nRarely (less than once per year) 38 (97%) 18 (100%) 20 (95%)\nRegularly (once per year) 0 (0%) 0 (0%) 0 (0%)\nOften (more than once per year) 1 (2.6%) 0 (0%) 1 (4.8%)\n**Previous tanning in the solarium,**\n38 (33%) 13 (22%) 25 (45%)\n**n (%)**\n\n**Usage of sunscreen (SPF), n (%)**\nSPF 6\u201310 2 (1.8%) 1 (1.7%) 1 (1.8%)\nSPF 15\u201325 10 (8.8%) 3 (5.1%) 7 (13%)\nSPF 30\u201350 64 (56%) 30 (51%) 34 (62%)\nSPF 50+ 38 (33%) 25 (42%) 13 (24%)\n\n1 Median (Range); n (%).\n\nA total of 1204 pigmented skin lesions were assessed in this study (Figure 1). In 61 cases (5.1%), we performed a histopathology examination, while 1129 lesions (94%) were diagnosed to be clinically clearly benign based on the combination of a risk assessment by a dermatologist and AI-based risk scores below the cut-off (2D and 3D TBP) and had no indication for obtaining histopathology.\n\n**Figure 1. Flow chart of all included pigmented skin lesions and their histopathological outcome.**\nAI = artificial intelligence; TBP = total body photography.\n\n_3.2. Diagnostic Accuracy and Performance of the Smartphone App SkinVision[\u00ae]_\n\n3.2.1. Comparison of all Risk Assessments\n\nThe smartphone app SkinVision[\u00ae] classified 980 (81%) lesions as benign and indicated an increased risk for melanoma in 224 (19%) cases, while the dermatologists diagnosed 1195 (99.3%) lesions as benign and only nine (0.7%) as suspicious. Consequently, the CE-certified app had a 27-fold higher rate of melanoma-suspicious lesions compared to dermatologists. The AI scores of the 2D and 3D CNN devices classified most lesions as benign and 47 (3.9%) lesions (FotoFinder ATBM[\u00ae]) resp. 39 (3.2%) lesions (Vectra[\u00ae] WB360) as suspicious for melanoma (Figure 2).\n\n-----\n\n_Cancers 2022, 14, 3829_ 7 of 21\n\n**Figure 2. Comparison of risk assessments with the highest rate of suspected melanoma cases by the**\nsmartphone app (n = 1204).\n\nAmong the 224 lesions classified as suspicious by SkinVision[\u00ae], 193 were considered clinically benign by the physician and the 2D and 3D CNN AI-risk assessment scores, whereas four pigmented skin lesions were classified suspicious by the smartphone app, dermatologists, 2D, and 3D TBP devices (Table 2, Supplementary Table S1). The knowledge of AI-based risk assessment scores did not meaningfully affect dermatologists\u2019 classification of skin lesions, as they changed their decision towards the indication for excision in only three lesions that later turned out to be benign (Table 2).\n\n**Table 2. Risk assessments of 1204 pigmented skin lesions by the smartphone app SkinVision[\u00ae], 2D**\nimaging FotoFinder ATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in\ncombination with knowledge of FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360 AI-scores.\n\n**Characteristic** **N = 1204** **[1]**\n\n**Smartphone app SkinVision[\u00ae]**\n\nbenign 980 (81%)\nsuspicious 224 (19%)\n**2D Imaging FotoFinder ATBM[\u00ae]**\n\nbenign 1157 (96%)\nsuspicious 47 (3.9%)\n**3D Imaging VECTRA[\u00ae]** **WB360**\nbenign 1165 (97%)\nsuspicious 39 (3.2%)\n**Dermatologists**\nbenign 1195 (99%)\nsuspicious 9 (0.7%)\n**Dermatologists informed about risk assessment scores by**\n**FotoFinder ATBM[\u00ae]** **+ VECTRA[\u00ae]** **WB360**\nbenign 1192 (99%)\nsuspicious 12 (1.0%)\n\n1 n (%); AI = artificial intelligence.\n\n-----\n\n_Cancers 2022, 14, 3829_ 8 of 21\n\n3.2.2. Diagnostic Accuracy of the Smartphone App Based on the Combination of the\nDermatologist\u2019s Evaluation plus the AI Risk-Assessment Scores of Two Independent\nMedical Devices\n\nReceiver operating characteristic (ROC) analysis of the classification of benign and suspicious lesions of the SkinVision[\u00ae] app compared to the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologists had an area under the curve (AUC) score of 0.621, sensitivity of 0.41, and specificity of 0.83 (Figure 3). Although the specificity is reasonable, the sensitivity is low, thus suggesting that the SkinVision[\u00ae] app has poor diagnostic accuracy.\n\n**Figure 3. Receiver operating characteristic curve of the smartphone app in relation to the results**\nof the combination of risk assessments by dermatologists, FotoFinder ATBM[\u00ae], and Vectra[\u00ae] WB360\n(sensitivity: 41.3%, specificity: 82.9%); AUC = area under the curve.\n\n3.2.3. Diagnostic Accuracy of the Smartphone App Based on Histopathology\n\nAmong 61 pigmented skin lesions examined histologically, we detected six melanomas, 19 melanocytic nevi, 20 dysplastic nevi, as well as 16 otherwise classified lesions were diagnosed (Table 3, Figure 1).\n\n**Table 3. Diagnostic accuracy of the AI-based smartphone app SkinVision[\u00ae], 2D imaging FotoFinder**\nATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in combination with AI in\nmelanoma detection based on histopathology: sensitivity and specificity.\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**Smartphone app SkinVision[\u00ae]** 61\nbenign 13 (68%) 10 (50%) 1 (17%) 10 (62%)\nsuspicious 6 (32%) 10 (50%) 5 (83%) 6 (38%)\n**2D imaging FotoFinder ATBM[\u00ae]** 61\nbenign 7 (37%) 11 (55%) 1 (17%) 4 (25%)\nsuspicious 12 (63%) 9 (45%) 5 (83%) 12 (75%)\n\n-----\n\n_Cancers 2022, 14, 3829_ 9 of 21\n\n**Table 3. Cont.**\n\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**3D imaging VECTRA[\u00ae]** **WB360** 61\nbenign 18 (95%) 9 (45%) 1 (17%) 8 (50%)\nsuspicious 1 (5.3%) 11 (55%) 5 (83%) 8 (50%)\n**Dermatologists** 61\nbenign 17 (89%) 18 (90%) 1 (17%) 16 (100%)\nsuspicious 2 (11%) 2 (10%) 5 (83%) 0 (0%)\n_Beginner: <2 years\u2019 work_\n44 N = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 14 (93%) 10 (83%) 1 (20%) 13 (100%)\nsuspicious 1 (6.7%) 2 (17%) 4 (80%) 0 (0%)\n_Intermediate: 2\u20135 years\u2019 work_\n5 N = 2 N = 3 N = 0 N = 0\n_experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\n11 N = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n**Dermatologists informed about**\n61\n**AI scores** **[2]**\n\nbenign 16 (84%) 17 (85%) 1 (17%) 15 (94%)\nsuspicious 3 (16%) 3 (15%) 5 (83%) 1 (6.2%)\n_Beginner: <2 years\u2019 work_\nN = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 13 (87%) 9 (75%) 1 (20%) 12 (92%)\nsuspicious 2 (13%) 3 (25%) 4 (80%) 1 (7.7%)\n_Intermediate: 2\u20135 years\u2019_\nN = 2 N = 3 N = 0 N = 0\n_work experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\nN = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n\n1 n (%); * Other = pigmented basal cell carcinoma, histiocytoma, lentigo solaris, pigmented seborrhoic keratosis,\nfolliculitis with perifolliculitis, collisional tumor: seborrhoic keratosis and nevus, collisional tumor: actinic\nkeratosis and lentigo solaris; [2] risk assessment scores by FotoFinder ATBM[\u00ae] and VECTRA[\u00ae] WB360; AI = artificial\nintelligence.\n\nBased on histopathology, Figure 4 represents the risk assessments of the dermatologists, the mHealth app, and the combination of AI and dermatologists. Both the SkinVision[\u00ae]\n\napp and the 2D and 3D CNN devices indicated for 5 of 6 histological verified melanomas\n(83% sensitivity) an elevated score indicating suspicion of melanoma. The false-negative rate for all AI-based medical devices was 17%. The three different false-negative classified melanomas were all superficial spreading stage IA melanomas\u2014SkinVision[\u00ae]: superficial spreading melanoma, 0.9 mm Breslow thickness, AJCC stage IA (pT1bN0M0); 3D TBP Vectra[\u00ae] WB360: superficial spreading melanoma, 0.3 mm Breslow thickness, AJCC stage IA (pT1aN0M0); 2D TBP FotoFinder ATBM[\u00ae]: superficial spreading melanoma, 0.7 mm Breslow thickness, AJCC stage IA (pT1aN0M0). The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses was 60.0% of the smartphone app, 63.6% of the 3D Vectra[\u00ae] WB360 imaging device, and 40.0% of the 2D FotoFinder ATBM[\u00ae] risk assessment tool. Dermatologists correctly identified five of six melanomas (83% sensitivity). The false-negative rate was 17%. The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses among all dermatologists was 92.7%. The performance of the dermatologists strongly correlated with their professional experience. Hence, the true-positive rates for melanoma were greater for experts and dermatologists with intermediate experience (100% sensitivity) compared to beginners (80% sensitivity); whereas the true-negative rates were similar (93.3% resp. 92.5% specificity) (Table 3).\n\n**Figure 4. Histology and corresponding diagnosis of lesions assessed by dermatologists, smartphone**\napp SkinVision[\u00ae], and dermatologists and AI (n = 61).\n\nFigure 5 reveals several pigmented skin lesions that were correctly and incorrectly classified as melanomas by the smartphone app SkinVision[\u00ae].\n\n**Figure 5. Correctly and falsely classified melanomas by the smartphone app SkinVision[\u00ae]: (A,B). True-**\npositive classified melanoma; (C,D). False-positive classified melanoma; (E,F). True-negative classified\nmelanoma; (G). False-negative classified melanoma.\n\nROC analysis of the classification of melanoma of the SkinVision[\u00ae] app compared\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision[\u00ae] app had\nhigh sensitivity but mediocre specificity.\n\nROC analysis of the classification of melanoma of the SkinVision\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision\nhigh sensitivity but mediocre specificity.\n\n-----\n\n_Cancers 2022, 14, 3829_ 11 of 21\n\n**Figure 6. Receiver operating characteristic curve of the smartphone app in relation to the results of**\nthe histology (sensitivity of 83.3%, specificity 60.0%); AUC = area under the curve.\n\n_3.3. Patient Perspective on AI in Melanoma Screening_\n3.3.1. Confidence in Dermatologists vs. Smartphone App\n\nMost of the patients at high-risk for melanoma (55% (30/55)) and patients with\nmelanoma (53% (31/59)) reported being very confident about a mole examination by\na dermatologist (rating scale 10/10). In contrast, only a minority of high-risk patients (16%\n(9/55)) and melanoma patients (12% (7/59)) felt very safe when being investigated by a\nsmartphone app alone. No significant difference in ratings was identified between the two\nrisk groups (p < 0.9; p < 0.7) (Supplementary Table S2).\n\n3.3.2. Trustworthiness of the Smartphone App\n\nAll participants primarily rated physician examination (100% (55/55) among highrisk patients, 100% (59/59) among melanoma patients) as trustworthy, and the majority\ndid so for 2D TBP imaging (93% (51/55) resp. 88% (52/59)) and 3D TBP imaging (91%\n(50/55) resp. 90% (53/59)). The smartphone app was less frequently rated as trustworthy,\nwith only 36% (20/55) among patients at high-risk for melanoma and 49% (29/59) among\nmelanoma patients (Table 4). The age revealed a significant correlation with the evaluation\nof smartphones\u2019 trustworthiness (p < 0.004), with older patients (>60 years old) having\ntrusted the app three times more than younger patients (\u226460 years old) (Figure 7). Neither\nprevious melanoma vs. high-risk criteria for melanoma nor sex significantly influenced the\nevaluation of the trustworthiness of the smartphone app.\n\n-----\n\n_Cancers 2022, 14, 3829_ 12 of 21\n\n**Table 4. Assessment of trustworthiness of the AI-based smartphone app SkinVision[\u00ae], 2D imaging**\nFotoFinder ATBM[\u00ae], and 3D imaging Vectra[\u00ae] WB360 compared to dermatologists.\n\n\n**Patients at**\n**High-Risk**\n**for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Characteristic** **N**\n\n\n**Patients**\n**with**\n**Melanoma,**\n**N = 59** **[1]**\n\n\n**_p-Value_** **[2]**\n\n\n**The following examination was**\n**trustworthy: Smartphone** 114 0.3\n**app assessment**\n\nYes 29 (49%) 20 (36%)\nNo 5 (8.5%) 8 (15%)\nI don\u2019t know 23 (39%) 22 (40%)\nNo answer 2 (3.4%) 5 (9.1%)\n**Dermatologist assessment** 114\nYes 59 (100%) 55 (100%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 0 (0%) 0 (0%)\nNo answer 0 (0%) 0 (0%)\n**2D TBP assessment** 114 0.3\nYes 52 (88%) 51 (93%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 7 (12%) 3 (5.5%)\nNo answer 0 (0%) 1 (1.8%)\n**3D TBP assessment** 114 0.3\nYes 53 (90%) 50 (91%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 6 (10%) 3 (5.5%)\nNo answer 0 (0%) 2 (3.6%)\n\n1 n (%); 2 Fisher\u2019s exact test; Pearson\u2019s Chi-squared test; TBP = total body photography.\n\n**Figure 7. Odds ratio for variables influencing the trustworthiness of smartphones\u2019 risk assessment in**\nmelanoma detection.\n\n-----\n\n_Cancers 2022, 14, 3829_ 13 of 21\n\n3.3.3. Impact of AI vs. Dermatologists\u2019 Examination on Patients\u2019 Fear of Developing\nSkin Cancer\n\nMost participants indicated that an examination by a physician reduced their fear of\ndeveloping skin cancer, namely in 89% (49/55) among high-risk patients and 81% (48/59)\namong melanoma patients. Comparably, the 2D TBP imaging achieved the same effect in\n78% (43/55) resp. 76% (45/59), and the 3D TBP device in 82% (45/55) resp. 75% (44/59). In\ncontrast, the assessment with the smartphone app appeased fear of skin cancer in only 33%\n(18/55) of patients at high-risk for melanoma and in 32% (19/59) of patients with melanoma.\nIndeed, one high-risk patient (1.8%) even reported an increased fear of developing skin\ncancer by using the smartphone app (Supplementary Table S2).\n\n3.3.4. Patients\u2019 Subjective Assessment of the Accuracy of AI vs. Dermatologists\n\nPatients expected reliable results with the highest accuracy by both the assessment\nby a physician (98% (54/55) among high-risk patients, 92% (54/59) among melanoma\npatients) and by the 2D TBP imaging (82% (45/55) resp. 86% (51/59)) as well as the 3D\nTBP device (89% (49/55) resp. 88% (52/59)). Only 16% (9/55) of high-risk patients and\n31% (18/59) of melanoma patients expected reliable results from the smartphone app\n(Supplementary Table S2). No significant differences were identified between the two risk\ngroups in the evaluations.\n\n3.3.5. Patient Preference for Skin Cancer Screening\n\nBoth cohorts favored a combination of dermatologist and 3D TBP risk assessment for\nthe examination of pigmented skin lesions (64% (35/55) among patients at high-risk for\nmelanoma, 51% (30/59) among melanoma patients), while neither preferred assessment\nby a smartphone app alone (Figure 8A, Supplementary Table S2). The combination of\ndermatologist and smartphone app was favored by only 1.8% (1/55) of patients at high-risk\nfor melanoma and 3.4% (2/59) of patients with melanoma.\n\n**Figure 8. Cont.**\n\n-----\n\n_Cancers 2022, 14, 3829_ 14 of 21\n\n**Figure 8. (A) Patient preference for mole assessment (patients at high-risk for melanoma, n = 55;**\npatients with melanoma, n = 59); (B) Patient preference for AI in skin cancer screening (patients at\nhigh-risk for melanoma, n = 55; patients with melanoma, n = 59).\n\nRegarding patient preference for skin cancer screenings, almost all high-risk patients\n(98% (54/55)) and melanoma patients (95% (56/59)) indicated their belief that AI can improve\na physician\u2019s diagnostic performance. Most patients (64% (35/55) among high-risk patients,\n54% (32/59) among melanoma patients) would prefer that the physicians would always\nconsider the result of AI in their diagnosis (Figure 8B) (Supplementary Table S2).\n\n3.3.6. Dermatologists\u2019 Perspective of Smartphone Apps for Melanoma Screening\n\nAmong the seven dermatologists, they stated in only 5.3% of the skin cancer screenings\n(6/114) that the smartphone app increased diagnostic confidence and in only 8.8% of the\nassessments (10/114) they trusted the app (Supplementary Table S3).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Over-Detection of Melanoma-Suspect Lesions by a CE-Certified Smartphone App: Performance in Comparison to Dermatologists, 2D and 3D Convolutional Neural Networks in a Prospective Data Set of 1204 Pigmented Skin Lesions Involving Patients' Perception.",
        "abstract": "The exponential increase in algorithm-based mobile health (mHealth) applications (apps) for melanoma screening is a reaction to a growing market. However, the performance of available apps remains to be investigated. In this prospective study, we investigated the diagnostic accuracy of a class 1 CE-certified smartphone app in melanoma risk stratification and its patient and dermatologist satisfaction. Pigmented skin lesions \u2265 3 mm and any suspicious smaller lesions were assessed by the smartphone app SkinVision\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Study Design and Participants_\n\nWe performed this prospective, single-center, comparative observational cohort study at the Department of Dermatology at the University Hospital Basel in Switzerland between January 2021 to June 2021. Seven dermatologists (beginners: <2 years professional experience, n = 4; intermediate: 2\u20135 years professional experience, n = 1; experts: >5 years professional experience, n = 2), as well as 114 patients were included in the study with an age of \u226518 years with a high risk for developing a primary cutaneous malignant melanoma defined as previous melanoma (including melanoma in situ), \u2265100 melanocytic nevi, \u22655 atypical nevi, strong family history for melanoma, diagnosis of dysplastic nevus syndrome, or known CDKN2A mutation. Exclusion criteria were acute psychiatric illness or acute crisis and Fitzpatrick sign type V-VI.\n\n_2.2. Procedures_\n\nIn the first step, the patients signed the informed consent after they were informed by a dermatologist about the study design and the following examinations. Subsequently, the dermatologist obtained a medical history, including a history of melanoma to enable classification into the different risk groups. The classification of all high-risk patients into further risk cohorts are defined by how often the patients require skin cancer screening. Based on this, we have classified our melanoma patients as high risk, as they come to dermatological control every 3\u201312 months according to the corresponding AJCC stage. Patients without melanoma in their own medical history are seen every 12 months. All study participants underwent a standard skin cancer screening by a dermatologist (Supplementary Figure S1). The pigmented skin lesions were assessed with dermatoscopy and classified as either benign or malignant. The dermatologist defined pigmented skin lesions that were suspected melanoma and indicated whether lesions smaller than 3 mm should be classified with the different AI modalities due to suspicion. The regular skin cancer screening was performed before any AI-based risk assessments were achieved, so the dermatologists had no knowledge about AI classification of the melanocytic lesions up to this point. In the next step, all melanocytic nevi \u2265 3 mm or any smaller suspicious melanocytic lesions were assessed using an iOS-based iPhone SE smartphone equipped with a 12-megapixel camera and the app SkinVision[\u00ae] (SkinVision[\u00ae] B.V., Amsterdam, the Netherlands, App-Version 6.8.1), which is based on a machine-learning algorithm classifying pigmented lesions into low, medium, and high risk for melanoma. Only the lesions classified as high-risk\u2014indicating a recommendation to consult a dermatologist\u2014were recorded. Afterwards, patients received a whole-body screening using the 2D automated total body mapping (ATBM[\u00ae]) master of FotoFinder (FotoFinder ATBM[\u00ae] Systems GmbH, Bad Birnbach, Germany, Version 3.3.1.0, risk scoring with MoleAnalyzer Pro), and the 3D TBP Vectra[\u00ae] WB360 System (Canfield Scientific, Parsippany, NJ, USA, Version 4.7.1, risk scoring with DEXI) which was performed by medical students and a study nurse. Subsequently, all lesions over 3 mm or any smaller suspicious lesions were assessed by the 2D and 3D AI devices using their dermoscopes (dermoscope medicam 1000 (FotoFinder ATBM[\u00ae]) and VISIOMED[\u00ae] D200evo dermoscope (Vectra[\u00ae] WB360)). For dermoscopic imaged melanocytic nevi, we obtained AI-based risk scores between 0.0\u20131.0 for 2D FotoFinder ATBM[\u00ae] and 0.0\u201310.0 for 3D TBP Vectra[\u00ae] WB360. FotoFinder\u2019s ATBM[\u00ae] master creates a 2D-image of the entire surface of the patient using 20 photos taken from 8 different parts of the body, while Vectra[\u00ae] WB360 generates a 3D-image using 92 individual photos that are converted into a 3D Avatar. Finally, the dermatologists assessed the skin lesions a second time with knowledge of the AI risk assessment scores of 2D FotoFinder ATBM[\u00ae] and 3D TBP Vectra[\u00ae] WB360. The patient was informed about the lesions with suspicion for melanoma by a dermatologist and about the 2D and 3D CNN classifications above the study-defined cut-off scores. The indication for excision of suspicious lesions was based on the dermatologists\u2019 suspicion for melanoma and/or malignancy risk assessment scores of the 2D imaging tool FotoFinder ATBM[\u00ae] and/or 3D imaging tool Vectra[\u00ae] WB360 indicating values over the study-defined cut-off score (>0.5 resp. >5.0). The patient was informed which lesions would be removed with the appropriate rationale. A maximum of two excisions were performed per visit and per patient due to ethical reasons, with an exception for further biopsies in cases of high-grade suspicion of malignancy by the dermatologist or patient request. The reference standard for a benign lesion was either the histological diagnosis (no melanoma) or the combination of the dermatologist\u2019s evaluation (benign lesion) plus the AI scores of two independent medical devices (FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360) each below the cut-off score. After all the assessments, the participating patients completed an anonymous 9-item questionnaire that surveyed their attitudes, personal preferences, and concerns about the use of AI in melanoma detection compared with skin cancer screenings by a dermatologist in addition to sociodemographic data. The dermatologists completed a 2-item questionnaire assessing their attitudes toward the smartphone app in melanoma screening, which was answered specifically for each of the 114 included patients. Neither the patients nor dermatologists were informed about the result of the app\u2019s risk assessment due to the extremely high number of false-positive findings to prevent response bias. Our survey was designed de novo after literature research. All answers were optional. The questionnaire was available in German, English and French and included binary questions (yes/no), multiple choice questions and visual analogue scales (VAS) with scores from 0\u201310.\n\n_2.3. Statistical Analysis_\n\nComparison of continuous variables was tested with the Wilcoxon rank test. Comparison of categorical variables was tested with Fishers Exact test. A p-value of less than 0.05 was deemed as significant. Receiver Operating Characteristics (ROC) analysis was used to assess the performance of the SkinVision[\u00ae] app against the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologist and the SkinVision[\u00ae] app against histology. All analyses were performed with R (version 4.1) and visualized with ggplot2.\n\n_2.4. Ethics_\n\nThe study was approved by the local ethics committee (22020-02482), registered with ClinicalTrials.gov (NCT04605822), and was conducted in compliance with the Declaration of Helsinki and Good Clinical Practice GCP-rules.",
        "results": "**3. Results**\n\n_3.1. Study Population_\n\nOverall, data from 114 patients were analyzed, including 55 patients at high-risk for developing a melanoma (mean age 55 years (age range): 22\u201385), 47% females) and 59 melanoma patients (mean age 60 years (age range: 29\u201381), 54% females) (Table 1). The family history for melanoma was positive in 19% of the melanoma patients and in 56% of the non-melanoma patients. Most of the patients used sunscreen SPF 30\u201350 (51% of melanoma patients vs. 62% of patients at high-risk for melanoma) and had previous sunburns in childhood (54% resp. 69%).\n\n**Table 1. Characteristics of the study population and their skin cancer awareness.**\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Age, n (age range)** 59 (22\u201385) 60 (29\u201381) 55 (22\u201385)\n**Sex, n (%)**\nFemale 58 (51%) 32 (54%) 26 (47%)\nMale 56 (49%) 27 (46%) 29 (53%)\n**Risk profile, n (%)**\nMultiple melanocytic nevi (\u2265100)\nand/or dysplastic nevi (\u22655) and/or positive\nfamily history for melanoma and/or 55 (48%) 0 (0%) 55 (100%)\ndiagnosis of dysplastic nevus syndrome\nand/or CDKN2A mutation\n\nPrevious resected melanoma in situ or\n57 (50%) 57 (97%) 0 (0%)\nprimary cutaneous melanoma\n\nMetastatic melanoma 2 (1.8%) 2 (3.4%) 0 (0%)\n**Positive family history for melanoma,**\n42 (37%) 11 (19%) 31 (56%)\n**n (%)**\n\n**Frequency of skin cancer screening,**\n**n (%)**\nSeveral times per year 40 (35%) 34 (58%) 6 (11%)\nEvery 12 months 39 (34%) 16 (27%) 23 (42%)\nEvery 1\u20132 years 8 (7%) 4 (6.8%) 4 (7.3%)\nEvery 2 years 9 (7.9%) 2 (3.4%) 7 (13%)\nLess than every 2 years 14 (12%) 3 (5.1%) 11 (20%)\nNever 4 (3.5%) 0 (0%) 4 (7.3%)\n**History of sunburns in childhood,**\n70 (61%) 32 (54%) 38 (69%)\n**n (%)**\n\n**Frequency of sunburns (Child), n (%)**\nRarely (less than once per year) 44 (63%) 20 (62%) 24 (63%)\nRegularly (once per year) 22 (31%) 10 (31%) 12 (32%)\nOften (more than once per year) 4 (5.7%) 2 (6.2%) 2 (5.3%)\n**History of sunburns in adulthood,**\n39 (34%) 18 (31%) 21 (38%)\n**n (%)**\n\n\n-----\n\n_Cancers 2022, 14, 3829_ 6 of 21\n\n**Table 1. Cont.**\n\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Frequency of sunburns (Adult), n (%)**\nRarely (less than once per year) 38 (97%) 18 (100%) 20 (95%)\nRegularly (once per year) 0 (0%) 0 (0%) 0 (0%)\nOften (more than once per year) 1 (2.6%) 0 (0%) 1 (4.8%)\n**Previous tanning in the solarium,**\n38 (33%) 13 (22%) 25 (45%)\n**n (%)**\n\n**Usage of sunscreen (SPF), n (%)**\nSPF 6\u201310 2 (1.8%) 1 (1.7%) 1 (1.8%)\nSPF 15\u201325 10 (8.8%) 3 (5.1%) 7 (13%)\nSPF 30\u201350 64 (56%) 30 (51%) 34 (62%)\nSPF 50+ 38 (33%) 25 (42%) 13 (24%)\n\n1 Median (Range); n (%).\n\nA total of 1204 pigmented skin lesions were assessed in this study (Figure 1). In 61 cases (5.1%), we performed a histopathology examination, while 1129 lesions (94%) were diagnosed to be clinically clearly benign based on the combination of a risk assessment by a dermatologist and AI-based risk scores below the cut-off (2D and 3D TBP) and had no indication for obtaining histopathology.\n\n**Figure 1. Flow chart of all included pigmented skin lesions and their histopathological outcome.**\nAI = artificial intelligence; TBP = total body photography.\n\n_3.2. Diagnostic Accuracy and Performance of the Smartphone App SkinVision[\u00ae]_\n\n3.2.1. Comparison of all Risk Assessments\n\nThe smartphone app SkinVision[\u00ae] classified 980 (81%) lesions as benign and indicated an increased risk for melanoma in 224 (19%) cases, while the dermatologists diagnosed 1195 (99.3%) lesions as benign and only nine (0.7%) as suspicious. Consequently, the CE-certified app had a 27-fold higher rate of melanoma-suspicious lesions compared to dermatologists. The AI scores of the 2D and 3D CNN devices classified most lesions as benign and 47 (3.9%) lesions (FotoFinder ATBM[\u00ae]) resp. 39 (3.2%) lesions (Vectra[\u00ae] WB360) as suspicious for melanoma (Figure 2).\n\n-----\n\n_Cancers 2022, 14, 3829_ 7 of 21\n\n**Figure 2. Comparison of risk assessments with the highest rate of suspected melanoma cases by the**\nsmartphone app (n = 1204).\n\nAmong the 224 lesions classified as suspicious by SkinVision[\u00ae], 193 were considered clinically benign by the physician and the 2D and 3D CNN AI-risk assessment scores, whereas four pigmented skin lesions were classified suspicious by the smartphone app, dermatologists, 2D, and 3D TBP devices (Table 2, Supplementary Table S1). The knowledge of AI-based risk assessment scores did not meaningfully affect dermatologists\u2019 classification of skin lesions, as they changed their decision towards the indication for excision in only three lesions that later turned out to be benign (Table 2).\n\n**Table 2. Risk assessments of 1204 pigmented skin lesions by the smartphone app SkinVision[\u00ae], 2D**\nimaging FotoFinder ATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in\ncombination with knowledge of FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360 AI-scores.\n\n**Characteristic** **N = 1204** **[1]**\n\n**Smartphone app SkinVision[\u00ae]**\n\nbenign 980 (81%)\nsuspicious 224 (19%)\n**2D Imaging FotoFinder ATBM[\u00ae]**\n\nbenign 1157 (96%)\nsuspicious 47 (3.9%)\n**3D Imaging VECTRA[\u00ae]** **WB360**\nbenign 1165 (97%)\nsuspicious 39 (3.2%)\n**Dermatologists**\nbenign 1195 (99%)\nsuspicious 9 (0.7%)\n**Dermatologists informed about risk assessment scores by**\n**FotoFinder ATBM[\u00ae]** **+ VECTRA[\u00ae]** **WB360**\nbenign 1192 (99%)\nsuspicious 12 (1.0%)\n\n1 n (%); AI = artificial intelligence.\n\n-----\n\n_Cancers 2022, 14, 3829_ 8 of 21\n\n3.2.2. Diagnostic Accuracy of the Smartphone App Based on the Combination of the\nDermatologist\u2019s Evaluation plus the AI Risk-Assessment Scores of Two Independent\nMedical Devices\n\nReceiver operating characteristic (ROC) analysis of the classification of benign and suspicious lesions of the SkinVision[\u00ae] app compared to the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologists had an area under the curve (AUC) score of 0.621, sensitivity of 0.41, and specificity of 0.83 (Figure 3). Although the specificity is reasonable, the sensitivity is low, thus suggesting that the SkinVision[\u00ae] app has poor diagnostic accuracy.\n\n**Figure 3. Receiver operating characteristic curve of the smartphone app in relation to the results**\nof the combination of risk assessments by dermatologists, FotoFinder ATBM[\u00ae], and Vectra[\u00ae] WB360\n(sensitivity: 41.3%, specificity: 82.9%); AUC = area under the curve.\n\n3.2.3. Diagnostic Accuracy of the Smartphone App Based on Histopathology\n\nAmong 61 pigmented skin lesions examined histologically, we detected six melanomas, 19 melanocytic nevi, 20 dysplastic nevi, as well as 16 otherwise classified lesions were diagnosed (Table 3, Figure 1).\n\n**Table 3. Diagnostic accuracy of the AI-based smartphone app SkinVision[\u00ae], 2D imaging FotoFinder**\nATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in combination with AI in\nmelanoma detection based on histopathology: sensitivity and specificity.\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**Smartphone app SkinVision[\u00ae]** 61\nbenign 13 (68%) 10 (50%) 1 (17%) 10 (62%)\nsuspicious 6 (32%) 10 (50%) 5 (83%) 6 (38%)\n**2D imaging FotoFinder ATBM[\u00ae]** 61\nbenign 7 (37%) 11 (55%) 1 (17%) 4 (25%)\nsuspicious 12 (63%) 9 (45%) 5 (83%) 12 (75%)\n\n-----\n\n_Cancers 2022, 14, 3829_ 9 of 21\n\n**Table 3. Cont.**\n\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**3D imaging VECTRA[\u00ae]** **WB360** 61\nbenign 18 (95%) 9 (45%) 1 (17%) 8 (50%)\nsuspicious 1 (5.3%) 11 (55%) 5 (83%) 8 (50%)\n**Dermatologists** 61\nbenign 17 (89%) 18 (90%) 1 (17%) 16 (100%)\nsuspicious 2 (11%) 2 (10%) 5 (83%) 0 (0%)\n_Beginner: <2 years\u2019 work_\n44 N = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 14 (93%) 10 (83%) 1 (20%) 13 (100%)\nsuspicious 1 (6.7%) 2 (17%) 4 (80%) 0 (0%)\n_Intermediate: 2\u20135 years\u2019 work_\n5 N = 2 N = 3 N = 0 N = 0\n_experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\n11 N = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n**Dermatologists informed about**\n61\n**AI scores** **[2]**\n\nbenign 16 (84%) 17 (85%) 1 (17%) 15 (94%)\nsuspicious 3 (16%) 3 (15%) 5 (83%) 1 (6.2%)\n_Beginner: <2 years\u2019 work_\nN = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 13 (87%) 9 (75%) 1 (20%) 12 (92%)\nsuspicious 2 (13%) 3 (25%) 4 (80%) 1 (7.7%)\n_Intermediate: 2\u20135 years\u2019_\nN = 2 N = 3 N = 0 N = 0\n_work experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\nN = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n\n1 n (%); * Other = pigmented basal cell carcinoma, histiocytoma, lentigo solaris, pigmented seborrhoic keratosis,\nfolliculitis with perifolliculitis, collisional tumor: seborrhoic keratosis and nevus, collisional tumor: actinic\nkeratosis and lentigo solaris; [2] risk assessment scores by FotoFinder ATBM[\u00ae] and VECTRA[\u00ae] WB360; AI = artificial\nintelligence.\n\nBased on histopathology, Figure 4 represents the risk assessments of the dermatologists, the mHealth app, and the combination of AI and dermatologists. Both the SkinVision[\u00ae]\n\napp and the 2D and 3D CNN devices indicated for 5 of 6 histological verified melanomas\n(83% sensitivity) an elevated score indicating suspicion of melanoma. The false-negative rate for all AI-based medical devices was 17%. The three different false-negative classified melanomas were all superficial spreading stage IA melanomas\u2014SkinVision[\u00ae]: superficial spreading melanoma, 0.9 mm Breslow thickness, AJCC stage IA (pT1bN0M0); 3D TBP Vectra[\u00ae] WB360: superficial spreading melanoma, 0.3 mm Breslow thickness, AJCC stage IA (pT1aN0M0); 2D TBP FotoFinder ATBM[\u00ae]: superficial spreading melanoma, 0.7 mm Breslow thickness, AJCC stage IA (pT1aN0M0). The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses was 60.0% of the smartphone app, 63.6% of the 3D Vectra[\u00ae] WB360 imaging device, and 40.0% of the 2D FotoFinder ATBM[\u00ae] risk assessment tool. Dermatologists correctly identified five of six melanomas (83% sensitivity). The false-negative rate was 17%. The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses among all dermatologists was 92.7%. The performance of the dermatologists strongly correlated with their professional experience. Hence, the true-positive rates for melanoma were greater for experts and dermatologists with intermediate experience (100% sensitivity) compared to beginners (80% sensitivity); whereas the true-negative rates were similar (93.3% resp. 92.5% specificity) (Table 3).\n\n**Figure 4. Histology and corresponding diagnosis of lesions assessed by dermatologists, smartphone**\napp SkinVision[\u00ae], and dermatologists and AI (n = 61).\n\nFigure 5 reveals several pigmented skin lesions that were correctly and incorrectly classified as melanomas by the smartphone app SkinVision[\u00ae].\n\n**Figure 5. Correctly and falsely classified melanomas by the smartphone app SkinVision[\u00ae]: (A,B). True-**\npositive classified melanoma; (C,D). False-positive classified melanoma; (E,F). True-negative classified\nmelanoma; (G). False-negative classified melanoma.\n\nROC analysis of the classification of melanoma of the SkinVision[\u00ae] app compared\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision[\u00ae] app had\nhigh sensitivity but mediocre specificity.\n\nROC analysis of the classification of melanoma of the SkinVision\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision\nhigh sensitivity but mediocre specificity.\n\n-----\n\n_Cancers 2022, 14, 3829_ 11 of 21\n\n**Figure 6. Receiver operating characteristic curve of the smartphone app in relation to the results of**\nthe histology (sensitivity of 83.3%, specificity 60.0%); AUC = area under the curve.\n\n_3.3. Patient Perspective on AI in Melanoma Screening_\n3.3.1. Confidence in Dermatologists vs. Smartphone App\n\nMost of the patients at high-risk for melanoma (55% (30/55)) and patients with\nmelanoma (53% (31/59)) reported being very confident about a mole examination by\na dermatologist (rating scale 10/10). In contrast, only a minority of high-risk patients (16%\n(9/55)) and melanoma patients (12% (7/59)) felt very safe when being investigated by a\nsmartphone app alone. No significant difference in ratings was identified between the two\nrisk groups (p < 0.9; p < 0.7) (Supplementary Table S2).\n\n3.3.2. Trustworthiness of the Smartphone App\n\nAll participants primarily rated physician examination (100% (55/55) among highrisk patients, 100% (59/59) among melanoma patients) as trustworthy, and the majority\ndid so for 2D TBP imaging (93% (51/55) resp. 88% (52/59)) and 3D TBP imaging (91%\n(50/55) resp. 90% (53/59)). The smartphone app was less frequently rated as trustworthy,\nwith only 36% (20/55) among patients at high-risk for melanoma and 49% (29/59) among\nmelanoma patients (Table 4). The age revealed a significant correlation with the evaluation\nof smartphones\u2019 trustworthiness (p < 0.004), with older patients (>60 years old) having\ntrusted the app three times more than younger patients (\u226460 years old) (Figure 7). Neither\nprevious melanoma vs. high-risk criteria for melanoma nor sex significantly influenced the\nevaluation of the trustworthiness of the smartphone app.\n\n-----\n\n_Cancers 2022, 14, 3829_ 12 of 21\n\n**Table 4. Assessment of trustworthiness of the AI-based smartphone app SkinVision[\u00ae], 2D imaging**\nFotoFinder ATBM[\u00ae], and 3D imaging Vectra[\u00ae] WB360 compared to dermatologists.\n\n\n**Patients at**\n**High-Risk**\n**for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Characteristic** **N**\n\n\n**Patients**\n**with**\n**Melanoma,**\n**N = 59** **[1]**\n\n\n**_p-Value_** **[2]**\n\n\n**The following examination was**\n**trustworthy: Smartphone** 114 0.3\n**app assessment**\n\nYes 29 (49%) 20 (36%)\nNo 5 (8.5%) 8 (15%)\nI don\u2019t know 23 (39%) 22 (40%)\nNo answer 2 (3.4%) 5 (9.1%)\n**Dermatologist assessment** 114\nYes 59 (100%) 55 (100%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 0 (0%) 0 (0%)\nNo answer 0 (0%) 0 (0%)\n**2D TBP assessment** 114 0.3\nYes 52 (88%) 51 (93%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 7 (12%) 3 (5.5%)\nNo answer 0 (0%) 1 (1.8%)\n**3D TBP assessment** 114 0.3\nYes 53 (90%) 50 (91%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 6 (10%) 3 (5.5%)\nNo answer 0 (0%) 2 (3.6%)\n\n1 n (%); 2 Fisher\u2019s exact test; Pearson\u2019s Chi-squared test; TBP = total body photography.\n\n**Figure 7. Odds ratio for variables influencing the trustworthiness of smartphones\u2019 risk assessment in**\nmelanoma detection.\n\n-----\n\n_Cancers 2022, 14, 3829_ 13 of 21\n\n3.3.3. Impact of AI vs. Dermatologists\u2019 Examination on Patients\u2019 Fear of Developing\nSkin Cancer\n\nMost participants indicated that an examination by a physician reduced their fear of\ndeveloping skin cancer, namely in 89% (49/55) among high-risk patients and 81% (48/59)\namong melanoma patients. Comparably, the 2D TBP imaging achieved the same effect in\n78% (43/55) resp. 76% (45/59), and the 3D TBP device in 82% (45/55) resp. 75% (44/59). In\ncontrast, the assessment with the smartphone app appeased fear of skin cancer in only 33%\n(18/55) of patients at high-risk for melanoma and in 32% (19/59) of patients with melanoma.\nIndeed, one high-risk patient (1.8%) even reported an increased fear of developing skin\ncancer by using the smartphone app (Supplementary Table S2).\n\n3.3.4. Patients\u2019 Subjective Assessment of the Accuracy of AI vs. Dermatologists\n\nPatients expected reliable results with the highest accuracy by both the assessment\nby a physician (98% (54/55) among high-risk patients, 92% (54/59) among melanoma\npatients) and by the 2D TBP imaging (82% (45/55) resp. 86% (51/59)) as well as the 3D\nTBP device (89% (49/55) resp. 88% (52/59)). Only 16% (9/55) of high-risk patients and\n31% (18/59) of melanoma patients expected reliable results from the smartphone app\n(Supplementary Table S2). No significant differences were identified between the two risk\ngroups in the evaluations.\n\n3.3.5. Patient Preference for Skin Cancer Screening\n\nBoth cohorts favored a combination of dermatologist and 3D TBP risk assessment for\nthe examination of pigmented skin lesions (64% (35/55) among patients at high-risk for\nmelanoma, 51% (30/59) among melanoma patients), while neither preferred assessment\nby a smartphone app alone (Figure 8A, Supplementary Table S2). The combination of\ndermatologist and smartphone app was favored by only 1.8% (1/55) of patients at high-risk\nfor melanoma and 3.4% (2/59) of patients with melanoma.\n\n**Figure 8. Cont.**\n\n-----\n\n_Cancers 2022, 14, 3829_ 14 of 21\n\n**Figure 8. (A) Patient preference for mole assessment (patients at high-risk for melanoma, n = 55;**\npatients with melanoma, n = 59); (B) Patient preference for AI in skin cancer screening (patients at\nhigh-risk for melanoma, n = 55; patients with melanoma, n = 59).\n\nRegarding patient preference for skin cancer screenings, almost all high-risk patients\n(98% (54/55)) and melanoma patients (95% (56/59)) indicated their belief that AI can improve\na physician\u2019s diagnostic performance. Most patients (64% (35/55) among high-risk patients,\n54% (32/59) among melanoma patients) would prefer that the physicians would always\nconsider the result of AI in their diagnosis (Figure 8B) (Supplementary Table S2).\n\n3.3.6. Dermatologists\u2019 Perspective of Smartphone Apps for Melanoma Screening\n\nAmong the seven dermatologists, they stated in only 5.3% of the skin cancer screenings\n(6/114) that the smartphone app increased diagnostic confidence and in only 8.8% of the\nassessments (10/114) they trusted the app (Supplementary Table S3).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "High"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Over-Detection of Melanoma-Suspect Lesions by a CE-Certified Smartphone App: Performance in Comparison to Dermatologists, 2D and 3D Convolutional Neural Networks in a Prospective Data Set of 1204 Pigmented Skin Lesions Involving Patients' Perception.",
        "abstract": "The exponential increase in algorithm-based mobile health (mHealth) applications (apps) for melanoma screening is a reaction to a growing market. However, the performance of available apps remains to be investigated. In this prospective study, we investigated the diagnostic accuracy of a class 1 CE-certified smartphone app in melanoma risk stratification and its patient and dermatologist satisfaction. Pigmented skin lesions \u2265 3 mm and any suspicious smaller lesions were assessed by the smartphone app SkinVision\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Study Design and Participants_\n\nWe performed this prospective, single-center, comparative observational cohort study at the Department of Dermatology at the University Hospital Basel in Switzerland between January 2021 to June 2021. Seven dermatologists (beginners: <2 years professional experience, n = 4; intermediate: 2\u20135 years professional experience, n = 1; experts: >5 years professional experience, n = 2), as well as 114 patients were included in the study with an age of \u226518 years with a high risk for developing a primary cutaneous malignant melanoma defined as previous melanoma (including melanoma in situ), \u2265100 melanocytic nevi, \u22655 atypical nevi, strong family history for melanoma, diagnosis of dysplastic nevus syndrome, or known CDKN2A mutation. Exclusion criteria were acute psychiatric illness or acute crisis and Fitzpatrick sign type V-VI.\n\n_2.2. Procedures_\n\nIn the first step, the patients signed the informed consent after they were informed by a dermatologist about the study design and the following examinations. Subsequently, the dermatologist obtained a medical history, including a history of melanoma to enable classification into the different risk groups. The classification of all high-risk patients into further risk cohorts are defined by how often the patients require skin cancer screening. Based on this, we have classified our melanoma patients as high risk, as they come to dermatological control every 3\u201312 months according to the corresponding AJCC stage. Patients without melanoma in their own medical history are seen every 12 months. All study participants underwent a standard skin cancer screening by a dermatologist (Supplementary Figure S1). The pigmented skin lesions were assessed with dermatoscopy and classified as either benign or malignant. The dermatologist defined pigmented skin lesions that were suspected melanoma and indicated whether lesions smaller than 3 mm should be classified with the different AI modalities due to suspicion. The regular skin cancer screening was performed before any AI-based risk assessments were achieved, so the dermatologists had no knowledge about AI classification of the melanocytic lesions up to this point. In the next step, all melanocytic nevi \u2265 3 mm or any smaller suspicious melanocytic lesions were assessed using an iOS-based iPhone SE smartphone equipped with a 12-megapixel camera and the app SkinVision[\u00ae] (SkinVision[\u00ae] B.V., Amsterdam, the Netherlands, App-Version 6.8.1), which is based on a machine-learning algorithm classifying pigmented lesions into low, medium, and high risk for melanoma. Only the lesions classified as high-risk\u2014indicating a recommendation to consult a dermatologist\u2014were recorded. Afterwards, patients received a whole-body screening using the 2D automated total body mapping (ATBM[\u00ae]) master of FotoFinder (FotoFinder ATBM[\u00ae] Systems GmbH, Bad Birnbach, Germany, Version 3.3.1.0, risk scoring with MoleAnalyzer Pro), and the 3D TBP Vectra[\u00ae] WB360 System (Canfield Scientific, Parsippany, NJ, USA, Version 4.7.1, risk scoring with DEXI) which was performed by medical students and a study nurse. Subsequently, all lesions over 3 mm or any smaller suspicious lesions were assessed by the 2D and 3D AI devices using their dermoscopes (dermoscope medicam 1000 (FotoFinder ATBM[\u00ae]) and VISIOMED[\u00ae] D200evo dermoscope (Vectra[\u00ae] WB360)). For dermoscopic imaged melanocytic nevi, we obtained AI-based risk scores between 0.0\u20131.0 for 2D FotoFinder ATBM[\u00ae] and 0.0\u201310.0 for 3D TBP Vectra[\u00ae] WB360. FotoFinder\u2019s ATBM[\u00ae] master creates a 2D-image of the entire surface of the patient using 20 photos taken from 8 different parts of the body, while Vectra[\u00ae] WB360 generates a 3D-image using 92 individual photos that are converted into a 3D Avatar. Finally, the dermatologists assessed the skin lesions a second time with knowledge of the AI risk assessment scores of 2D FotoFinder ATBM[\u00ae] and 3D TBP Vectra[\u00ae] WB360. The patient was informed about the lesions with suspicion for melanoma by a dermatologist and about the 2D and 3D CNN classifications above the study-defined cut-off scores. The indication for excision of suspicious lesions was based on the dermatologists\u2019 suspicion for melanoma and/or malignancy risk assessment scores of the 2D imaging tool FotoFinder ATBM[\u00ae] and/or 3D imaging tool Vectra[\u00ae] WB360 indicating values over the study-defined cut-off score (>0.5 resp. >5.0). The patient was informed which lesions would be removed with the appropriate rationale. A maximum of two excisions were performed per visit and per patient due to ethical reasons, with an exception for further biopsies in cases of high-grade suspicion of malignancy by the dermatologist or patient request. The reference standard for a benign lesion was either the histological diagnosis (no melanoma) or the combination of the dermatologist\u2019s evaluation (benign lesion) plus the AI scores of two independent medical devices (FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360) each below the cut-off score. After all the assessments, the participating patients completed an anonymous 9-item questionnaire that surveyed their attitudes, personal preferences, and concerns about the use of AI in melanoma detection compared with skin cancer screenings by a dermatologist in addition to sociodemographic data. The dermatologists completed a 2-item questionnaire assessing their attitudes toward the smartphone app in melanoma screening, which was answered specifically for each of the 114 included patients. Neither the patients nor dermatologists were informed about the result of the app\u2019s risk assessment due to the extremely high number of false-positive findings to prevent response bias. Our survey was designed de novo after literature research. All answers were optional. The questionnaire was available in German, English and French and included binary questions (yes/no), multiple choice questions and visual analogue scales (VAS) with scores from 0\u201310.\n\n_2.3. Statistical Analysis_\n\nComparison of continuous variables was tested with the Wilcoxon rank test. Comparison of categorical variables was tested with Fishers Exact test. A p-value of less than 0.05 was deemed as significant. Receiver Operating Characteristics (ROC) analysis was used to assess the performance of the SkinVision[\u00ae] app against the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologist and the SkinVision[\u00ae] app against histology. All analyses were performed with R (version 4.1) and visualized with ggplot2.\n\n_2.4. Ethics_\n\nThe study was approved by the local ethics committee (22020-02482), registered with ClinicalTrials.gov (NCT04605822), and was conducted in compliance with the Declaration of Helsinki and Good Clinical Practice GCP-rules.",
        "results": "**3. Results**\n\n_3.1. Study Population_\n\nOverall, data from 114 patients were analyzed, including 55 patients at high-risk for developing a melanoma (mean age 55 years (age range): 22\u201385), 47% females) and 59 melanoma patients (mean age 60 years (age range: 29\u201381), 54% females) (Table 1). The family history for melanoma was positive in 19% of the melanoma patients and in 56% of the non-melanoma patients. Most of the patients used sunscreen SPF 30\u201350 (51% of melanoma patients vs. 62% of patients at high-risk for melanoma) and had previous sunburns in childhood (54% resp. 69%).\n\n**Table 1. Characteristics of the study population and their skin cancer awareness.**\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Age, n (age range)** 59 (22\u201385) 60 (29\u201381) 55 (22\u201385)\n**Sex, n (%)**\nFemale 58 (51%) 32 (54%) 26 (47%)\nMale 56 (49%) 27 (46%) 29 (53%)\n**Risk profile, n (%)**\nMultiple melanocytic nevi (\u2265100)\nand/or dysplastic nevi (\u22655) and/or positive\nfamily history for melanoma and/or 55 (48%) 0 (0%) 55 (100%)\ndiagnosis of dysplastic nevus syndrome\nand/or CDKN2A mutation\n\nPrevious resected melanoma in situ or\n57 (50%) 57 (97%) 0 (0%)\nprimary cutaneous melanoma\n\nMetastatic melanoma 2 (1.8%) 2 (3.4%) 0 (0%)\n**Positive family history for melanoma,**\n42 (37%) 11 (19%) 31 (56%)\n**n (%)**\n\n**Frequency of skin cancer screening,**\n**n (%)**\nSeveral times per year 40 (35%) 34 (58%) 6 (11%)\nEvery 12 months 39 (34%) 16 (27%) 23 (42%)\nEvery 1\u20132 years 8 (7%) 4 (6.8%) 4 (7.3%)\nEvery 2 years 9 (7.9%) 2 (3.4%) 7 (13%)\nLess than every 2 years 14 (12%) 3 (5.1%) 11 (20%)\nNever 4 (3.5%) 0 (0%) 4 (7.3%)\n**History of sunburns in childhood,**\n70 (61%) 32 (54%) 38 (69%)\n**n (%)**\n\n**Frequency of sunburns (Child), n (%)**\nRarely (less than once per year) 44 (63%) 20 (62%) 24 (63%)\nRegularly (once per year) 22 (31%) 10 (31%) 12 (32%)\nOften (more than once per year) 4 (5.7%) 2 (6.2%) 2 (5.3%)\n**History of sunburns in adulthood,**\n39 (34%) 18 (31%) 21 (38%)\n**n (%)**\n\n\n-----\n\n_Cancers 2022, 14, 3829_ 6 of 21\n\n**Table 1. Cont.**\n\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Frequency of sunburns (Adult), n (%)**\nRarely (less than once per year) 38 (97%) 18 (100%) 20 (95%)\nRegularly (once per year) 0 (0%) 0 (0%) 0 (0%)\nOften (more than once per year) 1 (2.6%) 0 (0%) 1 (4.8%)\n**Previous tanning in the solarium,**\n38 (33%) 13 (22%) 25 (45%)\n**n (%)**\n\n**Usage of sunscreen (SPF), n (%)**\nSPF 6\u201310 2 (1.8%) 1 (1.7%) 1 (1.8%)\nSPF 15\u201325 10 (8.8%) 3 (5.1%) 7 (13%)\nSPF 30\u201350 64 (56%) 30 (51%) 34 (62%)\nSPF 50+ 38 (33%) 25 (42%) 13 (24%)\n\n1 Median (Range); n (%).\n\nA total of 1204 pigmented skin lesions were assessed in this study (Figure 1). In 61 cases (5.1%), we performed a histopathology examination, while 1129 lesions (94%) were diagnosed to be clinically clearly benign based on the combination of a risk assessment by a dermatologist and AI-based risk scores below the cut-off (2D and 3D TBP) and had no indication for obtaining histopathology.\n\n**Figure 1. Flow chart of all included pigmented skin lesions and their histopathological outcome.**\nAI = artificial intelligence; TBP = total body photography.\n\n_3.2. Diagnostic Accuracy and Performance of the Smartphone App SkinVision[\u00ae]_\n\n3.2.1. Comparison of all Risk Assessments\n\nThe smartphone app SkinVision[\u00ae] classified 980 (81%) lesions as benign and indicated an increased risk for melanoma in 224 (19%) cases, while the dermatologists diagnosed 1195 (99.3%) lesions as benign and only nine (0.7%) as suspicious. Consequently, the CE-certified app had a 27-fold higher rate of melanoma-suspicious lesions compared to dermatologists. The AI scores of the 2D and 3D CNN devices classified most lesions as benign and 47 (3.9%) lesions (FotoFinder ATBM[\u00ae]) resp. 39 (3.2%) lesions (Vectra[\u00ae] WB360) as suspicious for melanoma (Figure 2).\n\n-----\n\n_Cancers 2022, 14, 3829_ 7 of 21\n\n**Figure 2. Comparison of risk assessments with the highest rate of suspected melanoma cases by the**\nsmartphone app (n = 1204).\n\nAmong the 224 lesions classified as suspicious by SkinVision[\u00ae], 193 were considered clinically benign by the physician and the 2D and 3D CNN AI-risk assessment scores, whereas four pigmented skin lesions were classified suspicious by the smartphone app, dermatologists, 2D, and 3D TBP devices (Table 2, Supplementary Table S1). The knowledge of AI-based risk assessment scores did not meaningfully affect dermatologists\u2019 classification of skin lesions, as they changed their decision towards the indication for excision in only three lesions that later turned out to be benign (Table 2).\n\n**Table 2. Risk assessments of 1204 pigmented skin lesions by the smartphone app SkinVision[\u00ae], 2D**\nimaging FotoFinder ATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in\ncombination with knowledge of FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360 AI-scores.\n\n**Characteristic** **N = 1204** **[1]**\n\n**Smartphone app SkinVision[\u00ae]**\n\nbenign 980 (81%)\nsuspicious 224 (19%)\n**2D Imaging FotoFinder ATBM[\u00ae]**\n\nbenign 1157 (96%)\nsuspicious 47 (3.9%)\n**3D Imaging VECTRA[\u00ae]** **WB360**\nbenign 1165 (97%)\nsuspicious 39 (3.2%)\n**Dermatologists**\nbenign 1195 (99%)\nsuspicious 9 (0.7%)\n**Dermatologists informed about risk assessment scores by**\n**FotoFinder ATBM[\u00ae]** **+ VECTRA[\u00ae]** **WB360**\nbenign 1192 (99%)\nsuspicious 12 (1.0%)\n\n1 n (%); AI = artificial intelligence.\n\n-----\n\n_Cancers 2022, 14, 3829_ 8 of 21\n\n3.2.2. Diagnostic Accuracy of the Smartphone App Based on the Combination of the\nDermatologist\u2019s Evaluation plus the AI Risk-Assessment Scores of Two Independent\nMedical Devices\n\nReceiver operating characteristic (ROC) analysis of the classification of benign and suspicious lesions of the SkinVision[\u00ae] app compared to the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologists had an area under the curve (AUC) score of 0.621, sensitivity of 0.41, and specificity of 0.83 (Figure 3). Although the specificity is reasonable, the sensitivity is low, thus suggesting that the SkinVision[\u00ae] app has poor diagnostic accuracy.\n\n**Figure 3. Receiver operating characteristic curve of the smartphone app in relation to the results**\nof the combination of risk assessments by dermatologists, FotoFinder ATBM[\u00ae], and Vectra[\u00ae] WB360\n(sensitivity: 41.3%, specificity: 82.9%); AUC = area under the curve.\n\n3.2.3. Diagnostic Accuracy of the Smartphone App Based on Histopathology\n\nAmong 61 pigmented skin lesions examined histologically, we detected six melanomas, 19 melanocytic nevi, 20 dysplastic nevi, as well as 16 otherwise classified lesions were diagnosed (Table 3, Figure 1).\n\n**Table 3. Diagnostic accuracy of the AI-based smartphone app SkinVision[\u00ae], 2D imaging FotoFinder**\nATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in combination with AI in\nmelanoma detection based on histopathology: sensitivity and specificity.\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**Smartphone app SkinVision[\u00ae]** 61\nbenign 13 (68%) 10 (50%) 1 (17%) 10 (62%)\nsuspicious 6 (32%) 10 (50%) 5 (83%) 6 (38%)\n**2D imaging FotoFinder ATBM[\u00ae]** 61\nbenign 7 (37%) 11 (55%) 1 (17%) 4 (25%)\nsuspicious 12 (63%) 9 (45%) 5 (83%) 12 (75%)\n\n-----\n\n_Cancers 2022, 14, 3829_ 9 of 21\n\n**Table 3. Cont.**\n\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**3D imaging VECTRA[\u00ae]** **WB360** 61\nbenign 18 (95%) 9 (45%) 1 (17%) 8 (50%)\nsuspicious 1 (5.3%) 11 (55%) 5 (83%) 8 (50%)\n**Dermatologists** 61\nbenign 17 (89%) 18 (90%) 1 (17%) 16 (100%)\nsuspicious 2 (11%) 2 (10%) 5 (83%) 0 (0%)\n_Beginner: <2 years\u2019 work_\n44 N = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 14 (93%) 10 (83%) 1 (20%) 13 (100%)\nsuspicious 1 (6.7%) 2 (17%) 4 (80%) 0 (0%)\n_Intermediate: 2\u20135 years\u2019 work_\n5 N = 2 N = 3 N = 0 N = 0\n_experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\n11 N = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n**Dermatologists informed about**\n61\n**AI scores** **[2]**\n\nbenign 16 (84%) 17 (85%) 1 (17%) 15 (94%)\nsuspicious 3 (16%) 3 (15%) 5 (83%) 1 (6.2%)\n_Beginner: <2 years\u2019 work_\nN = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 13 (87%) 9 (75%) 1 (20%) 12 (92%)\nsuspicious 2 (13%) 3 (25%) 4 (80%) 1 (7.7%)\n_Intermediate: 2\u20135 years\u2019_\nN = 2 N = 3 N = 0 N = 0\n_work experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\nN = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n\n1 n (%); * Other = pigmented basal cell carcinoma, histiocytoma, lentigo solaris, pigmented seborrhoic keratosis,\nfolliculitis with perifolliculitis, collisional tumor: seborrhoic keratosis and nevus, collisional tumor: actinic\nkeratosis and lentigo solaris; [2] risk assessment scores by FotoFinder ATBM[\u00ae] and VECTRA[\u00ae] WB360; AI = artificial\nintelligence.\n\nBased on histopathology, Figure 4 represents the risk assessments of the dermatologists, the mHealth app, and the combination of AI and dermatologists. Both the SkinVision[\u00ae]\n\napp and the 2D and 3D CNN devices indicated for 5 of 6 histological verified melanomas\n(83% sensitivity) an elevated score indicating suspicion of melanoma. The false-negative rate for all AI-based medical devices was 17%. The three different false-negative classified melanomas were all superficial spreading stage IA melanomas\u2014SkinVision[\u00ae]: superficial spreading melanoma, 0.9 mm Breslow thickness, AJCC stage IA (pT1bN0M0); 3D TBP Vectra[\u00ae] WB360: superficial spreading melanoma, 0.3 mm Breslow thickness, AJCC stage IA (pT1aN0M0); 2D TBP FotoFinder ATBM[\u00ae]: superficial spreading melanoma, 0.7 mm Breslow thickness, AJCC stage IA (pT1aN0M0). The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses was 60.0% of the smartphone app, 63.6% of the 3D Vectra[\u00ae] WB360 imaging device, and 40.0% of the 2D FotoFinder ATBM[\u00ae] risk assessment tool. Dermatologists correctly identified five of six melanomas (83% sensitivity). The false-negative rate was 17%. The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses among all dermatologists was 92.7%. The performance of the dermatologists strongly correlated with their professional experience. Hence, the true-positive rates for melanoma were greater for experts and dermatologists with intermediate experience (100% sensitivity) compared to beginners (80% sensitivity); whereas the true-negative rates were similar (93.3% resp. 92.5% specificity) (Table 3).\n\n**Figure 4. Histology and corresponding diagnosis of lesions assessed by dermatologists, smartphone**\napp SkinVision[\u00ae], and dermatologists and AI (n = 61).\n\nFigure 5 reveals several pigmented skin lesions that were correctly and incorrectly classified as melanomas by the smartphone app SkinVision[\u00ae].\n\n**Figure 5. Correctly and falsely classified melanomas by the smartphone app SkinVision[\u00ae]: (A,B). True-**\npositive classified melanoma; (C,D). False-positive classified melanoma; (E,F). True-negative classified\nmelanoma; (G). False-negative classified melanoma.\n\nROC analysis of the classification of melanoma of the SkinVision[\u00ae] app compared\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision[\u00ae] app had\nhigh sensitivity but mediocre specificity.\n\nROC analysis of the classification of melanoma of the SkinVision\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision\nhigh sensitivity but mediocre specificity.\n\n-----\n\n_Cancers 2022, 14, 3829_ 11 of 21\n\n**Figure 6. Receiver operating characteristic curve of the smartphone app in relation to the results of**\nthe histology (sensitivity of 83.3%, specificity 60.0%); AUC = area under the curve.\n\n_3.3. Patient Perspective on AI in Melanoma Screening_\n3.3.1. Confidence in Dermatologists vs. Smartphone App\n\nMost of the patients at high-risk for melanoma (55% (30/55)) and patients with\nmelanoma (53% (31/59)) reported being very confident about a mole examination by\na dermatologist (rating scale 10/10). In contrast, only a minority of high-risk patients (16%\n(9/55)) and melanoma patients (12% (7/59)) felt very safe when being investigated by a\nsmartphone app alone. No significant difference in ratings was identified between the two\nrisk groups (p < 0.9; p < 0.7) (Supplementary Table S2).\n\n3.3.2. Trustworthiness of the Smartphone App\n\nAll participants primarily rated physician examination (100% (55/55) among highrisk patients, 100% (59/59) among melanoma patients) as trustworthy, and the majority\ndid so for 2D TBP imaging (93% (51/55) resp. 88% (52/59)) and 3D TBP imaging (91%\n(50/55) resp. 90% (53/59)). The smartphone app was less frequently rated as trustworthy,\nwith only 36% (20/55) among patients at high-risk for melanoma and 49% (29/59) among\nmelanoma patients (Table 4). The age revealed a significant correlation with the evaluation\nof smartphones\u2019 trustworthiness (p < 0.004), with older patients (>60 years old) having\ntrusted the app three times more than younger patients (\u226460 years old) (Figure 7). Neither\nprevious melanoma vs. high-risk criteria for melanoma nor sex significantly influenced the\nevaluation of the trustworthiness of the smartphone app.\n\n-----\n\n_Cancers 2022, 14, 3829_ 12 of 21\n\n**Table 4. Assessment of trustworthiness of the AI-based smartphone app SkinVision[\u00ae], 2D imaging**\nFotoFinder ATBM[\u00ae], and 3D imaging Vectra[\u00ae] WB360 compared to dermatologists.\n\n\n**Patients at**\n**High-Risk**\n**for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Characteristic** **N**\n\n\n**Patients**\n**with**\n**Melanoma,**\n**N = 59** **[1]**\n\n\n**_p-Value_** **[2]**\n\n\n**The following examination was**\n**trustworthy: Smartphone** 114 0.3\n**app assessment**\n\nYes 29 (49%) 20 (36%)\nNo 5 (8.5%) 8 (15%)\nI don\u2019t know 23 (39%) 22 (40%)\nNo answer 2 (3.4%) 5 (9.1%)\n**Dermatologist assessment** 114\nYes 59 (100%) 55 (100%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 0 (0%) 0 (0%)\nNo answer 0 (0%) 0 (0%)\n**2D TBP assessment** 114 0.3\nYes 52 (88%) 51 (93%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 7 (12%) 3 (5.5%)\nNo answer 0 (0%) 1 (1.8%)\n**3D TBP assessment** 114 0.3\nYes 53 (90%) 50 (91%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 6 (10%) 3 (5.5%)\nNo answer 0 (0%) 2 (3.6%)\n\n1 n (%); 2 Fisher\u2019s exact test; Pearson\u2019s Chi-squared test; TBP = total body photography.\n\n**Figure 7. Odds ratio for variables influencing the trustworthiness of smartphones\u2019 risk assessment in**\nmelanoma detection.\n\n-----\n\n_Cancers 2022, 14, 3829_ 13 of 21\n\n3.3.3. Impact of AI vs. Dermatologists\u2019 Examination on Patients\u2019 Fear of Developing\nSkin Cancer\n\nMost participants indicated that an examination by a physician reduced their fear of\ndeveloping skin cancer, namely in 89% (49/55) among high-risk patients and 81% (48/59)\namong melanoma patients. Comparably, the 2D TBP imaging achieved the same effect in\n78% (43/55) resp. 76% (45/59), and the 3D TBP device in 82% (45/55) resp. 75% (44/59). In\ncontrast, the assessment with the smartphone app appeased fear of skin cancer in only 33%\n(18/55) of patients at high-risk for melanoma and in 32% (19/59) of patients with melanoma.\nIndeed, one high-risk patient (1.8%) even reported an increased fear of developing skin\ncancer by using the smartphone app (Supplementary Table S2).\n\n3.3.4. Patients\u2019 Subjective Assessment of the Accuracy of AI vs. Dermatologists\n\nPatients expected reliable results with the highest accuracy by both the assessment\nby a physician (98% (54/55) among high-risk patients, 92% (54/59) among melanoma\npatients) and by the 2D TBP imaging (82% (45/55) resp. 86% (51/59)) as well as the 3D\nTBP device (89% (49/55) resp. 88% (52/59)). Only 16% (9/55) of high-risk patients and\n31% (18/59) of melanoma patients expected reliable results from the smartphone app\n(Supplementary Table S2). No significant differences were identified between the two risk\ngroups in the evaluations.\n\n3.3.5. Patient Preference for Skin Cancer Screening\n\nBoth cohorts favored a combination of dermatologist and 3D TBP risk assessment for\nthe examination of pigmented skin lesions (64% (35/55) among patients at high-risk for\nmelanoma, 51% (30/59) among melanoma patients), while neither preferred assessment\nby a smartphone app alone (Figure 8A, Supplementary Table S2). The combination of\ndermatologist and smartphone app was favored by only 1.8% (1/55) of patients at high-risk\nfor melanoma and 3.4% (2/59) of patients with melanoma.\n\n**Figure 8. Cont.**\n\n-----\n\n_Cancers 2022, 14, 3829_ 14 of 21\n\n**Figure 8. (A) Patient preference for mole assessment (patients at high-risk for melanoma, n = 55;**\npatients with melanoma, n = 59); (B) Patient preference for AI in skin cancer screening (patients at\nhigh-risk for melanoma, n = 55; patients with melanoma, n = 59).\n\nRegarding patient preference for skin cancer screenings, almost all high-risk patients\n(98% (54/55)) and melanoma patients (95% (56/59)) indicated their belief that AI can improve\na physician\u2019s diagnostic performance. Most patients (64% (35/55) among high-risk patients,\n54% (32/59) among melanoma patients) would prefer that the physicians would always\nconsider the result of AI in their diagnosis (Figure 8B) (Supplementary Table S2).\n\n3.3.6. Dermatologists\u2019 Perspective of Smartphone Apps for Melanoma Screening\n\nAmong the seven dermatologists, they stated in only 5.3% of the skin cancer screenings\n(6/114) that the smartphone app increased diagnostic confidence and in only 8.8% of the\nassessments (10/114) they trusted the app (Supplementary Table S3).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "High"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Over-Detection of Melanoma-Suspect Lesions by a CE-Certified Smartphone App: Performance in Comparison to Dermatologists, 2D and 3D Convolutional Neural Networks in a Prospective Data Set of 1204 Pigmented Skin Lesions Involving Patients' Perception.",
        "abstract": "The exponential increase in algorithm-based mobile health (mHealth) applications (apps) for melanoma screening is a reaction to a growing market. However, the performance of available apps remains to be investigated. In this prospective study, we investigated the diagnostic accuracy of a class 1 CE-certified smartphone app in melanoma risk stratification and its patient and dermatologist satisfaction. Pigmented skin lesions \u2265 3 mm and any suspicious smaller lesions were assessed by the smartphone app SkinVision\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Study Design and Participants_\n\nWe performed this prospective, single-center, comparative observational cohort study at the Department of Dermatology at the University Hospital Basel in Switzerland between January 2021 to June 2021. Seven dermatologists (beginners: <2 years professional experience, n = 4; intermediate: 2\u20135 years professional experience, n = 1; experts: >5 years professional experience, n = 2), as well as 114 patients were included in the study with an age of \u226518 years with a high risk for developing a primary cutaneous malignant melanoma defined as previous melanoma (including melanoma in situ), \u2265100 melanocytic nevi, \u22655 atypical nevi, strong family history for melanoma, diagnosis of dysplastic nevus syndrome, or known CDKN2A mutation. Exclusion criteria were acute psychiatric illness or acute crisis and Fitzpatrick sign type V-VI.\n\n_2.2. Procedures_\n\nIn the first step, the patients signed the informed consent after they were informed by a dermatologist about the study design and the following examinations. Subsequently, the dermatologist obtained a medical history, including a history of melanoma to enable classification into the different risk groups. The classification of all high-risk patients into further risk cohorts are defined by how often the patients require skin cancer screening. Based on this, we have classified our melanoma patients as high risk, as they come to dermatological control every 3\u201312 months according to the corresponding AJCC stage. Patients without melanoma in their own medical history are seen every 12 months. All study participants underwent a standard skin cancer screening by a dermatologist (Supplementary Figure S1). The pigmented skin lesions were assessed with dermatoscopy and classified as either benign or malignant. The dermatologist defined pigmented skin lesions that were suspected melanoma and indicated whether lesions smaller than 3 mm should be classified with the different AI modalities due to suspicion. The regular skin cancer screening was performed before any AI-based risk assessments were achieved, so the dermatologists had no knowledge about AI classification of the melanocytic lesions up to this point. In the next step, all melanocytic nevi \u2265 3 mm or any smaller suspicious melanocytic lesions were assessed using an iOS-based iPhone SE smartphone equipped with a 12-megapixel camera and the app SkinVision[\u00ae] (SkinVision[\u00ae] B.V., Amsterdam, the Netherlands, App-Version 6.8.1), which is based on a machine-learning algorithm classifying pigmented lesions into low, medium, and high risk for melanoma. Only the lesions classified as high-risk\u2014indicating a recommendation to consult a dermatologist\u2014were recorded. Afterwards, patients received a whole-body screening using the 2D automated total body mapping (ATBM[\u00ae]) master of FotoFinder (FotoFinder ATBM[\u00ae] Systems GmbH, Bad Birnbach, Germany, Version 3.3.1.0, risk scoring with MoleAnalyzer Pro), and the 3D TBP Vectra[\u00ae] WB360 System (Canfield Scientific, Parsippany, NJ, USA, Version 4.7.1, risk scoring with DEXI) which was performed by medical students and a study nurse. Subsequently, all lesions over 3 mm or any smaller suspicious lesions were assessed by the 2D and 3D AI devices using their dermoscopes (dermoscope medicam 1000 (FotoFinder ATBM[\u00ae]) and VISIOMED[\u00ae] D200evo dermoscope (Vectra[\u00ae] WB360)). For dermoscopic imaged melanocytic nevi, we obtained AI-based risk scores between 0.0\u20131.0 for 2D FotoFinder ATBM[\u00ae] and 0.0\u201310.0 for 3D TBP Vectra[\u00ae] WB360. FotoFinder\u2019s ATBM[\u00ae] master creates a 2D-image of the entire surface of the patient using 20 photos taken from 8 different parts of the body, while Vectra[\u00ae] WB360 generates a 3D-image using 92 individual photos that are converted into a 3D Avatar. Finally, the dermatologists assessed the skin lesions a second time with knowledge of the AI risk assessment scores of 2D FotoFinder ATBM[\u00ae] and 3D TBP Vectra[\u00ae] WB360. The patient was informed about the lesions with suspicion for melanoma by a dermatologist and about the 2D and 3D CNN classifications above the study-defined cut-off scores. The indication for excision of suspicious lesions was based on the dermatologists\u2019 suspicion for melanoma and/or malignancy risk assessment scores of the 2D imaging tool FotoFinder ATBM[\u00ae] and/or 3D imaging tool Vectra[\u00ae] WB360 indicating values over the study-defined cut-off score (>0.5 resp. >5.0). The patient was informed which lesions would be removed with the appropriate rationale. A maximum of two excisions were performed per visit and per patient due to ethical reasons, with an exception for further biopsies in cases of high-grade suspicion of malignancy by the dermatologist or patient request. The reference standard for a benign lesion was either the histological diagnosis (no melanoma) or the combination of the dermatologist\u2019s evaluation (benign lesion) plus the AI scores of two independent medical devices (FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360) each below the cut-off score. After all the assessments, the participating patients completed an anonymous 9-item questionnaire that surveyed their attitudes, personal preferences, and concerns about the use of AI in melanoma detection compared with skin cancer screenings by a dermatologist in addition to sociodemographic data. The dermatologists completed a 2-item questionnaire assessing their attitudes toward the smartphone app in melanoma screening, which was answered specifically for each of the 114 included patients. Neither the patients nor dermatologists were informed about the result of the app\u2019s risk assessment due to the extremely high number of false-positive findings to prevent response bias. Our survey was designed de novo after literature research. All answers were optional. The questionnaire was available in German, English and French and included binary questions (yes/no), multiple choice questions and visual analogue scales (VAS) with scores from 0\u201310.\n\n_2.3. Statistical Analysis_\n\nComparison of continuous variables was tested with the Wilcoxon rank test. Comparison of categorical variables was tested with Fishers Exact test. A p-value of less than 0.05 was deemed as significant. Receiver Operating Characteristics (ROC) analysis was used to assess the performance of the SkinVision[\u00ae] app against the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologist and the SkinVision[\u00ae] app against histology. All analyses were performed with R (version 4.1) and visualized with ggplot2.\n\n_2.4. Ethics_\n\nThe study was approved by the local ethics committee (22020-02482), registered with ClinicalTrials.gov (NCT04605822), and was conducted in compliance with the Declaration of Helsinki and Good Clinical Practice GCP-rules.",
        "results": "**3. Results**\n\n_3.1. Study Population_\n\nOverall, data from 114 patients were analyzed, including 55 patients at high-risk for developing a melanoma (mean age 55 years (age range): 22\u201385), 47% females) and 59 melanoma patients (mean age 60 years (age range: 29\u201381), 54% females) (Table 1). The family history for melanoma was positive in 19% of the melanoma patients and in 56% of the non-melanoma patients. Most of the patients used sunscreen SPF 30\u201350 (51% of melanoma patients vs. 62% of patients at high-risk for melanoma) and had previous sunburns in childhood (54% resp. 69%).\n\n**Table 1. Characteristics of the study population and their skin cancer awareness.**\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Age, n (age range)** 59 (22\u201385) 60 (29\u201381) 55 (22\u201385)\n**Sex, n (%)**\nFemale 58 (51%) 32 (54%) 26 (47%)\nMale 56 (49%) 27 (46%) 29 (53%)\n**Risk profile, n (%)**\nMultiple melanocytic nevi (\u2265100)\nand/or dysplastic nevi (\u22655) and/or positive\nfamily history for melanoma and/or 55 (48%) 0 (0%) 55 (100%)\ndiagnosis of dysplastic nevus syndrome\nand/or CDKN2A mutation\n\nPrevious resected melanoma in situ or\n57 (50%) 57 (97%) 0 (0%)\nprimary cutaneous melanoma\n\nMetastatic melanoma 2 (1.8%) 2 (3.4%) 0 (0%)\n**Positive family history for melanoma,**\n42 (37%) 11 (19%) 31 (56%)\n**n (%)**\n\n**Frequency of skin cancer screening,**\n**n (%)**\nSeveral times per year 40 (35%) 34 (58%) 6 (11%)\nEvery 12 months 39 (34%) 16 (27%) 23 (42%)\nEvery 1\u20132 years 8 (7%) 4 (6.8%) 4 (7.3%)\nEvery 2 years 9 (7.9%) 2 (3.4%) 7 (13%)\nLess than every 2 years 14 (12%) 3 (5.1%) 11 (20%)\nNever 4 (3.5%) 0 (0%) 4 (7.3%)\n**History of sunburns in childhood,**\n70 (61%) 32 (54%) 38 (69%)\n**n (%)**\n\n**Frequency of sunburns (Child), n (%)**\nRarely (less than once per year) 44 (63%) 20 (62%) 24 (63%)\nRegularly (once per year) 22 (31%) 10 (31%) 12 (32%)\nOften (more than once per year) 4 (5.7%) 2 (6.2%) 2 (5.3%)\n**History of sunburns in adulthood,**\n39 (34%) 18 (31%) 21 (38%)\n**n (%)**\n\n\n-----\n\n_Cancers 2022, 14, 3829_ 6 of 21\n\n**Table 1. Cont.**\n\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Frequency of sunburns (Adult), n (%)**\nRarely (less than once per year) 38 (97%) 18 (100%) 20 (95%)\nRegularly (once per year) 0 (0%) 0 (0%) 0 (0%)\nOften (more than once per year) 1 (2.6%) 0 (0%) 1 (4.8%)\n**Previous tanning in the solarium,**\n38 (33%) 13 (22%) 25 (45%)\n**n (%)**\n\n**Usage of sunscreen (SPF), n (%)**\nSPF 6\u201310 2 (1.8%) 1 (1.7%) 1 (1.8%)\nSPF 15\u201325 10 (8.8%) 3 (5.1%) 7 (13%)\nSPF 30\u201350 64 (56%) 30 (51%) 34 (62%)\nSPF 50+ 38 (33%) 25 (42%) 13 (24%)\n\n1 Median (Range); n (%).\n\nA total of 1204 pigmented skin lesions were assessed in this study (Figure 1). In 61 cases (5.1%), we performed a histopathology examination, while 1129 lesions (94%) were diagnosed to be clinically clearly benign based on the combination of a risk assessment by a dermatologist and AI-based risk scores below the cut-off (2D and 3D TBP) and had no indication for obtaining histopathology.\n\n**Figure 1. Flow chart of all included pigmented skin lesions and their histopathological outcome.**\nAI = artificial intelligence; TBP = total body photography.\n\n_3.2. Diagnostic Accuracy and Performance of the Smartphone App SkinVision[\u00ae]_\n\n3.2.1. Comparison of all Risk Assessments\n\nThe smartphone app SkinVision[\u00ae] classified 980 (81%) lesions as benign and indicated an increased risk for melanoma in 224 (19%) cases, while the dermatologists diagnosed 1195 (99.3%) lesions as benign and only nine (0.7%) as suspicious. Consequently, the CE-certified app had a 27-fold higher rate of melanoma-suspicious lesions compared to dermatologists. The AI scores of the 2D and 3D CNN devices classified most lesions as benign and 47 (3.9%) lesions (FotoFinder ATBM[\u00ae]) resp. 39 (3.2%) lesions (Vectra[\u00ae] WB360) as suspicious for melanoma (Figure 2).\n\n-----\n\n_Cancers 2022, 14, 3829_ 7 of 21\n\n**Figure 2. Comparison of risk assessments with the highest rate of suspected melanoma cases by the**\nsmartphone app (n = 1204).\n\nAmong the 224 lesions classified as suspicious by SkinVision[\u00ae], 193 were considered clinically benign by the physician and the 2D and 3D CNN AI-risk assessment scores, whereas four pigmented skin lesions were classified suspicious by the smartphone app, dermatologists, 2D, and 3D TBP devices (Table 2, Supplementary Table S1). The knowledge of AI-based risk assessment scores did not meaningfully affect dermatologists\u2019 classification of skin lesions, as they changed their decision towards the indication for excision in only three lesions that later turned out to be benign (Table 2).\n\n**Table 2. Risk assessments of 1204 pigmented skin lesions by the smartphone app SkinVision[\u00ae], 2D**\nimaging FotoFinder ATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in\ncombination with knowledge of FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360 AI-scores.\n\n**Characteristic** **N = 1204** **[1]**\n\n**Smartphone app SkinVision[\u00ae]**\n\nbenign 980 (81%)\nsuspicious 224 (19%)\n**2D Imaging FotoFinder ATBM[\u00ae]**\n\nbenign 1157 (96%)\nsuspicious 47 (3.9%)\n**3D Imaging VECTRA[\u00ae]** **WB360**\nbenign 1165 (97%)\nsuspicious 39 (3.2%)\n**Dermatologists**\nbenign 1195 (99%)\nsuspicious 9 (0.7%)\n**Dermatologists informed about risk assessment scores by**\n**FotoFinder ATBM[\u00ae]** **+ VECTRA[\u00ae]** **WB360**\nbenign 1192 (99%)\nsuspicious 12 (1.0%)\n\n1 n (%); AI = artificial intelligence.\n\n-----\n\n_Cancers 2022, 14, 3829_ 8 of 21\n\n3.2.2. Diagnostic Accuracy of the Smartphone App Based on the Combination of the\nDermatologist\u2019s Evaluation plus the AI Risk-Assessment Scores of Two Independent\nMedical Devices\n\nReceiver operating characteristic (ROC) analysis of the classification of benign and suspicious lesions of the SkinVision[\u00ae] app compared to the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologists had an area under the curve (AUC) score of 0.621, sensitivity of 0.41, and specificity of 0.83 (Figure 3). Although the specificity is reasonable, the sensitivity is low, thus suggesting that the SkinVision[\u00ae] app has poor diagnostic accuracy.\n\n**Figure 3. Receiver operating characteristic curve of the smartphone app in relation to the results**\nof the combination of risk assessments by dermatologists, FotoFinder ATBM[\u00ae], and Vectra[\u00ae] WB360\n(sensitivity: 41.3%, specificity: 82.9%); AUC = area under the curve.\n\n3.2.3. Diagnostic Accuracy of the Smartphone App Based on Histopathology\n\nAmong 61 pigmented skin lesions examined histologically, we detected six melanomas, 19 melanocytic nevi, 20 dysplastic nevi, as well as 16 otherwise classified lesions were diagnosed (Table 3, Figure 1).\n\n**Table 3. Diagnostic accuracy of the AI-based smartphone app SkinVision[\u00ae], 2D imaging FotoFinder**\nATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in combination with AI in\nmelanoma detection based on histopathology: sensitivity and specificity.\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**Smartphone app SkinVision[\u00ae]** 61\nbenign 13 (68%) 10 (50%) 1 (17%) 10 (62%)\nsuspicious 6 (32%) 10 (50%) 5 (83%) 6 (38%)\n**2D imaging FotoFinder ATBM[\u00ae]** 61\nbenign 7 (37%) 11 (55%) 1 (17%) 4 (25%)\nsuspicious 12 (63%) 9 (45%) 5 (83%) 12 (75%)\n\n-----\n\n_Cancers 2022, 14, 3829_ 9 of 21\n\n**Table 3. Cont.**\n\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**3D imaging VECTRA[\u00ae]** **WB360** 61\nbenign 18 (95%) 9 (45%) 1 (17%) 8 (50%)\nsuspicious 1 (5.3%) 11 (55%) 5 (83%) 8 (50%)\n**Dermatologists** 61\nbenign 17 (89%) 18 (90%) 1 (17%) 16 (100%)\nsuspicious 2 (11%) 2 (10%) 5 (83%) 0 (0%)\n_Beginner: <2 years\u2019 work_\n44 N = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 14 (93%) 10 (83%) 1 (20%) 13 (100%)\nsuspicious 1 (6.7%) 2 (17%) 4 (80%) 0 (0%)\n_Intermediate: 2\u20135 years\u2019 work_\n5 N = 2 N = 3 N = 0 N = 0\n_experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\n11 N = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n**Dermatologists informed about**\n61\n**AI scores** **[2]**\n\nbenign 16 (84%) 17 (85%) 1 (17%) 15 (94%)\nsuspicious 3 (16%) 3 (15%) 5 (83%) 1 (6.2%)\n_Beginner: <2 years\u2019 work_\nN = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 13 (87%) 9 (75%) 1 (20%) 12 (92%)\nsuspicious 2 (13%) 3 (25%) 4 (80%) 1 (7.7%)\n_Intermediate: 2\u20135 years\u2019_\nN = 2 N = 3 N = 0 N = 0\n_work experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\nN = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n\n1 n (%); * Other = pigmented basal cell carcinoma, histiocytoma, lentigo solaris, pigmented seborrhoic keratosis,\nfolliculitis with perifolliculitis, collisional tumor: seborrhoic keratosis and nevus, collisional tumor: actinic\nkeratosis and lentigo solaris; [2] risk assessment scores by FotoFinder ATBM[\u00ae] and VECTRA[\u00ae] WB360; AI = artificial\nintelligence.\n\nBased on histopathology, Figure 4 represents the risk assessments of the dermatologists, the mHealth app, and the combination of AI and dermatologists. Both the SkinVision[\u00ae]\n\napp and the 2D and 3D CNN devices indicated for 5 of 6 histological verified melanomas\n(83% sensitivity) an elevated score indicating suspicion of melanoma. The false-negative rate for all AI-based medical devices was 17%. The three different false-negative classified melanomas were all superficial spreading stage IA melanomas\u2014SkinVision[\u00ae]: superficial spreading melanoma, 0.9 mm Breslow thickness, AJCC stage IA (pT1bN0M0); 3D TBP Vectra[\u00ae] WB360: superficial spreading melanoma, 0.3 mm Breslow thickness, AJCC stage IA (pT1aN0M0); 2D TBP FotoFinder ATBM[\u00ae]: superficial spreading melanoma, 0.7 mm Breslow thickness, AJCC stage IA (pT1aN0M0). The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses was 60.0% of the smartphone app, 63.6% of the 3D Vectra[\u00ae] WB360 imaging device, and 40.0% of the 2D FotoFinder ATBM[\u00ae] risk assessment tool. Dermatologists correctly identified five of six melanomas (83% sensitivity). The false-negative rate was 17%. The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses among all dermatologists was 92.7%. The performance of the dermatologists strongly correlated with their professional experience. Hence, the true-positive rates for melanoma were greater for experts and dermatologists with intermediate experience (100% sensitivity) compared to beginners (80% sensitivity); whereas the true-negative rates were similar (93.3% resp. 92.5% specificity) (Table 3).\n\n**Figure 4. Histology and corresponding diagnosis of lesions assessed by dermatologists, smartphone**\napp SkinVision[\u00ae], and dermatologists and AI (n = 61).\n\nFigure 5 reveals several pigmented skin lesions that were correctly and incorrectly classified as melanomas by the smartphone app SkinVision[\u00ae].\n\n**Figure 5. Correctly and falsely classified melanomas by the smartphone app SkinVision[\u00ae]: (A,B). True-**\npositive classified melanoma; (C,D). False-positive classified melanoma; (E,F). True-negative classified\nmelanoma; (G). False-negative classified melanoma.\n\nROC analysis of the classification of melanoma of the SkinVision[\u00ae] app compared\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision[\u00ae] app had\nhigh sensitivity but mediocre specificity.\n\nROC analysis of the classification of melanoma of the SkinVision\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision\nhigh sensitivity but mediocre specificity.\n\n-----\n\n_Cancers 2022, 14, 3829_ 11 of 21\n\n**Figure 6. Receiver operating characteristic curve of the smartphone app in relation to the results of**\nthe histology (sensitivity of 83.3%, specificity 60.0%); AUC = area under the curve.\n\n_3.3. Patient Perspective on AI in Melanoma Screening_\n3.3.1. Confidence in Dermatologists vs. Smartphone App\n\nMost of the patients at high-risk for melanoma (55% (30/55)) and patients with\nmelanoma (53% (31/59)) reported being very confident about a mole examination by\na dermatologist (rating scale 10/10). In contrast, only a minority of high-risk patients (16%\n(9/55)) and melanoma patients (12% (7/59)) felt very safe when being investigated by a\nsmartphone app alone. No significant difference in ratings was identified between the two\nrisk groups (p < 0.9; p < 0.7) (Supplementary Table S2).\n\n3.3.2. Trustworthiness of the Smartphone App\n\nAll participants primarily rated physician examination (100% (55/55) among highrisk patients, 100% (59/59) among melanoma patients) as trustworthy, and the majority\ndid so for 2D TBP imaging (93% (51/55) resp. 88% (52/59)) and 3D TBP imaging (91%\n(50/55) resp. 90% (53/59)). The smartphone app was less frequently rated as trustworthy,\nwith only 36% (20/55) among patients at high-risk for melanoma and 49% (29/59) among\nmelanoma patients (Table 4). The age revealed a significant correlation with the evaluation\nof smartphones\u2019 trustworthiness (p < 0.004), with older patients (>60 years old) having\ntrusted the app three times more than younger patients (\u226460 years old) (Figure 7). Neither\nprevious melanoma vs. high-risk criteria for melanoma nor sex significantly influenced the\nevaluation of the trustworthiness of the smartphone app.\n\n-----\n\n_Cancers 2022, 14, 3829_ 12 of 21\n\n**Table 4. Assessment of trustworthiness of the AI-based smartphone app SkinVision[\u00ae], 2D imaging**\nFotoFinder ATBM[\u00ae], and 3D imaging Vectra[\u00ae] WB360 compared to dermatologists.\n\n\n**Patients at**\n**High-Risk**\n**for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Characteristic** **N**\n\n\n**Patients**\n**with**\n**Melanoma,**\n**N = 59** **[1]**\n\n\n**_p-Value_** **[2]**\n\n\n**The following examination was**\n**trustworthy: Smartphone** 114 0.3\n**app assessment**\n\nYes 29 (49%) 20 (36%)\nNo 5 (8.5%) 8 (15%)\nI don\u2019t know 23 (39%) 22 (40%)\nNo answer 2 (3.4%) 5 (9.1%)\n**Dermatologist assessment** 114\nYes 59 (100%) 55 (100%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 0 (0%) 0 (0%)\nNo answer 0 (0%) 0 (0%)\n**2D TBP assessment** 114 0.3\nYes 52 (88%) 51 (93%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 7 (12%) 3 (5.5%)\nNo answer 0 (0%) 1 (1.8%)\n**3D TBP assessment** 114 0.3\nYes 53 (90%) 50 (91%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 6 (10%) 3 (5.5%)\nNo answer 0 (0%) 2 (3.6%)\n\n1 n (%); 2 Fisher\u2019s exact test; Pearson\u2019s Chi-squared test; TBP = total body photography.\n\n**Figure 7. Odds ratio for variables influencing the trustworthiness of smartphones\u2019 risk assessment in**\nmelanoma detection.\n\n-----\n\n_Cancers 2022, 14, 3829_ 13 of 21\n\n3.3.3. Impact of AI vs. Dermatologists\u2019 Examination on Patients\u2019 Fear of Developing\nSkin Cancer\n\nMost participants indicated that an examination by a physician reduced their fear of\ndeveloping skin cancer, namely in 89% (49/55) among high-risk patients and 81% (48/59)\namong melanoma patients. Comparably, the 2D TBP imaging achieved the same effect in\n78% (43/55) resp. 76% (45/59), and the 3D TBP device in 82% (45/55) resp. 75% (44/59). In\ncontrast, the assessment with the smartphone app appeased fear of skin cancer in only 33%\n(18/55) of patients at high-risk for melanoma and in 32% (19/59) of patients with melanoma.\nIndeed, one high-risk patient (1.8%) even reported an increased fear of developing skin\ncancer by using the smartphone app (Supplementary Table S2).\n\n3.3.4. Patients\u2019 Subjective Assessment of the Accuracy of AI vs. Dermatologists\n\nPatients expected reliable results with the highest accuracy by both the assessment\nby a physician (98% (54/55) among high-risk patients, 92% (54/59) among melanoma\npatients) and by the 2D TBP imaging (82% (45/55) resp. 86% (51/59)) as well as the 3D\nTBP device (89% (49/55) resp. 88% (52/59)). Only 16% (9/55) of high-risk patients and\n31% (18/59) of melanoma patients expected reliable results from the smartphone app\n(Supplementary Table S2). No significant differences were identified between the two risk\ngroups in the evaluations.\n\n3.3.5. Patient Preference for Skin Cancer Screening\n\nBoth cohorts favored a combination of dermatologist and 3D TBP risk assessment for\nthe examination of pigmented skin lesions (64% (35/55) among patients at high-risk for\nmelanoma, 51% (30/59) among melanoma patients), while neither preferred assessment\nby a smartphone app alone (Figure 8A, Supplementary Table S2). The combination of\ndermatologist and smartphone app was favored by only 1.8% (1/55) of patients at high-risk\nfor melanoma and 3.4% (2/59) of patients with melanoma.\n\n**Figure 8. Cont.**\n\n-----\n\n_Cancers 2022, 14, 3829_ 14 of 21\n\n**Figure 8. (A) Patient preference for mole assessment (patients at high-risk for melanoma, n = 55;**\npatients with melanoma, n = 59); (B) Patient preference for AI in skin cancer screening (patients at\nhigh-risk for melanoma, n = 55; patients with melanoma, n = 59).\n\nRegarding patient preference for skin cancer screenings, almost all high-risk patients\n(98% (54/55)) and melanoma patients (95% (56/59)) indicated their belief that AI can improve\na physician\u2019s diagnostic performance. Most patients (64% (35/55) among high-risk patients,\n54% (32/59) among melanoma patients) would prefer that the physicians would always\nconsider the result of AI in their diagnosis (Figure 8B) (Supplementary Table S2).\n\n3.3.6. Dermatologists\u2019 Perspective of Smartphone Apps for Melanoma Screening\n\nAmong the seven dermatologists, they stated in only 5.3% of the skin cancer screenings\n(6/114) that the smartphone app increased diagnostic confidence and in only 8.8% of the\nassessments (10/114) they trusted the app (Supplementary Table S3).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "High"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Over-Detection of Melanoma-Suspect Lesions by a CE-Certified Smartphone App: Performance in Comparison to Dermatologists, 2D and 3D Convolutional Neural Networks in a Prospective Data Set of 1204 Pigmented Skin Lesions Involving Patients' Perception.",
        "abstract": "The exponential increase in algorithm-based mobile health (mHealth) applications (apps) for melanoma screening is a reaction to a growing market. However, the performance of available apps remains to be investigated. In this prospective study, we investigated the diagnostic accuracy of a class 1 CE-certified smartphone app in melanoma risk stratification and its patient and dermatologist satisfaction. Pigmented skin lesions \u2265 3 mm and any suspicious smaller lesions were assessed by the smartphone app SkinVision\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Study Design and Participants_\n\nWe performed this prospective, single-center, comparative observational cohort study at the Department of Dermatology at the University Hospital Basel in Switzerland between January 2021 to June 2021. Seven dermatologists (beginners: <2 years professional experience, n = 4; intermediate: 2\u20135 years professional experience, n = 1; experts: >5 years professional experience, n = 2), as well as 114 patients were included in the study with an age of \u226518 years with a high risk for developing a primary cutaneous malignant melanoma defined as previous melanoma (including melanoma in situ), \u2265100 melanocytic nevi, \u22655 atypical nevi, strong family history for melanoma, diagnosis of dysplastic nevus syndrome, or known CDKN2A mutation. Exclusion criteria were acute psychiatric illness or acute crisis and Fitzpatrick sign type V-VI.\n\n_2.2. Procedures_\n\nIn the first step, the patients signed the informed consent after they were informed by a dermatologist about the study design and the following examinations. Subsequently, the dermatologist obtained a medical history, including a history of melanoma to enable classification into the different risk groups. The classification of all high-risk patients into further risk cohorts are defined by how often the patients require skin cancer screening. Based on this, we have classified our melanoma patients as high risk, as they come to dermatological control every 3\u201312 months according to the corresponding AJCC stage. Patients without melanoma in their own medical history are seen every 12 months. All study participants underwent a standard skin cancer screening by a dermatologist (Supplementary Figure S1). The pigmented skin lesions were assessed with dermatoscopy and classified as either benign or malignant. The dermatologist defined pigmented skin lesions that were suspected melanoma and indicated whether lesions smaller than 3 mm should be classified with the different AI modalities due to suspicion. The regular skin cancer screening was performed before any AI-based risk assessments were achieved, so the dermatologists had no knowledge about AI classification of the melanocytic lesions up to this point. In the next step, all melanocytic nevi \u2265 3 mm or any smaller suspicious melanocytic lesions were assessed using an iOS-based iPhone SE smartphone equipped with a 12-megapixel camera and the app SkinVision[\u00ae] (SkinVision[\u00ae] B.V., Amsterdam, the Netherlands, App-Version 6.8.1), which is based on a machine-learning algorithm classifying pigmented lesions into low, medium, and high risk for melanoma. Only the lesions classified as high-risk\u2014indicating a recommendation to consult a dermatologist\u2014were recorded. Afterwards, patients received a whole-body screening using the 2D automated total body mapping (ATBM[\u00ae]) master of FotoFinder (FotoFinder ATBM[\u00ae] Systems GmbH, Bad Birnbach, Germany, Version 3.3.1.0, risk scoring with MoleAnalyzer Pro), and the 3D TBP Vectra[\u00ae] WB360 System (Canfield Scientific, Parsippany, NJ, USA, Version 4.7.1, risk scoring with DEXI) which was performed by medical students and a study nurse. Subsequently, all lesions over 3 mm or any smaller suspicious lesions were assessed by the 2D and 3D AI devices using their dermoscopes (dermoscope medicam 1000 (FotoFinder ATBM[\u00ae]) and VISIOMED[\u00ae] D200evo dermoscope (Vectra[\u00ae] WB360)). For dermoscopic imaged melanocytic nevi, we obtained AI-based risk scores between 0.0\u20131.0 for 2D FotoFinder ATBM[\u00ae] and 0.0\u201310.0 for 3D TBP Vectra[\u00ae] WB360. FotoFinder\u2019s ATBM[\u00ae] master creates a 2D-image of the entire surface of the patient using 20 photos taken from 8 different parts of the body, while Vectra[\u00ae] WB360 generates a 3D-image using 92 individual photos that are converted into a 3D Avatar. Finally, the dermatologists assessed the skin lesions a second time with knowledge of the AI risk assessment scores of 2D FotoFinder ATBM[\u00ae] and 3D TBP Vectra[\u00ae] WB360. The patient was informed about the lesions with suspicion for melanoma by a dermatologist and about the 2D and 3D CNN classifications above the study-defined cut-off scores. The indication for excision of suspicious lesions was based on the dermatologists\u2019 suspicion for melanoma and/or malignancy risk assessment scores of the 2D imaging tool FotoFinder ATBM[\u00ae] and/or 3D imaging tool Vectra[\u00ae] WB360 indicating values over the study-defined cut-off score (>0.5 resp. >5.0). The patient was informed which lesions would be removed with the appropriate rationale. A maximum of two excisions were performed per visit and per patient due to ethical reasons, with an exception for further biopsies in cases of high-grade suspicion of malignancy by the dermatologist or patient request. The reference standard for a benign lesion was either the histological diagnosis (no melanoma) or the combination of the dermatologist\u2019s evaluation (benign lesion) plus the AI scores of two independent medical devices (FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360) each below the cut-off score. After all the assessments, the participating patients completed an anonymous 9-item questionnaire that surveyed their attitudes, personal preferences, and concerns about the use of AI in melanoma detection compared with skin cancer screenings by a dermatologist in addition to sociodemographic data. The dermatologists completed a 2-item questionnaire assessing their attitudes toward the smartphone app in melanoma screening, which was answered specifically for each of the 114 included patients. Neither the patients nor dermatologists were informed about the result of the app\u2019s risk assessment due to the extremely high number of false-positive findings to prevent response bias. Our survey was designed de novo after literature research. All answers were optional. The questionnaire was available in German, English and French and included binary questions (yes/no), multiple choice questions and visual analogue scales (VAS) with scores from 0\u201310.\n\n_2.3. Statistical Analysis_\n\nComparison of continuous variables was tested with the Wilcoxon rank test. Comparison of categorical variables was tested with Fishers Exact test. A p-value of less than 0.05 was deemed as significant. Receiver Operating Characteristics (ROC) analysis was used to assess the performance of the SkinVision[\u00ae] app against the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologist and the SkinVision[\u00ae] app against histology. All analyses were performed with R (version 4.1) and visualized with ggplot2.\n\n_2.4. Ethics_\n\nThe study was approved by the local ethics committee (22020-02482), registered with ClinicalTrials.gov (NCT04605822), and was conducted in compliance with the Declaration of Helsinki and Good Clinical Practice GCP-rules.",
        "results": "**3. Results**\n\n_3.1. Study Population_\n\nOverall, data from 114 patients were analyzed, including 55 patients at high-risk for developing a melanoma (mean age 55 years (age range): 22\u201385), 47% females) and 59 melanoma patients (mean age 60 years (age range: 29\u201381), 54% females) (Table 1). The family history for melanoma was positive in 19% of the melanoma patients and in 56% of the non-melanoma patients. Most of the patients used sunscreen SPF 30\u201350 (51% of melanoma patients vs. 62% of patients at high-risk for melanoma) and had previous sunburns in childhood (54% resp. 69%).\n\n**Table 1. Characteristics of the study population and their skin cancer awareness.**\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Age, n (age range)** 59 (22\u201385) 60 (29\u201381) 55 (22\u201385)\n**Sex, n (%)**\nFemale 58 (51%) 32 (54%) 26 (47%)\nMale 56 (49%) 27 (46%) 29 (53%)\n**Risk profile, n (%)**\nMultiple melanocytic nevi (\u2265100)\nand/or dysplastic nevi (\u22655) and/or positive\nfamily history for melanoma and/or 55 (48%) 0 (0%) 55 (100%)\ndiagnosis of dysplastic nevus syndrome\nand/or CDKN2A mutation\n\nPrevious resected melanoma in situ or\n57 (50%) 57 (97%) 0 (0%)\nprimary cutaneous melanoma\n\nMetastatic melanoma 2 (1.8%) 2 (3.4%) 0 (0%)\n**Positive family history for melanoma,**\n42 (37%) 11 (19%) 31 (56%)\n**n (%)**\n\n**Frequency of skin cancer screening,**\n**n (%)**\nSeveral times per year 40 (35%) 34 (58%) 6 (11%)\nEvery 12 months 39 (34%) 16 (27%) 23 (42%)\nEvery 1\u20132 years 8 (7%) 4 (6.8%) 4 (7.3%)\nEvery 2 years 9 (7.9%) 2 (3.4%) 7 (13%)\nLess than every 2 years 14 (12%) 3 (5.1%) 11 (20%)\nNever 4 (3.5%) 0 (0%) 4 (7.3%)\n**History of sunburns in childhood,**\n70 (61%) 32 (54%) 38 (69%)\n**n (%)**\n\n**Frequency of sunburns (Child), n (%)**\nRarely (less than once per year) 44 (63%) 20 (62%) 24 (63%)\nRegularly (once per year) 22 (31%) 10 (31%) 12 (32%)\nOften (more than once per year) 4 (5.7%) 2 (6.2%) 2 (5.3%)\n**History of sunburns in adulthood,**\n39 (34%) 18 (31%) 21 (38%)\n**n (%)**\n\n\n-----\n\n_Cancers 2022, 14, 3829_ 6 of 21\n\n**Table 1. Cont.**\n\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Frequency of sunburns (Adult), n (%)**\nRarely (less than once per year) 38 (97%) 18 (100%) 20 (95%)\nRegularly (once per year) 0 (0%) 0 (0%) 0 (0%)\nOften (more than once per year) 1 (2.6%) 0 (0%) 1 (4.8%)\n**Previous tanning in the solarium,**\n38 (33%) 13 (22%) 25 (45%)\n**n (%)**\n\n**Usage of sunscreen (SPF), n (%)**\nSPF 6\u201310 2 (1.8%) 1 (1.7%) 1 (1.8%)\nSPF 15\u201325 10 (8.8%) 3 (5.1%) 7 (13%)\nSPF 30\u201350 64 (56%) 30 (51%) 34 (62%)\nSPF 50+ 38 (33%) 25 (42%) 13 (24%)\n\n1 Median (Range); n (%).\n\nA total of 1204 pigmented skin lesions were assessed in this study (Figure 1). In 61 cases (5.1%), we performed a histopathology examination, while 1129 lesions (94%) were diagnosed to be clinically clearly benign based on the combination of a risk assessment by a dermatologist and AI-based risk scores below the cut-off (2D and 3D TBP) and had no indication for obtaining histopathology.\n\n**Figure 1. Flow chart of all included pigmented skin lesions and their histopathological outcome.**\nAI = artificial intelligence; TBP = total body photography.\n\n_3.2. Diagnostic Accuracy and Performance of the Smartphone App SkinVision[\u00ae]_\n\n3.2.1. Comparison of all Risk Assessments\n\nThe smartphone app SkinVision[\u00ae] classified 980 (81%) lesions as benign and indicated an increased risk for melanoma in 224 (19%) cases, while the dermatologists diagnosed 1195 (99.3%) lesions as benign and only nine (0.7%) as suspicious. Consequently, the CE-certified app had a 27-fold higher rate of melanoma-suspicious lesions compared to dermatologists. The AI scores of the 2D and 3D CNN devices classified most lesions as benign and 47 (3.9%) lesions (FotoFinder ATBM[\u00ae]) resp. 39 (3.2%) lesions (Vectra[\u00ae] WB360) as suspicious for melanoma (Figure 2).\n\n-----\n\n_Cancers 2022, 14, 3829_ 7 of 21\n\n**Figure 2. Comparison of risk assessments with the highest rate of suspected melanoma cases by the**\nsmartphone app (n = 1204).\n\nAmong the 224 lesions classified as suspicious by SkinVision[\u00ae], 193 were considered clinically benign by the physician and the 2D and 3D CNN AI-risk assessment scores, whereas four pigmented skin lesions were classified suspicious by the smartphone app, dermatologists, 2D, and 3D TBP devices (Table 2, Supplementary Table S1). The knowledge of AI-based risk assessment scores did not meaningfully affect dermatologists\u2019 classification of skin lesions, as they changed their decision towards the indication for excision in only three lesions that later turned out to be benign (Table 2).\n\n**Table 2. Risk assessments of 1204 pigmented skin lesions by the smartphone app SkinVision[\u00ae], 2D**\nimaging FotoFinder ATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in\ncombination with knowledge of FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360 AI-scores.\n\n**Characteristic** **N = 1204** **[1]**\n\n**Smartphone app SkinVision[\u00ae]**\n\nbenign 980 (81%)\nsuspicious 224 (19%)\n**2D Imaging FotoFinder ATBM[\u00ae]**\n\nbenign 1157 (96%)\nsuspicious 47 (3.9%)\n**3D Imaging VECTRA[\u00ae]** **WB360**\nbenign 1165 (97%)\nsuspicious 39 (3.2%)\n**Dermatologists**\nbenign 1195 (99%)\nsuspicious 9 (0.7%)\n**Dermatologists informed about risk assessment scores by**\n**FotoFinder ATBM[\u00ae]** **+ VECTRA[\u00ae]** **WB360**\nbenign 1192 (99%)\nsuspicious 12 (1.0%)\n\n1 n (%); AI = artificial intelligence.\n\n-----\n\n_Cancers 2022, 14, 3829_ 8 of 21\n\n3.2.2. Diagnostic Accuracy of the Smartphone App Based on the Combination of the\nDermatologist\u2019s Evaluation plus the AI Risk-Assessment Scores of Two Independent\nMedical Devices\n\nReceiver operating characteristic (ROC) analysis of the classification of benign and suspicious lesions of the SkinVision[\u00ae] app compared to the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologists had an area under the curve (AUC) score of 0.621, sensitivity of 0.41, and specificity of 0.83 (Figure 3). Although the specificity is reasonable, the sensitivity is low, thus suggesting that the SkinVision[\u00ae] app has poor diagnostic accuracy.\n\n**Figure 3. Receiver operating characteristic curve of the smartphone app in relation to the results**\nof the combination of risk assessments by dermatologists, FotoFinder ATBM[\u00ae], and Vectra[\u00ae] WB360\n(sensitivity: 41.3%, specificity: 82.9%); AUC = area under the curve.\n\n3.2.3. Diagnostic Accuracy of the Smartphone App Based on Histopathology\n\nAmong 61 pigmented skin lesions examined histologically, we detected six melanomas, 19 melanocytic nevi, 20 dysplastic nevi, as well as 16 otherwise classified lesions were diagnosed (Table 3, Figure 1).\n\n**Table 3. Diagnostic accuracy of the AI-based smartphone app SkinVision[\u00ae], 2D imaging FotoFinder**\nATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in combination with AI in\nmelanoma detection based on histopathology: sensitivity and specificity.\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**Smartphone app SkinVision[\u00ae]** 61\nbenign 13 (68%) 10 (50%) 1 (17%) 10 (62%)\nsuspicious 6 (32%) 10 (50%) 5 (83%) 6 (38%)\n**2D imaging FotoFinder ATBM[\u00ae]** 61\nbenign 7 (37%) 11 (55%) 1 (17%) 4 (25%)\nsuspicious 12 (63%) 9 (45%) 5 (83%) 12 (75%)\n\n-----\n\n_Cancers 2022, 14, 3829_ 9 of 21\n\n**Table 3. Cont.**\n\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**3D imaging VECTRA[\u00ae]** **WB360** 61\nbenign 18 (95%) 9 (45%) 1 (17%) 8 (50%)\nsuspicious 1 (5.3%) 11 (55%) 5 (83%) 8 (50%)\n**Dermatologists** 61\nbenign 17 (89%) 18 (90%) 1 (17%) 16 (100%)\nsuspicious 2 (11%) 2 (10%) 5 (83%) 0 (0%)\n_Beginner: <2 years\u2019 work_\n44 N = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 14 (93%) 10 (83%) 1 (20%) 13 (100%)\nsuspicious 1 (6.7%) 2 (17%) 4 (80%) 0 (0%)\n_Intermediate: 2\u20135 years\u2019 work_\n5 N = 2 N = 3 N = 0 N = 0\n_experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\n11 N = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n**Dermatologists informed about**\n61\n**AI scores** **[2]**\n\nbenign 16 (84%) 17 (85%) 1 (17%) 15 (94%)\nsuspicious 3 (16%) 3 (15%) 5 (83%) 1 (6.2%)\n_Beginner: <2 years\u2019 work_\nN = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 13 (87%) 9 (75%) 1 (20%) 12 (92%)\nsuspicious 2 (13%) 3 (25%) 4 (80%) 1 (7.7%)\n_Intermediate: 2\u20135 years\u2019_\nN = 2 N = 3 N = 0 N = 0\n_work experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\nN = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n\n1 n (%); * Other = pigmented basal cell carcinoma, histiocytoma, lentigo solaris, pigmented seborrhoic keratosis,\nfolliculitis with perifolliculitis, collisional tumor: seborrhoic keratosis and nevus, collisional tumor: actinic\nkeratosis and lentigo solaris; [2] risk assessment scores by FotoFinder ATBM[\u00ae] and VECTRA[\u00ae] WB360; AI = artificial\nintelligence.\n\nBased on histopathology, Figure 4 represents the risk assessments of the dermatologists, the mHealth app, and the combination of AI and dermatologists. Both the SkinVision[\u00ae]\n\napp and the 2D and 3D CNN devices indicated for 5 of 6 histological verified melanomas\n(83% sensitivity) an elevated score indicating suspicion of melanoma. The false-negative rate for all AI-based medical devices was 17%. The three different false-negative classified melanomas were all superficial spreading stage IA melanomas\u2014SkinVision[\u00ae]: superficial spreading melanoma, 0.9 mm Breslow thickness, AJCC stage IA (pT1bN0M0); 3D TBP Vectra[\u00ae] WB360: superficial spreading melanoma, 0.3 mm Breslow thickness, AJCC stage IA (pT1aN0M0); 2D TBP FotoFinder ATBM[\u00ae]: superficial spreading melanoma, 0.7 mm Breslow thickness, AJCC stage IA (pT1aN0M0). The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses was 60.0% of the smartphone app, 63.6% of the 3D Vectra[\u00ae] WB360 imaging device, and 40.0% of the 2D FotoFinder ATBM[\u00ae] risk assessment tool. Dermatologists correctly identified five of six melanomas (83% sensitivity). The false-negative rate was 17%. The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses among all dermatologists was 92.7%. The performance of the dermatologists strongly correlated with their professional experience. Hence, the true-positive rates for melanoma were greater for experts and dermatologists with intermediate experience (100% sensitivity) compared to beginners (80% sensitivity); whereas the true-negative rates were similar (93.3% resp. 92.5% specificity) (Table 3).\n\n**Figure 4. Histology and corresponding diagnosis of lesions assessed by dermatologists, smartphone**\napp SkinVision[\u00ae], and dermatologists and AI (n = 61).\n\nFigure 5 reveals several pigmented skin lesions that were correctly and incorrectly classified as melanomas by the smartphone app SkinVision[\u00ae].\n\n**Figure 5. Correctly and falsely classified melanomas by the smartphone app SkinVision[\u00ae]: (A,B). True-**\npositive classified melanoma; (C,D). False-positive classified melanoma; (E,F). True-negative classified\nmelanoma; (G). False-negative classified melanoma.\n\nROC analysis of the classification of melanoma of the SkinVision[\u00ae] app compared\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision[\u00ae] app had\nhigh sensitivity but mediocre specificity.\n\nROC analysis of the classification of melanoma of the SkinVision\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision\nhigh sensitivity but mediocre specificity.\n\n-----\n\n_Cancers 2022, 14, 3829_ 11 of 21\n\n**Figure 6. Receiver operating characteristic curve of the smartphone app in relation to the results of**\nthe histology (sensitivity of 83.3%, specificity 60.0%); AUC = area under the curve.\n\n_3.3. Patient Perspective on AI in Melanoma Screening_\n3.3.1. Confidence in Dermatologists vs. Smartphone App\n\nMost of the patients at high-risk for melanoma (55% (30/55)) and patients with\nmelanoma (53% (31/59)) reported being very confident about a mole examination by\na dermatologist (rating scale 10/10). In contrast, only a minority of high-risk patients (16%\n(9/55)) and melanoma patients (12% (7/59)) felt very safe when being investigated by a\nsmartphone app alone. No significant difference in ratings was identified between the two\nrisk groups (p < 0.9; p < 0.7) (Supplementary Table S2).\n\n3.3.2. Trustworthiness of the Smartphone App\n\nAll participants primarily rated physician examination (100% (55/55) among highrisk patients, 100% (59/59) among melanoma patients) as trustworthy, and the majority\ndid so for 2D TBP imaging (93% (51/55) resp. 88% (52/59)) and 3D TBP imaging (91%\n(50/55) resp. 90% (53/59)). The smartphone app was less frequently rated as trustworthy,\nwith only 36% (20/55) among patients at high-risk for melanoma and 49% (29/59) among\nmelanoma patients (Table 4). The age revealed a significant correlation with the evaluation\nof smartphones\u2019 trustworthiness (p < 0.004), with older patients (>60 years old) having\ntrusted the app three times more than younger patients (\u226460 years old) (Figure 7). Neither\nprevious melanoma vs. high-risk criteria for melanoma nor sex significantly influenced the\nevaluation of the trustworthiness of the smartphone app.\n\n-----\n\n_Cancers 2022, 14, 3829_ 12 of 21\n\n**Table 4. Assessment of trustworthiness of the AI-based smartphone app SkinVision[\u00ae], 2D imaging**\nFotoFinder ATBM[\u00ae], and 3D imaging Vectra[\u00ae] WB360 compared to dermatologists.\n\n\n**Patients at**\n**High-Risk**\n**for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Characteristic** **N**\n\n\n**Patients**\n**with**\n**Melanoma,**\n**N = 59** **[1]**\n\n\n**_p-Value_** **[2]**\n\n\n**The following examination was**\n**trustworthy: Smartphone** 114 0.3\n**app assessment**\n\nYes 29 (49%) 20 (36%)\nNo 5 (8.5%) 8 (15%)\nI don\u2019t know 23 (39%) 22 (40%)\nNo answer 2 (3.4%) 5 (9.1%)\n**Dermatologist assessment** 114\nYes 59 (100%) 55 (100%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 0 (0%) 0 (0%)\nNo answer 0 (0%) 0 (0%)\n**2D TBP assessment** 114 0.3\nYes 52 (88%) 51 (93%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 7 (12%) 3 (5.5%)\nNo answer 0 (0%) 1 (1.8%)\n**3D TBP assessment** 114 0.3\nYes 53 (90%) 50 (91%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 6 (10%) 3 (5.5%)\nNo answer 0 (0%) 2 (3.6%)\n\n1 n (%); 2 Fisher\u2019s exact test; Pearson\u2019s Chi-squared test; TBP = total body photography.\n\n**Figure 7. Odds ratio for variables influencing the trustworthiness of smartphones\u2019 risk assessment in**\nmelanoma detection.\n\n-----\n\n_Cancers 2022, 14, 3829_ 13 of 21\n\n3.3.3. Impact of AI vs. Dermatologists\u2019 Examination on Patients\u2019 Fear of Developing\nSkin Cancer\n\nMost participants indicated that an examination by a physician reduced their fear of\ndeveloping skin cancer, namely in 89% (49/55) among high-risk patients and 81% (48/59)\namong melanoma patients. Comparably, the 2D TBP imaging achieved the same effect in\n78% (43/55) resp. 76% (45/59), and the 3D TBP device in 82% (45/55) resp. 75% (44/59). In\ncontrast, the assessment with the smartphone app appeased fear of skin cancer in only 33%\n(18/55) of patients at high-risk for melanoma and in 32% (19/59) of patients with melanoma.\nIndeed, one high-risk patient (1.8%) even reported an increased fear of developing skin\ncancer by using the smartphone app (Supplementary Table S2).\n\n3.3.4. Patients\u2019 Subjective Assessment of the Accuracy of AI vs. Dermatologists\n\nPatients expected reliable results with the highest accuracy by both the assessment\nby a physician (98% (54/55) among high-risk patients, 92% (54/59) among melanoma\npatients) and by the 2D TBP imaging (82% (45/55) resp. 86% (51/59)) as well as the 3D\nTBP device (89% (49/55) resp. 88% (52/59)). Only 16% (9/55) of high-risk patients and\n31% (18/59) of melanoma patients expected reliable results from the smartphone app\n(Supplementary Table S2). No significant differences were identified between the two risk\ngroups in the evaluations.\n\n3.3.5. Patient Preference for Skin Cancer Screening\n\nBoth cohorts favored a combination of dermatologist and 3D TBP risk assessment for\nthe examination of pigmented skin lesions (64% (35/55) among patients at high-risk for\nmelanoma, 51% (30/59) among melanoma patients), while neither preferred assessment\nby a smartphone app alone (Figure 8A, Supplementary Table S2). The combination of\ndermatologist and smartphone app was favored by only 1.8% (1/55) of patients at high-risk\nfor melanoma and 3.4% (2/59) of patients with melanoma.\n\n**Figure 8. Cont.**\n\n-----\n\n_Cancers 2022, 14, 3829_ 14 of 21\n\n**Figure 8. (A) Patient preference for mole assessment (patients at high-risk for melanoma, n = 55;**\npatients with melanoma, n = 59); (B) Patient preference for AI in skin cancer screening (patients at\nhigh-risk for melanoma, n = 55; patients with melanoma, n = 59).\n\nRegarding patient preference for skin cancer screenings, almost all high-risk patients\n(98% (54/55)) and melanoma patients (95% (56/59)) indicated their belief that AI can improve\na physician\u2019s diagnostic performance. Most patients (64% (35/55) among high-risk patients,\n54% (32/59) among melanoma patients) would prefer that the physicians would always\nconsider the result of AI in their diagnosis (Figure 8B) (Supplementary Table S2).\n\n3.3.6. Dermatologists\u2019 Perspective of Smartphone Apps for Melanoma Screening\n\nAmong the seven dermatologists, they stated in only 5.3% of the skin cancer screenings\n(6/114) that the smartphone app increased diagnostic confidence and in only 8.8% of the\nassessments (10/114) they trusted the app (Supplementary Table S3).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Unclear"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Over-Detection of Melanoma-Suspect Lesions by a CE-Certified Smartphone App: Performance in Comparison to Dermatologists, 2D and 3D Convolutional Neural Networks in a Prospective Data Set of 1204 Pigmented Skin Lesions Involving Patients' Perception.",
        "abstract": "The exponential increase in algorithm-based mobile health (mHealth) applications (apps) for melanoma screening is a reaction to a growing market. However, the performance of available apps remains to be investigated. In this prospective study, we investigated the diagnostic accuracy of a class 1 CE-certified smartphone app in melanoma risk stratification and its patient and dermatologist satisfaction. Pigmented skin lesions \u2265 3 mm and any suspicious smaller lesions were assessed by the smartphone app SkinVision\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Study Design and Participants_\n\nWe performed this prospective, single-center, comparative observational cohort study at the Department of Dermatology at the University Hospital Basel in Switzerland between January 2021 to June 2021. Seven dermatologists (beginners: <2 years professional experience, n = 4; intermediate: 2\u20135 years professional experience, n = 1; experts: >5 years professional experience, n = 2), as well as 114 patients were included in the study with an age of \u226518 years with a high risk for developing a primary cutaneous malignant melanoma defined as previous melanoma (including melanoma in situ), \u2265100 melanocytic nevi, \u22655 atypical nevi, strong family history for melanoma, diagnosis of dysplastic nevus syndrome, or known CDKN2A mutation. Exclusion criteria were acute psychiatric illness or acute crisis and Fitzpatrick sign type V-VI.\n\n_2.2. Procedures_\n\nIn the first step, the patients signed the informed consent after they were informed by a dermatologist about the study design and the following examinations. Subsequently, the dermatologist obtained a medical history, including a history of melanoma to enable classification into the different risk groups. The classification of all high-risk patients into further risk cohorts are defined by how often the patients require skin cancer screening. Based on this, we have classified our melanoma patients as high risk, as they come to dermatological control every 3\u201312 months according to the corresponding AJCC stage. Patients without melanoma in their own medical history are seen every 12 months. All study participants underwent a standard skin cancer screening by a dermatologist (Supplementary Figure S1). The pigmented skin lesions were assessed with dermatoscopy and classified as either benign or malignant. The dermatologist defined pigmented skin lesions that were suspected melanoma and indicated whether lesions smaller than 3 mm should be classified with the different AI modalities due to suspicion. The regular skin cancer screening was performed before any AI-based risk assessments were achieved, so the dermatologists had no knowledge about AI classification of the melanocytic lesions up to this point. In the next step, all melanocytic nevi \u2265 3 mm or any smaller suspicious melanocytic lesions were assessed using an iOS-based iPhone SE smartphone equipped with a 12-megapixel camera and the app SkinVision[\u00ae] (SkinVision[\u00ae] B.V., Amsterdam, the Netherlands, App-Version 6.8.1), which is based on a machine-learning algorithm classifying pigmented lesions into low, medium, and high risk for melanoma. Only the lesions classified as high-risk\u2014indicating a recommendation to consult a dermatologist\u2014were recorded. Afterwards, patients received a whole-body screening using the 2D automated total body mapping (ATBM[\u00ae]) master of FotoFinder (FotoFinder ATBM[\u00ae] Systems GmbH, Bad Birnbach, Germany, Version 3.3.1.0, risk scoring with MoleAnalyzer Pro), and the 3D TBP Vectra[\u00ae] WB360 System (Canfield Scientific, Parsippany, NJ, USA, Version 4.7.1, risk scoring with DEXI) which was performed by medical students and a study nurse. Subsequently, all lesions over 3 mm or any smaller suspicious lesions were assessed by the 2D and 3D AI devices using their dermoscopes (dermoscope medicam 1000 (FotoFinder ATBM[\u00ae]) and VISIOMED[\u00ae] D200evo dermoscope (Vectra[\u00ae] WB360)). For dermoscopic imaged melanocytic nevi, we obtained AI-based risk scores between 0.0\u20131.0 for 2D FotoFinder ATBM[\u00ae] and 0.0\u201310.0 for 3D TBP Vectra[\u00ae] WB360. FotoFinder\u2019s ATBM[\u00ae] master creates a 2D-image of the entire surface of the patient using 20 photos taken from 8 different parts of the body, while Vectra[\u00ae] WB360 generates a 3D-image using 92 individual photos that are converted into a 3D Avatar. Finally, the dermatologists assessed the skin lesions a second time with knowledge of the AI risk assessment scores of 2D FotoFinder ATBM[\u00ae] and 3D TBP Vectra[\u00ae] WB360. The patient was informed about the lesions with suspicion for melanoma by a dermatologist and about the 2D and 3D CNN classifications above the study-defined cut-off scores. The indication for excision of suspicious lesions was based on the dermatologists\u2019 suspicion for melanoma and/or malignancy risk assessment scores of the 2D imaging tool FotoFinder ATBM[\u00ae] and/or 3D imaging tool Vectra[\u00ae] WB360 indicating values over the study-defined cut-off score (>0.5 resp. >5.0). The patient was informed which lesions would be removed with the appropriate rationale. A maximum of two excisions were performed per visit and per patient due to ethical reasons, with an exception for further biopsies in cases of high-grade suspicion of malignancy by the dermatologist or patient request. The reference standard for a benign lesion was either the histological diagnosis (no melanoma) or the combination of the dermatologist\u2019s evaluation (benign lesion) plus the AI scores of two independent medical devices (FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360) each below the cut-off score. After all the assessments, the participating patients completed an anonymous 9-item questionnaire that surveyed their attitudes, personal preferences, and concerns about the use of AI in melanoma detection compared with skin cancer screenings by a dermatologist in addition to sociodemographic data. The dermatologists completed a 2-item questionnaire assessing their attitudes toward the smartphone app in melanoma screening, which was answered specifically for each of the 114 included patients. Neither the patients nor dermatologists were informed about the result of the app\u2019s risk assessment due to the extremely high number of false-positive findings to prevent response bias. Our survey was designed de novo after literature research. All answers were optional. The questionnaire was available in German, English and French and included binary questions (yes/no), multiple choice questions and visual analogue scales (VAS) with scores from 0\u201310.\n\n_2.3. Statistical Analysis_\n\nComparison of continuous variables was tested with the Wilcoxon rank test. Comparison of categorical variables was tested with Fishers Exact test. A p-value of less than 0.05 was deemed as significant. Receiver Operating Characteristics (ROC) analysis was used to assess the performance of the SkinVision[\u00ae] app against the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologist and the SkinVision[\u00ae] app against histology. All analyses were performed with R (version 4.1) and visualized with ggplot2.\n\n_2.4. Ethics_\n\nThe study was approved by the local ethics committee (22020-02482), registered with ClinicalTrials.gov (NCT04605822), and was conducted in compliance with the Declaration of Helsinki and Good Clinical Practice GCP-rules.",
        "results": "**3. Results**\n\n_3.1. Study Population_\n\nOverall, data from 114 patients were analyzed, including 55 patients at high-risk for developing a melanoma (mean age 55 years (age range): 22\u201385), 47% females) and 59 melanoma patients (mean age 60 years (age range: 29\u201381), 54% females) (Table 1). The family history for melanoma was positive in 19% of the melanoma patients and in 56% of the non-melanoma patients. Most of the patients used sunscreen SPF 30\u201350 (51% of melanoma patients vs. 62% of patients at high-risk for melanoma) and had previous sunburns in childhood (54% resp. 69%).\n\n**Table 1. Characteristics of the study population and their skin cancer awareness.**\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Age, n (age range)** 59 (22\u201385) 60 (29\u201381) 55 (22\u201385)\n**Sex, n (%)**\nFemale 58 (51%) 32 (54%) 26 (47%)\nMale 56 (49%) 27 (46%) 29 (53%)\n**Risk profile, n (%)**\nMultiple melanocytic nevi (\u2265100)\nand/or dysplastic nevi (\u22655) and/or positive\nfamily history for melanoma and/or 55 (48%) 0 (0%) 55 (100%)\ndiagnosis of dysplastic nevus syndrome\nand/or CDKN2A mutation\n\nPrevious resected melanoma in situ or\n57 (50%) 57 (97%) 0 (0%)\nprimary cutaneous melanoma\n\nMetastatic melanoma 2 (1.8%) 2 (3.4%) 0 (0%)\n**Positive family history for melanoma,**\n42 (37%) 11 (19%) 31 (56%)\n**n (%)**\n\n**Frequency of skin cancer screening,**\n**n (%)**\nSeveral times per year 40 (35%) 34 (58%) 6 (11%)\nEvery 12 months 39 (34%) 16 (27%) 23 (42%)\nEvery 1\u20132 years 8 (7%) 4 (6.8%) 4 (7.3%)\nEvery 2 years 9 (7.9%) 2 (3.4%) 7 (13%)\nLess than every 2 years 14 (12%) 3 (5.1%) 11 (20%)\nNever 4 (3.5%) 0 (0%) 4 (7.3%)\n**History of sunburns in childhood,**\n70 (61%) 32 (54%) 38 (69%)\n**n (%)**\n\n**Frequency of sunburns (Child), n (%)**\nRarely (less than once per year) 44 (63%) 20 (62%) 24 (63%)\nRegularly (once per year) 22 (31%) 10 (31%) 12 (32%)\nOften (more than once per year) 4 (5.7%) 2 (6.2%) 2 (5.3%)\n**History of sunburns in adulthood,**\n39 (34%) 18 (31%) 21 (38%)\n**n (%)**\n\n\n-----\n\n_Cancers 2022, 14, 3829_ 6 of 21\n\n**Table 1. Cont.**\n\n\n**Patients with**\n**All Patients,**\n**Characteristic** **Melanoma,**\n**N = 114** **[1]**\n**N = 59** **[1]**\n\n\n**Patients at**\n**High-Risk for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Frequency of sunburns (Adult), n (%)**\nRarely (less than once per year) 38 (97%) 18 (100%) 20 (95%)\nRegularly (once per year) 0 (0%) 0 (0%) 0 (0%)\nOften (more than once per year) 1 (2.6%) 0 (0%) 1 (4.8%)\n**Previous tanning in the solarium,**\n38 (33%) 13 (22%) 25 (45%)\n**n (%)**\n\n**Usage of sunscreen (SPF), n (%)**\nSPF 6\u201310 2 (1.8%) 1 (1.7%) 1 (1.8%)\nSPF 15\u201325 10 (8.8%) 3 (5.1%) 7 (13%)\nSPF 30\u201350 64 (56%) 30 (51%) 34 (62%)\nSPF 50+ 38 (33%) 25 (42%) 13 (24%)\n\n1 Median (Range); n (%).\n\nA total of 1204 pigmented skin lesions were assessed in this study (Figure 1). In 61 cases (5.1%), we performed a histopathology examination, while 1129 lesions (94%) were diagnosed to be clinically clearly benign based on the combination of a risk assessment by a dermatologist and AI-based risk scores below the cut-off (2D and 3D TBP) and had no indication for obtaining histopathology.\n\n**Figure 1. Flow chart of all included pigmented skin lesions and their histopathological outcome.**\nAI = artificial intelligence; TBP = total body photography.\n\n_3.2. Diagnostic Accuracy and Performance of the Smartphone App SkinVision[\u00ae]_\n\n3.2.1. Comparison of all Risk Assessments\n\nThe smartphone app SkinVision[\u00ae] classified 980 (81%) lesions as benign and indicated an increased risk for melanoma in 224 (19%) cases, while the dermatologists diagnosed 1195 (99.3%) lesions as benign and only nine (0.7%) as suspicious. Consequently, the CE-certified app had a 27-fold higher rate of melanoma-suspicious lesions compared to dermatologists. The AI scores of the 2D and 3D CNN devices classified most lesions as benign and 47 (3.9%) lesions (FotoFinder ATBM[\u00ae]) resp. 39 (3.2%) lesions (Vectra[\u00ae] WB360) as suspicious for melanoma (Figure 2).\n\n-----\n\n_Cancers 2022, 14, 3829_ 7 of 21\n\n**Figure 2. Comparison of risk assessments with the highest rate of suspected melanoma cases by the**\nsmartphone app (n = 1204).\n\nAmong the 224 lesions classified as suspicious by SkinVision[\u00ae], 193 were considered clinically benign by the physician and the 2D and 3D CNN AI-risk assessment scores, whereas four pigmented skin lesions were classified suspicious by the smartphone app, dermatologists, 2D, and 3D TBP devices (Table 2, Supplementary Table S1). The knowledge of AI-based risk assessment scores did not meaningfully affect dermatologists\u2019 classification of skin lesions, as they changed their decision towards the indication for excision in only three lesions that later turned out to be benign (Table 2).\n\n**Table 2. Risk assessments of 1204 pigmented skin lesions by the smartphone app SkinVision[\u00ae], 2D**\nimaging FotoFinder ATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in\ncombination with knowledge of FotoFinder ATBM[\u00ae] and Vectra[\u00ae] WB360 AI-scores.\n\n**Characteristic** **N = 1204** **[1]**\n\n**Smartphone app SkinVision[\u00ae]**\n\nbenign 980 (81%)\nsuspicious 224 (19%)\n**2D Imaging FotoFinder ATBM[\u00ae]**\n\nbenign 1157 (96%)\nsuspicious 47 (3.9%)\n**3D Imaging VECTRA[\u00ae]** **WB360**\nbenign 1165 (97%)\nsuspicious 39 (3.2%)\n**Dermatologists**\nbenign 1195 (99%)\nsuspicious 9 (0.7%)\n**Dermatologists informed about risk assessment scores by**\n**FotoFinder ATBM[\u00ae]** **+ VECTRA[\u00ae]** **WB360**\nbenign 1192 (99%)\nsuspicious 12 (1.0%)\n\n1 n (%); AI = artificial intelligence.\n\n-----\n\n_Cancers 2022, 14, 3829_ 8 of 21\n\n3.2.2. Diagnostic Accuracy of the Smartphone App Based on the Combination of the\nDermatologist\u2019s Evaluation plus the AI Risk-Assessment Scores of Two Independent\nMedical Devices\n\nReceiver operating characteristic (ROC) analysis of the classification of benign and suspicious lesions of the SkinVision[\u00ae] app compared to the combined evaluation of FotoFinder ATBM[\u00ae], Vectra[\u00ae] WB360, and the dermatologists had an area under the curve (AUC) score of 0.621, sensitivity of 0.41, and specificity of 0.83 (Figure 3). Although the specificity is reasonable, the sensitivity is low, thus suggesting that the SkinVision[\u00ae] app has poor diagnostic accuracy.\n\n**Figure 3. Receiver operating characteristic curve of the smartphone app in relation to the results**\nof the combination of risk assessments by dermatologists, FotoFinder ATBM[\u00ae], and Vectra[\u00ae] WB360\n(sensitivity: 41.3%, specificity: 82.9%); AUC = area under the curve.\n\n3.2.3. Diagnostic Accuracy of the Smartphone App Based on Histopathology\n\nAmong 61 pigmented skin lesions examined histologically, we detected six melanomas, 19 melanocytic nevi, 20 dysplastic nevi, as well as 16 otherwise classified lesions were diagnosed (Table 3, Figure 1).\n\n**Table 3. Diagnostic accuracy of the AI-based smartphone app SkinVision[\u00ae], 2D imaging FotoFinder**\nATBM[\u00ae], 3D imaging Vectra[\u00ae] WB360, dermatologists, and dermatologists in combination with AI in\nmelanoma detection based on histopathology: sensitivity and specificity.\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**Smartphone app SkinVision[\u00ae]** 61\nbenign 13 (68%) 10 (50%) 1 (17%) 10 (62%)\nsuspicious 6 (32%) 10 (50%) 5 (83%) 6 (38%)\n**2D imaging FotoFinder ATBM[\u00ae]** 61\nbenign 7 (37%) 11 (55%) 1 (17%) 4 (25%)\nsuspicious 12 (63%) 9 (45%) 5 (83%) 12 (75%)\n\n-----\n\n_Cancers 2022, 14, 3829_ 9 of 21\n\n**Table 3. Cont.**\n\n\n**Melanocytic**\n**Histopathologic Diagnosis** **N** **Nevus,**\n**N = 19** **[1]**\n\n\n**Dysplastic**\n**Melanoma,** **Other *,**\n**Nevus,**\n**N = 6** **[1]** **N = 16** **[1]**\n**N = 20** **[1]**\n\n\n**3D imaging VECTRA[\u00ae]** **WB360** 61\nbenign 18 (95%) 9 (45%) 1 (17%) 8 (50%)\nsuspicious 1 (5.3%) 11 (55%) 5 (83%) 8 (50%)\n**Dermatologists** 61\nbenign 17 (89%) 18 (90%) 1 (17%) 16 (100%)\nsuspicious 2 (11%) 2 (10%) 5 (83%) 0 (0%)\n_Beginner: <2 years\u2019 work_\n44 N = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 14 (93%) 10 (83%) 1 (20%) 13 (100%)\nsuspicious 1 (6.7%) 2 (17%) 4 (80%) 0 (0%)\n_Intermediate: 2\u20135 years\u2019 work_\n5 N = 2 N = 3 N = 0 N = 0\n_experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\n11 N = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n**Dermatologists informed about**\n61\n**AI scores** **[2]**\n\nbenign 16 (84%) 17 (85%) 1 (17%) 15 (94%)\nsuspicious 3 (16%) 3 (15%) 5 (83%) 1 (6.2%)\n_Beginner: <2 years\u2019 work_\nN = 15 N = 12 N = 5 N = 13\n_experience_\n\nbenign 13 (87%) 9 (75%) 1 (20%) 12 (92%)\nsuspicious 2 (13%) 3 (25%) 4 (80%) 1 (7.7%)\n_Intermediate: 2\u20135 years\u2019_\nN = 2 N = 3 N = 0 N = 0\n_work experience_\n\nbenign 1 (50%) 3 (100%) 0 (0%) 0 (0%)\nsuspicious 1 (50%) 0 (0%) 0 (0%) 0 (0%)\n_Experts: >5 years\u2019 work_\nN = 2 N = 5 N = 1 N = 3\n_experience_\n\nbenign 2 (100%) 5 (100%) 0 (0%) 3 (100%)\nsuspicious 0 (0%) 0 (0%) 1 (100%) 0 (0%)\n\n1 n (%); * Other = pigmented basal cell carcinoma, histiocytoma, lentigo solaris, pigmented seborrhoic keratosis,\nfolliculitis with perifolliculitis, collisional tumor: seborrhoic keratosis and nevus, collisional tumor: actinic\nkeratosis and lentigo solaris; [2] risk assessment scores by FotoFinder ATBM[\u00ae] and VECTRA[\u00ae] WB360; AI = artificial\nintelligence.\n\nBased on histopathology, Figure 4 represents the risk assessments of the dermatologists, the mHealth app, and the combination of AI and dermatologists. Both the SkinVision[\u00ae]\n\napp and the 2D and 3D CNN devices indicated for 5 of 6 histological verified melanomas\n(83% sensitivity) an elevated score indicating suspicion of melanoma. The false-negative rate for all AI-based medical devices was 17%. The three different false-negative classified melanomas were all superficial spreading stage IA melanomas\u2014SkinVision[\u00ae]: superficial spreading melanoma, 0.9 mm Breslow thickness, AJCC stage IA (pT1bN0M0); 3D TBP Vectra[\u00ae] WB360: superficial spreading melanoma, 0.3 mm Breslow thickness, AJCC stage IA (pT1aN0M0); 2D TBP FotoFinder ATBM[\u00ae]: superficial spreading melanoma, 0.7 mm Breslow thickness, AJCC stage IA (pT1aN0M0). The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses was 60.0% of the smartphone app, 63.6% of the 3D Vectra[\u00ae] WB360 imaging device, and 40.0% of the 2D FotoFinder ATBM[\u00ae] risk assessment tool. Dermatologists correctly identified five of six melanomas (83% sensitivity). The false-negative rate was 17%. The true-negative rate for melanomas (specificity) including melanocytic nevi, dysplastic nevi, and otherwise classified diagnoses among all dermatologists was 92.7%. The performance of the dermatologists strongly correlated with their professional experience. Hence, the true-positive rates for melanoma were greater for experts and dermatologists with intermediate experience (100% sensitivity) compared to beginners (80% sensitivity); whereas the true-negative rates were similar (93.3% resp. 92.5% specificity) (Table 3).\n\n**Figure 4. Histology and corresponding diagnosis of lesions assessed by dermatologists, smartphone**\napp SkinVision[\u00ae], and dermatologists and AI (n = 61).\n\nFigure 5 reveals several pigmented skin lesions that were correctly and incorrectly classified as melanomas by the smartphone app SkinVision[\u00ae].\n\n**Figure 5. Correctly and falsely classified melanomas by the smartphone app SkinVision[\u00ae]: (A,B). True-**\npositive classified melanoma; (C,D). False-positive classified melanoma; (E,F). True-negative classified\nmelanoma; (G). False-negative classified melanoma.\n\nROC analysis of the classification of melanoma of the SkinVision[\u00ae] app compared\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision[\u00ae] app had\nhigh sensitivity but mediocre specificity.\n\nROC analysis of the classification of melanoma of the SkinVision\nto histology had an area under the curve (AUC) score of 0.717, sensitivity of 0.83, and\nspecificity of 0.6 (Figure 6). Using histology as the gold standard, the SkinVision\nhigh sensitivity but mediocre specificity.\n\n-----\n\n_Cancers 2022, 14, 3829_ 11 of 21\n\n**Figure 6. Receiver operating characteristic curve of the smartphone app in relation to the results of**\nthe histology (sensitivity of 83.3%, specificity 60.0%); AUC = area under the curve.\n\n_3.3. Patient Perspective on AI in Melanoma Screening_\n3.3.1. Confidence in Dermatologists vs. Smartphone App\n\nMost of the patients at high-risk for melanoma (55% (30/55)) and patients with\nmelanoma (53% (31/59)) reported being very confident about a mole examination by\na dermatologist (rating scale 10/10). In contrast, only a minority of high-risk patients (16%\n(9/55)) and melanoma patients (12% (7/59)) felt very safe when being investigated by a\nsmartphone app alone. No significant difference in ratings was identified between the two\nrisk groups (p < 0.9; p < 0.7) (Supplementary Table S2).\n\n3.3.2. Trustworthiness of the Smartphone App\n\nAll participants primarily rated physician examination (100% (55/55) among highrisk patients, 100% (59/59) among melanoma patients) as trustworthy, and the majority\ndid so for 2D TBP imaging (93% (51/55) resp. 88% (52/59)) and 3D TBP imaging (91%\n(50/55) resp. 90% (53/59)). The smartphone app was less frequently rated as trustworthy,\nwith only 36% (20/55) among patients at high-risk for melanoma and 49% (29/59) among\nmelanoma patients (Table 4). The age revealed a significant correlation with the evaluation\nof smartphones\u2019 trustworthiness (p < 0.004), with older patients (>60 years old) having\ntrusted the app three times more than younger patients (\u226460 years old) (Figure 7). Neither\nprevious melanoma vs. high-risk criteria for melanoma nor sex significantly influenced the\nevaluation of the trustworthiness of the smartphone app.\n\n-----\n\n_Cancers 2022, 14, 3829_ 12 of 21\n\n**Table 4. Assessment of trustworthiness of the AI-based smartphone app SkinVision[\u00ae], 2D imaging**\nFotoFinder ATBM[\u00ae], and 3D imaging Vectra[\u00ae] WB360 compared to dermatologists.\n\n\n**Patients at**\n**High-Risk**\n**for**\n**Melanoma,**\n**N = 55** **[1]**\n\n\n**Characteristic** **N**\n\n\n**Patients**\n**with**\n**Melanoma,**\n**N = 59** **[1]**\n\n\n**_p-Value_** **[2]**\n\n\n**The following examination was**\n**trustworthy: Smartphone** 114 0.3\n**app assessment**\n\nYes 29 (49%) 20 (36%)\nNo 5 (8.5%) 8 (15%)\nI don\u2019t know 23 (39%) 22 (40%)\nNo answer 2 (3.4%) 5 (9.1%)\n**Dermatologist assessment** 114\nYes 59 (100%) 55 (100%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 0 (0%) 0 (0%)\nNo answer 0 (0%) 0 (0%)\n**2D TBP assessment** 114 0.3\nYes 52 (88%) 51 (93%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 7 (12%) 3 (5.5%)\nNo answer 0 (0%) 1 (1.8%)\n**3D TBP assessment** 114 0.3\nYes 53 (90%) 50 (91%)\nNo 0 (0%) 0 (0%)\nI don\u2019t know 6 (10%) 3 (5.5%)\nNo answer 0 (0%) 2 (3.6%)\n\n1 n (%); 2 Fisher\u2019s exact test; Pearson\u2019s Chi-squared test; TBP = total body photography.\n\n**Figure 7. Odds ratio for variables influencing the trustworthiness of smartphones\u2019 risk assessment in**\nmelanoma detection.\n\n-----\n\n_Cancers 2022, 14, 3829_ 13 of 21\n\n3.3.3. Impact of AI vs. Dermatologists\u2019 Examination on Patients\u2019 Fear of Developing\nSkin Cancer\n\nMost participants indicated that an examination by a physician reduced their fear of\ndeveloping skin cancer, namely in 89% (49/55) among high-risk patients and 81% (48/59)\namong melanoma patients. Comparably, the 2D TBP imaging achieved the same effect in\n78% (43/55) resp. 76% (45/59), and the 3D TBP device in 82% (45/55) resp. 75% (44/59). In\ncontrast, the assessment with the smartphone app appeased fear of skin cancer in only 33%\n(18/55) of patients at high-risk for melanoma and in 32% (19/59) of patients with melanoma.\nIndeed, one high-risk patient (1.8%) even reported an increased fear of developing skin\ncancer by using the smartphone app (Supplementary Table S2).\n\n3.3.4. Patients\u2019 Subjective Assessment of the Accuracy of AI vs. Dermatologists\n\nPatients expected reliable results with the highest accuracy by both the assessment\nby a physician (98% (54/55) among high-risk patients, 92% (54/59) among melanoma\npatients) and by the 2D TBP imaging (82% (45/55) resp. 86% (51/59)) as well as the 3D\nTBP device (89% (49/55) resp. 88% (52/59)). Only 16% (9/55) of high-risk patients and\n31% (18/59) of melanoma patients expected reliable results from the smartphone app\n(Supplementary Table S2). No significant differences were identified between the two risk\ngroups in the evaluations.\n\n3.3.5. Patient Preference for Skin Cancer Screening\n\nBoth cohorts favored a combination of dermatologist and 3D TBP risk assessment for\nthe examination of pigmented skin lesions (64% (35/55) among patients at high-risk for\nmelanoma, 51% (30/59) among melanoma patients), while neither preferred assessment\nby a smartphone app alone (Figure 8A, Supplementary Table S2). The combination of\ndermatologist and smartphone app was favored by only 1.8% (1/55) of patients at high-risk\nfor melanoma and 3.4% (2/59) of patients with melanoma.\n\n**Figure 8. Cont.**\n\n-----\n\n_Cancers 2022, 14, 3829_ 14 of 21\n\n**Figure 8. (A) Patient preference for mole assessment (patients at high-risk for melanoma, n = 55;**\npatients with melanoma, n = 59); (B) Patient preference for AI in skin cancer screening (patients at\nhigh-risk for melanoma, n = 55; patients with melanoma, n = 59).\n\nRegarding patient preference for skin cancer screenings, almost all high-risk patients\n(98% (54/55)) and melanoma patients (95% (56/59)) indicated their belief that AI can improve\na physician\u2019s diagnostic performance. Most patients (64% (35/55) among high-risk patients,\n54% (32/59) among melanoma patients) would prefer that the physicians would always\nconsider the result of AI in their diagnosis (Figure 8B) (Supplementary Table S2).\n\n3.3.6. Dermatologists\u2019 Perspective of Smartphone Apps for Melanoma Screening\n\nAmong the seven dermatologists, they stated in only 5.3% of the skin cancer screenings\n(6/114) that the smartphone app increased diagnostic confidence and in only 8.8% of the\nassessments (10/114) they trusted the app (Supplementary Table S3).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "High"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmenting the accuracy of trainee doctors in diagnosing skin lesions suspected of skin neoplasms in a real-world setting: A prospective controlled before-and-after study.",
        "abstract": "## BACKGROUND\nAlthough deep neural networks have shown promising results in the diagnosis of skin cancer, a prospective evaluation in a real-world setting could confirm these results. This study aimed to evaluate whether an algorithm (http://b2019.modelderm.com) improves the accuracy of nondermatologists in diagnosing skin neoplasms.\n## METHODS\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n## RESULTS\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n## CONCLUSIONS\nIn real-world settings, AI augmented the diagnostic accuracy of trainee doctors. The limitation of this study is that the algorithm was tested only for Asians recruited from a single region. Additional international randomized controlled trials involving various ethnicities are required.\n",
        "methods": "##### Methods\n\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n\nTraining of the algorithm\n\nThe training history of our algorithm (Model Dermatology; http://modelderm.com) was described in previous studies [5, 9, 12, 15, 17]. Image crops of normal and benign disorders were annotated based on the image findings and these image crops were used for the training to reduce false positives for common benign disorders. The classifier of the algorithm was trained with 721,749 image crops of 178 disease classes. With NVIDIA Caffe (https://github.com/nvidia/caffe; version 0.17.2, CUDA 10.0, cuDNN 7.6.2), we trained our CNN models using a transfer learning method using ImageNet pretrained models. Histogram normalization was performed as a preprocessing step before training the models. The output values of SE-Net [18] and SE-ResNeXt-50 were arithmetically averaged to obtain a final model output. Along with three potential diagnoses, the algorithm reports a malignancy score (range: 0~100) using the following formula: Malignancy score = (basal cell carcinoma output + squamous cell carcinoma (SCC) output + SCC in situ output + keratoacanthoma output + malignant melanoma output) \u00d7 100 + (actinic keratosis output + ulcer output) \u00d7 20. The algorithm reports an overall risk of malignancy as \u201cLow\u201d, \u201cMedium,\u201d or \u201cHigh\u201d. The algorithm reports the risk of malignancy as \u201cLow\u201d when the malignancy score is below 10, \u201cMedium\u201d when the score is between 10 and 20, and \u201cHigh\u201d when the score is over 20.\n\nValidation of the algorithm\n\nAfter obtaining approval from the institutional review board of Asan Medical Center (2018\u20131130), a prospective study was performed at two tertiary care centers in Korea (230 cases from Department of Dermatology, Asan Medical Center, and 55 cases from Seoul National University, Bundang Hospital) between February 1, 2020, and November 7, 2020. The algorithm (Model Dermatology, build 2019; https://b2019.modelderm.com) developed in our previous study [5, 15] was used. The algorithm suggests the three most probable diagnosis of uploaded photographs and also reports a malignancy score (range: 0\u2013100) (Supplementary Methods). After obtaining informed consent, all patients (age > 19 years) who had skin neoplasms suspected of malignancy by either patient or physician were recruited. Exclusion criteria were patient refusal, broken blindness, the wrong version of the algorithm, non-real-time analysis, and exposure of the biopsy results in the referral note (Fig 1). If first impressions were recorded at >24 h after patients\u2019 visits, they were classified as non-real-time. There were no inconclusive cases in the prediction of the algorithm. Ultimately, 270 pathologically diagnosed cases and 15 clinically diagnosed cases were used in the final analysis (Table 1 and S1 Table). A total of 139 and 131 cases were pathologically diagnosed in the AI group and the control group, respectively. A total of 15 cases (5 cases = AI group, 10 cases = Control group) were clinically diagnosed because the attending physicians concluded that they were definitely benign cases and do not to be biopsied. A total of 10 attending physicians (11.4 \u00b1 8.8 years\u2019 experience after board certification), 11 dermatology trainees, and 7 intern doctors participated in this study (S2 Table). Attending physicians routinely recorded their diagnoses after thorough examinations. The trainees who were blinded to attending physicians\u2019 diagnoses evaluated the patients. After quasirandomization using odd/even patient ID, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and provided their diagnoses up to three predictions. In the AI group, trainees selected one photograph and uploaded on http://b2019.modelderm.com. After referring to the algorithm\u2019s three diagnoses and the malignancy score, they were given an opportunity to modify their initial diagnoses. In the control group, trainees just reviewed the photographs once again then provided the after-diagnoses. Top accuracy was calculated as an evaluating metric. Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u201ccorrect.\u201d Only an exact diagnosis was recorded as correct. For evaluating the sensitivity and specificity of malignancy prediction, the physicians\u2019 diagnoses were transformed into either malignant or benign. Top accuracies were compared using two-tailed paired Wilcoxon signed-rank tests (R version 3.5.3), and a P value of < .05 was considered statistically significant.",
        "results": "##### Results\n\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n\nResult of the AI group\n\nAfter analyzing the accuracies before and after assistance, it was noted that the Top-1/Top-2/Top-3 accuracies after assistance were significantly higher than those before assistance (before = 46.5%/ 54.2%/ 54.9%; after = 58.3%/ 70.1%/ 71.5%; P = .008/ < .001/ < .001). The Top-1/Top-2/Top-3 accuracies of the attending dermatologists were 61.8%/ 69.4%/ 71.5%, respectively, and those of the standalone algorithm were 53.5%/ 66.0%/ 70.8%, respectively. In 42.4% (61/144) cases, the Top-1 diagnosis of the algorithm was coherent with that of the trainees, and in 50.0% (72/144) cases, the Top-1 of the algorithm was coherent with that of the attending physicians. The Top-1 of the trainees was coherent with that of the attending physicians in 52.8% (76/144) cases. The trainees revised 28.5% (41/144) of their Top-1 diagnosis after reviewing three diagnoses of the algorithm. A total of 70% (29/41) of their revised answers were correct, whereas 29% (12/41) of their revised answers were incorrect. For determining malignancy, the sensitivity/specificity derived from the Top-1 was 78.3%/ 88.4% before the assistance and 73.9%/ 94.2% after the assistance (Table 2, P = .77/ = .06). The sensitivity/specificity of the attending dermatologists was 82.6%/ 91.7% and that of the patients were 56.5%/ 42.6%. The sensitivity/specificity derived from the Top-1 diagnosis of the algorithm was 52.2%/ 93.4%. The sensitivity/specificity at the threshold of the risk \u201cMedium\u201d using the malignancy score was 95.7%/ 60.3% and that at the threshold of the risk \u201cHigh\u201d was 82.6%/ 70.2%.\n\nResult of the control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between before and after reviewing photographs were not significant (Control-Before, 46.1%/ 64.5%/ 66.7%; Control-After, 51.8%/ 66.7/ 68.1%; P = .19/ = .42/ = .35). For determining malignancy, the sensitivity/specificity derived from the Top-1 diagnosis was 65.5%/ 81.3% before reviewing and 65.5%/ 86.6% after reviewing (Table 2, P = 1.00/ = .09). The sensitivity/specificity of the attending dermatologists was 79.3%/ 90.2% and that of the patients was 48.1%/ 44.5%.\n\nAI group versus control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between the AI group and the Control were not significant (AI Group = 58.3%/ 70.1%/ 71.5%; Control Group = 51.8%/ 66.7%/ 68.1%; P = .27/ = .53/ = .53). Summarized key results were described in S4 Table.",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmenting the accuracy of trainee doctors in diagnosing skin lesions suspected of skin neoplasms in a real-world setting: A prospective controlled before-and-after study.",
        "abstract": "## BACKGROUND\nAlthough deep neural networks have shown promising results in the diagnosis of skin cancer, a prospective evaluation in a real-world setting could confirm these results. This study aimed to evaluate whether an algorithm (http://b2019.modelderm.com) improves the accuracy of nondermatologists in diagnosing skin neoplasms.\n## METHODS\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n## RESULTS\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n## CONCLUSIONS\nIn real-world settings, AI augmented the diagnostic accuracy of trainee doctors. The limitation of this study is that the algorithm was tested only for Asians recruited from a single region. Additional international randomized controlled trials involving various ethnicities are required.\n",
        "methods": "##### Methods\n\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n\nTraining of the algorithm\n\nThe training history of our algorithm (Model Dermatology; http://modelderm.com) was described in previous studies [5, 9, 12, 15, 17]. Image crops of normal and benign disorders were annotated based on the image findings and these image crops were used for the training to reduce false positives for common benign disorders. The classifier of the algorithm was trained with 721,749 image crops of 178 disease classes. With NVIDIA Caffe (https://github.com/nvidia/caffe; version 0.17.2, CUDA 10.0, cuDNN 7.6.2), we trained our CNN models using a transfer learning method using ImageNet pretrained models. Histogram normalization was performed as a preprocessing step before training the models. The output values of SE-Net [18] and SE-ResNeXt-50 were arithmetically averaged to obtain a final model output. Along with three potential diagnoses, the algorithm reports a malignancy score (range: 0~100) using the following formula: Malignancy score = (basal cell carcinoma output + squamous cell carcinoma (SCC) output + SCC in situ output + keratoacanthoma output + malignant melanoma output) \u00d7 100 + (actinic keratosis output + ulcer output) \u00d7 20. The algorithm reports an overall risk of malignancy as \u201cLow\u201d, \u201cMedium,\u201d or \u201cHigh\u201d. The algorithm reports the risk of malignancy as \u201cLow\u201d when the malignancy score is below 10, \u201cMedium\u201d when the score is between 10 and 20, and \u201cHigh\u201d when the score is over 20.\n\nValidation of the algorithm\n\nAfter obtaining approval from the institutional review board of Asan Medical Center (2018\u20131130), a prospective study was performed at two tertiary care centers in Korea (230 cases from Department of Dermatology, Asan Medical Center, and 55 cases from Seoul National University, Bundang Hospital) between February 1, 2020, and November 7, 2020. The algorithm (Model Dermatology, build 2019; https://b2019.modelderm.com) developed in our previous study [5, 15] was used. The algorithm suggests the three most probable diagnosis of uploaded photographs and also reports a malignancy score (range: 0\u2013100) (Supplementary Methods). After obtaining informed consent, all patients (age > 19 years) who had skin neoplasms suspected of malignancy by either patient or physician were recruited. Exclusion criteria were patient refusal, broken blindness, the wrong version of the algorithm, non-real-time analysis, and exposure of the biopsy results in the referral note (Fig 1). If first impressions were recorded at >24 h after patients\u2019 visits, they were classified as non-real-time. There were no inconclusive cases in the prediction of the algorithm. Ultimately, 270 pathologically diagnosed cases and 15 clinically diagnosed cases were used in the final analysis (Table 1 and S1 Table). A total of 139 and 131 cases were pathologically diagnosed in the AI group and the control group, respectively. A total of 15 cases (5 cases = AI group, 10 cases = Control group) were clinically diagnosed because the attending physicians concluded that they were definitely benign cases and do not to be biopsied. A total of 10 attending physicians (11.4 \u00b1 8.8 years\u2019 experience after board certification), 11 dermatology trainees, and 7 intern doctors participated in this study (S2 Table). Attending physicians routinely recorded their diagnoses after thorough examinations. The trainees who were blinded to attending physicians\u2019 diagnoses evaluated the patients. After quasirandomization using odd/even patient ID, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and provided their diagnoses up to three predictions. In the AI group, trainees selected one photograph and uploaded on http://b2019.modelderm.com. After referring to the algorithm\u2019s three diagnoses and the malignancy score, they were given an opportunity to modify their initial diagnoses. In the control group, trainees just reviewed the photographs once again then provided the after-diagnoses. Top accuracy was calculated as an evaluating metric. Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u201ccorrect.\u201d Only an exact diagnosis was recorded as correct. For evaluating the sensitivity and specificity of malignancy prediction, the physicians\u2019 diagnoses were transformed into either malignant or benign. Top accuracies were compared using two-tailed paired Wilcoxon signed-rank tests (R version 3.5.3), and a P value of < .05 was considered statistically significant.",
        "results": "##### Results\n\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n\nResult of the AI group\n\nAfter analyzing the accuracies before and after assistance, it was noted that the Top-1/Top-2/Top-3 accuracies after assistance were significantly higher than those before assistance (before = 46.5%/ 54.2%/ 54.9%; after = 58.3%/ 70.1%/ 71.5%; P = .008/ < .001/ < .001). The Top-1/Top-2/Top-3 accuracies of the attending dermatologists were 61.8%/ 69.4%/ 71.5%, respectively, and those of the standalone algorithm were 53.5%/ 66.0%/ 70.8%, respectively. In 42.4% (61/144) cases, the Top-1 diagnosis of the algorithm was coherent with that of the trainees, and in 50.0% (72/144) cases, the Top-1 of the algorithm was coherent with that of the attending physicians. The Top-1 of the trainees was coherent with that of the attending physicians in 52.8% (76/144) cases. The trainees revised 28.5% (41/144) of their Top-1 diagnosis after reviewing three diagnoses of the algorithm. A total of 70% (29/41) of their revised answers were correct, whereas 29% (12/41) of their revised answers were incorrect. For determining malignancy, the sensitivity/specificity derived from the Top-1 was 78.3%/ 88.4% before the assistance and 73.9%/ 94.2% after the assistance (Table 2, P = .77/ = .06). The sensitivity/specificity of the attending dermatologists was 82.6%/ 91.7% and that of the patients were 56.5%/ 42.6%. The sensitivity/specificity derived from the Top-1 diagnosis of the algorithm was 52.2%/ 93.4%. The sensitivity/specificity at the threshold of the risk \u201cMedium\u201d using the malignancy score was 95.7%/ 60.3% and that at the threshold of the risk \u201cHigh\u201d was 82.6%/ 70.2%.\n\nResult of the control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between before and after reviewing photographs were not significant (Control-Before, 46.1%/ 64.5%/ 66.7%; Control-After, 51.8%/ 66.7/ 68.1%; P = .19/ = .42/ = .35). For determining malignancy, the sensitivity/specificity derived from the Top-1 diagnosis was 65.5%/ 81.3% before reviewing and 65.5%/ 86.6% after reviewing (Table 2, P = 1.00/ = .09). The sensitivity/specificity of the attending dermatologists was 79.3%/ 90.2% and that of the patients was 48.1%/ 44.5%.\n\nAI group versus control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between the AI group and the Control were not significant (AI Group = 58.3%/ 70.1%/ 71.5%; Control Group = 51.8%/ 66.7%/ 68.1%; P = .27/ = .53/ = .53). Summarized key results were described in S4 Table.",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmenting the accuracy of trainee doctors in diagnosing skin lesions suspected of skin neoplasms in a real-world setting: A prospective controlled before-and-after study.",
        "abstract": "## BACKGROUND\nAlthough deep neural networks have shown promising results in the diagnosis of skin cancer, a prospective evaluation in a real-world setting could confirm these results. This study aimed to evaluate whether an algorithm (http://b2019.modelderm.com) improves the accuracy of nondermatologists in diagnosing skin neoplasms.\n## METHODS\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n## RESULTS\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n## CONCLUSIONS\nIn real-world settings, AI augmented the diagnostic accuracy of trainee doctors. The limitation of this study is that the algorithm was tested only for Asians recruited from a single region. Additional international randomized controlled trials involving various ethnicities are required.\n",
        "methods": "##### Methods\n\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n\nTraining of the algorithm\n\nThe training history of our algorithm (Model Dermatology; http://modelderm.com) was described in previous studies [5, 9, 12, 15, 17]. Image crops of normal and benign disorders were annotated based on the image findings and these image crops were used for the training to reduce false positives for common benign disorders. The classifier of the algorithm was trained with 721,749 image crops of 178 disease classes. With NVIDIA Caffe (https://github.com/nvidia/caffe; version 0.17.2, CUDA 10.0, cuDNN 7.6.2), we trained our CNN models using a transfer learning method using ImageNet pretrained models. Histogram normalization was performed as a preprocessing step before training the models. The output values of SE-Net [18] and SE-ResNeXt-50 were arithmetically averaged to obtain a final model output. Along with three potential diagnoses, the algorithm reports a malignancy score (range: 0~100) using the following formula: Malignancy score = (basal cell carcinoma output + squamous cell carcinoma (SCC) output + SCC in situ output + keratoacanthoma output + malignant melanoma output) \u00d7 100 + (actinic keratosis output + ulcer output) \u00d7 20. The algorithm reports an overall risk of malignancy as \u201cLow\u201d, \u201cMedium,\u201d or \u201cHigh\u201d. The algorithm reports the risk of malignancy as \u201cLow\u201d when the malignancy score is below 10, \u201cMedium\u201d when the score is between 10 and 20, and \u201cHigh\u201d when the score is over 20.\n\nValidation of the algorithm\n\nAfter obtaining approval from the institutional review board of Asan Medical Center (2018\u20131130), a prospective study was performed at two tertiary care centers in Korea (230 cases from Department of Dermatology, Asan Medical Center, and 55 cases from Seoul National University, Bundang Hospital) between February 1, 2020, and November 7, 2020. The algorithm (Model Dermatology, build 2019; https://b2019.modelderm.com) developed in our previous study [5, 15] was used. The algorithm suggests the three most probable diagnosis of uploaded photographs and also reports a malignancy score (range: 0\u2013100) (Supplementary Methods). After obtaining informed consent, all patients (age > 19 years) who had skin neoplasms suspected of malignancy by either patient or physician were recruited. Exclusion criteria were patient refusal, broken blindness, the wrong version of the algorithm, non-real-time analysis, and exposure of the biopsy results in the referral note (Fig 1). If first impressions were recorded at >24 h after patients\u2019 visits, they were classified as non-real-time. There were no inconclusive cases in the prediction of the algorithm. Ultimately, 270 pathologically diagnosed cases and 15 clinically diagnosed cases were used in the final analysis (Table 1 and S1 Table). A total of 139 and 131 cases were pathologically diagnosed in the AI group and the control group, respectively. A total of 15 cases (5 cases = AI group, 10 cases = Control group) were clinically diagnosed because the attending physicians concluded that they were definitely benign cases and do not to be biopsied. A total of 10 attending physicians (11.4 \u00b1 8.8 years\u2019 experience after board certification), 11 dermatology trainees, and 7 intern doctors participated in this study (S2 Table). Attending physicians routinely recorded their diagnoses after thorough examinations. The trainees who were blinded to attending physicians\u2019 diagnoses evaluated the patients. After quasirandomization using odd/even patient ID, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and provided their diagnoses up to three predictions. In the AI group, trainees selected one photograph and uploaded on http://b2019.modelderm.com. After referring to the algorithm\u2019s three diagnoses and the malignancy score, they were given an opportunity to modify their initial diagnoses. In the control group, trainees just reviewed the photographs once again then provided the after-diagnoses. Top accuracy was calculated as an evaluating metric. Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u201ccorrect.\u201d Only an exact diagnosis was recorded as correct. For evaluating the sensitivity and specificity of malignancy prediction, the physicians\u2019 diagnoses were transformed into either malignant or benign. Top accuracies were compared using two-tailed paired Wilcoxon signed-rank tests (R version 3.5.3), and a P value of < .05 was considered statistically significant.",
        "results": "##### Results\n\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n\nResult of the AI group\n\nAfter analyzing the accuracies before and after assistance, it was noted that the Top-1/Top-2/Top-3 accuracies after assistance were significantly higher than those before assistance (before = 46.5%/ 54.2%/ 54.9%; after = 58.3%/ 70.1%/ 71.5%; P = .008/ < .001/ < .001). The Top-1/Top-2/Top-3 accuracies of the attending dermatologists were 61.8%/ 69.4%/ 71.5%, respectively, and those of the standalone algorithm were 53.5%/ 66.0%/ 70.8%, respectively. In 42.4% (61/144) cases, the Top-1 diagnosis of the algorithm was coherent with that of the trainees, and in 50.0% (72/144) cases, the Top-1 of the algorithm was coherent with that of the attending physicians. The Top-1 of the trainees was coherent with that of the attending physicians in 52.8% (76/144) cases. The trainees revised 28.5% (41/144) of their Top-1 diagnosis after reviewing three diagnoses of the algorithm. A total of 70% (29/41) of their revised answers were correct, whereas 29% (12/41) of their revised answers were incorrect. For determining malignancy, the sensitivity/specificity derived from the Top-1 was 78.3%/ 88.4% before the assistance and 73.9%/ 94.2% after the assistance (Table 2, P = .77/ = .06). The sensitivity/specificity of the attending dermatologists was 82.6%/ 91.7% and that of the patients were 56.5%/ 42.6%. The sensitivity/specificity derived from the Top-1 diagnosis of the algorithm was 52.2%/ 93.4%. The sensitivity/specificity at the threshold of the risk \u201cMedium\u201d using the malignancy score was 95.7%/ 60.3% and that at the threshold of the risk \u201cHigh\u201d was 82.6%/ 70.2%.\n\nResult of the control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between before and after reviewing photographs were not significant (Control-Before, 46.1%/ 64.5%/ 66.7%; Control-After, 51.8%/ 66.7/ 68.1%; P = .19/ = .42/ = .35). For determining malignancy, the sensitivity/specificity derived from the Top-1 diagnosis was 65.5%/ 81.3% before reviewing and 65.5%/ 86.6% after reviewing (Table 2, P = 1.00/ = .09). The sensitivity/specificity of the attending dermatologists was 79.3%/ 90.2% and that of the patients was 48.1%/ 44.5%.\n\nAI group versus control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between the AI group and the Control were not significant (AI Group = 58.3%/ 70.1%/ 71.5%; Control Group = 51.8%/ 66.7%/ 68.1%; P = .27/ = .53/ = .53). Summarized key results were described in S4 Table.",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmenting the accuracy of trainee doctors in diagnosing skin lesions suspected of skin neoplasms in a real-world setting: A prospective controlled before-and-after study.",
        "abstract": "## BACKGROUND\nAlthough deep neural networks have shown promising results in the diagnosis of skin cancer, a prospective evaluation in a real-world setting could confirm these results. This study aimed to evaluate whether an algorithm (http://b2019.modelderm.com) improves the accuracy of nondermatologists in diagnosing skin neoplasms.\n## METHODS\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n## RESULTS\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n## CONCLUSIONS\nIn real-world settings, AI augmented the diagnostic accuracy of trainee doctors. The limitation of this study is that the algorithm was tested only for Asians recruited from a single region. Additional international randomized controlled trials involving various ethnicities are required.\n",
        "methods": "##### Methods\n\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n\nTraining of the algorithm\n\nThe training history of our algorithm (Model Dermatology; http://modelderm.com) was described in previous studies [5, 9, 12, 15, 17]. Image crops of normal and benign disorders were annotated based on the image findings and these image crops were used for the training to reduce false positives for common benign disorders. The classifier of the algorithm was trained with 721,749 image crops of 178 disease classes. With NVIDIA Caffe (https://github.com/nvidia/caffe; version 0.17.2, CUDA 10.0, cuDNN 7.6.2), we trained our CNN models using a transfer learning method using ImageNet pretrained models. Histogram normalization was performed as a preprocessing step before training the models. The output values of SE-Net [18] and SE-ResNeXt-50 were arithmetically averaged to obtain a final model output. Along with three potential diagnoses, the algorithm reports a malignancy score (range: 0~100) using the following formula: Malignancy score = (basal cell carcinoma output + squamous cell carcinoma (SCC) output + SCC in situ output + keratoacanthoma output + malignant melanoma output) \u00d7 100 + (actinic keratosis output + ulcer output) \u00d7 20. The algorithm reports an overall risk of malignancy as \u201cLow\u201d, \u201cMedium,\u201d or \u201cHigh\u201d. The algorithm reports the risk of malignancy as \u201cLow\u201d when the malignancy score is below 10, \u201cMedium\u201d when the score is between 10 and 20, and \u201cHigh\u201d when the score is over 20.\n\nValidation of the algorithm\n\nAfter obtaining approval from the institutional review board of Asan Medical Center (2018\u20131130), a prospective study was performed at two tertiary care centers in Korea (230 cases from Department of Dermatology, Asan Medical Center, and 55 cases from Seoul National University, Bundang Hospital) between February 1, 2020, and November 7, 2020. The algorithm (Model Dermatology, build 2019; https://b2019.modelderm.com) developed in our previous study [5, 15] was used. The algorithm suggests the three most probable diagnosis of uploaded photographs and also reports a malignancy score (range: 0\u2013100) (Supplementary Methods). After obtaining informed consent, all patients (age > 19 years) who had skin neoplasms suspected of malignancy by either patient or physician were recruited. Exclusion criteria were patient refusal, broken blindness, the wrong version of the algorithm, non-real-time analysis, and exposure of the biopsy results in the referral note (Fig 1). If first impressions were recorded at >24 h after patients\u2019 visits, they were classified as non-real-time. There were no inconclusive cases in the prediction of the algorithm. Ultimately, 270 pathologically diagnosed cases and 15 clinically diagnosed cases were used in the final analysis (Table 1 and S1 Table). A total of 139 and 131 cases were pathologically diagnosed in the AI group and the control group, respectively. A total of 15 cases (5 cases = AI group, 10 cases = Control group) were clinically diagnosed because the attending physicians concluded that they were definitely benign cases and do not to be biopsied. A total of 10 attending physicians (11.4 \u00b1 8.8 years\u2019 experience after board certification), 11 dermatology trainees, and 7 intern doctors participated in this study (S2 Table). Attending physicians routinely recorded their diagnoses after thorough examinations. The trainees who were blinded to attending physicians\u2019 diagnoses evaluated the patients. After quasirandomization using odd/even patient ID, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and provided their diagnoses up to three predictions. In the AI group, trainees selected one photograph and uploaded on http://b2019.modelderm.com. After referring to the algorithm\u2019s three diagnoses and the malignancy score, they were given an opportunity to modify their initial diagnoses. In the control group, trainees just reviewed the photographs once again then provided the after-diagnoses. Top accuracy was calculated as an evaluating metric. Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u201ccorrect.\u201d Only an exact diagnosis was recorded as correct. For evaluating the sensitivity and specificity of malignancy prediction, the physicians\u2019 diagnoses were transformed into either malignant or benign. Top accuracies were compared using two-tailed paired Wilcoxon signed-rank tests (R version 3.5.3), and a P value of < .05 was considered statistically significant.",
        "results": "##### Results\n\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n\nResult of the AI group\n\nAfter analyzing the accuracies before and after assistance, it was noted that the Top-1/Top-2/Top-3 accuracies after assistance were significantly higher than those before assistance (before = 46.5%/ 54.2%/ 54.9%; after = 58.3%/ 70.1%/ 71.5%; P = .008/ < .001/ < .001). The Top-1/Top-2/Top-3 accuracies of the attending dermatologists were 61.8%/ 69.4%/ 71.5%, respectively, and those of the standalone algorithm were 53.5%/ 66.0%/ 70.8%, respectively. In 42.4% (61/144) cases, the Top-1 diagnosis of the algorithm was coherent with that of the trainees, and in 50.0% (72/144) cases, the Top-1 of the algorithm was coherent with that of the attending physicians. The Top-1 of the trainees was coherent with that of the attending physicians in 52.8% (76/144) cases. The trainees revised 28.5% (41/144) of their Top-1 diagnosis after reviewing three diagnoses of the algorithm. A total of 70% (29/41) of their revised answers were correct, whereas 29% (12/41) of their revised answers were incorrect. For determining malignancy, the sensitivity/specificity derived from the Top-1 was 78.3%/ 88.4% before the assistance and 73.9%/ 94.2% after the assistance (Table 2, P = .77/ = .06). The sensitivity/specificity of the attending dermatologists was 82.6%/ 91.7% and that of the patients were 56.5%/ 42.6%. The sensitivity/specificity derived from the Top-1 diagnosis of the algorithm was 52.2%/ 93.4%. The sensitivity/specificity at the threshold of the risk \u201cMedium\u201d using the malignancy score was 95.7%/ 60.3% and that at the threshold of the risk \u201cHigh\u201d was 82.6%/ 70.2%.\n\nResult of the control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between before and after reviewing photographs were not significant (Control-Before, 46.1%/ 64.5%/ 66.7%; Control-After, 51.8%/ 66.7/ 68.1%; P = .19/ = .42/ = .35). For determining malignancy, the sensitivity/specificity derived from the Top-1 diagnosis was 65.5%/ 81.3% before reviewing and 65.5%/ 86.6% after reviewing (Table 2, P = 1.00/ = .09). The sensitivity/specificity of the attending dermatologists was 79.3%/ 90.2% and that of the patients was 48.1%/ 44.5%.\n\nAI group versus control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between the AI group and the Control were not significant (AI Group = 58.3%/ 70.1%/ 71.5%; Control Group = 51.8%/ 66.7%/ 68.1%; P = .27/ = .53/ = .53). Summarized key results were described in S4 Table.",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmenting the accuracy of trainee doctors in diagnosing skin lesions suspected of skin neoplasms in a real-world setting: A prospective controlled before-and-after study.",
        "abstract": "## BACKGROUND\nAlthough deep neural networks have shown promising results in the diagnosis of skin cancer, a prospective evaluation in a real-world setting could confirm these results. This study aimed to evaluate whether an algorithm (http://b2019.modelderm.com) improves the accuracy of nondermatologists in diagnosing skin neoplasms.\n## METHODS\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n## RESULTS\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n## CONCLUSIONS\nIn real-world settings, AI augmented the diagnostic accuracy of trainee doctors. The limitation of this study is that the algorithm was tested only for Asians recruited from a single region. Additional international randomized controlled trials involving various ethnicities are required.\n",
        "methods": "##### Methods\n\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n\nTraining of the algorithm\n\nThe training history of our algorithm (Model Dermatology; http://modelderm.com) was described in previous studies [5, 9, 12, 15, 17]. Image crops of normal and benign disorders were annotated based on the image findings and these image crops were used for the training to reduce false positives for common benign disorders. The classifier of the algorithm was trained with 721,749 image crops of 178 disease classes. With NVIDIA Caffe (https://github.com/nvidia/caffe; version 0.17.2, CUDA 10.0, cuDNN 7.6.2), we trained our CNN models using a transfer learning method using ImageNet pretrained models. Histogram normalization was performed as a preprocessing step before training the models. The output values of SE-Net [18] and SE-ResNeXt-50 were arithmetically averaged to obtain a final model output. Along with three potential diagnoses, the algorithm reports a malignancy score (range: 0~100) using the following formula: Malignancy score = (basal cell carcinoma output + squamous cell carcinoma (SCC) output + SCC in situ output + keratoacanthoma output + malignant melanoma output) \u00d7 100 + (actinic keratosis output + ulcer output) \u00d7 20. The algorithm reports an overall risk of malignancy as \u201cLow\u201d, \u201cMedium,\u201d or \u201cHigh\u201d. The algorithm reports the risk of malignancy as \u201cLow\u201d when the malignancy score is below 10, \u201cMedium\u201d when the score is between 10 and 20, and \u201cHigh\u201d when the score is over 20.\n\nValidation of the algorithm\n\nAfter obtaining approval from the institutional review board of Asan Medical Center (2018\u20131130), a prospective study was performed at two tertiary care centers in Korea (230 cases from Department of Dermatology, Asan Medical Center, and 55 cases from Seoul National University, Bundang Hospital) between February 1, 2020, and November 7, 2020. The algorithm (Model Dermatology, build 2019; https://b2019.modelderm.com) developed in our previous study [5, 15] was used. The algorithm suggests the three most probable diagnosis of uploaded photographs and also reports a malignancy score (range: 0\u2013100) (Supplementary Methods). After obtaining informed consent, all patients (age > 19 years) who had skin neoplasms suspected of malignancy by either patient or physician were recruited. Exclusion criteria were patient refusal, broken blindness, the wrong version of the algorithm, non-real-time analysis, and exposure of the biopsy results in the referral note (Fig 1). If first impressions were recorded at >24 h after patients\u2019 visits, they were classified as non-real-time. There were no inconclusive cases in the prediction of the algorithm. Ultimately, 270 pathologically diagnosed cases and 15 clinically diagnosed cases were used in the final analysis (Table 1 and S1 Table). A total of 139 and 131 cases were pathologically diagnosed in the AI group and the control group, respectively. A total of 15 cases (5 cases = AI group, 10 cases = Control group) were clinically diagnosed because the attending physicians concluded that they were definitely benign cases and do not to be biopsied. A total of 10 attending physicians (11.4 \u00b1 8.8 years\u2019 experience after board certification), 11 dermatology trainees, and 7 intern doctors participated in this study (S2 Table). Attending physicians routinely recorded their diagnoses after thorough examinations. The trainees who were blinded to attending physicians\u2019 diagnoses evaluated the patients. After quasirandomization using odd/even patient ID, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and provided their diagnoses up to three predictions. In the AI group, trainees selected one photograph and uploaded on http://b2019.modelderm.com. After referring to the algorithm\u2019s three diagnoses and the malignancy score, they were given an opportunity to modify their initial diagnoses. In the control group, trainees just reviewed the photographs once again then provided the after-diagnoses. Top accuracy was calculated as an evaluating metric. Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u201ccorrect.\u201d Only an exact diagnosis was recorded as correct. For evaluating the sensitivity and specificity of malignancy prediction, the physicians\u2019 diagnoses were transformed into either malignant or benign. Top accuracies were compared using two-tailed paired Wilcoxon signed-rank tests (R version 3.5.3), and a P value of < .05 was considered statistically significant.",
        "results": "##### Results\n\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n\nResult of the AI group\n\nAfter analyzing the accuracies before and after assistance, it was noted that the Top-1/Top-2/Top-3 accuracies after assistance were significantly higher than those before assistance (before = 46.5%/ 54.2%/ 54.9%; after = 58.3%/ 70.1%/ 71.5%; P = .008/ < .001/ < .001). The Top-1/Top-2/Top-3 accuracies of the attending dermatologists were 61.8%/ 69.4%/ 71.5%, respectively, and those of the standalone algorithm were 53.5%/ 66.0%/ 70.8%, respectively. In 42.4% (61/144) cases, the Top-1 diagnosis of the algorithm was coherent with that of the trainees, and in 50.0% (72/144) cases, the Top-1 of the algorithm was coherent with that of the attending physicians. The Top-1 of the trainees was coherent with that of the attending physicians in 52.8% (76/144) cases. The trainees revised 28.5% (41/144) of their Top-1 diagnosis after reviewing three diagnoses of the algorithm. A total of 70% (29/41) of their revised answers were correct, whereas 29% (12/41) of their revised answers were incorrect. For determining malignancy, the sensitivity/specificity derived from the Top-1 was 78.3%/ 88.4% before the assistance and 73.9%/ 94.2% after the assistance (Table 2, P = .77/ = .06). The sensitivity/specificity of the attending dermatologists was 82.6%/ 91.7% and that of the patients were 56.5%/ 42.6%. The sensitivity/specificity derived from the Top-1 diagnosis of the algorithm was 52.2%/ 93.4%. The sensitivity/specificity at the threshold of the risk \u201cMedium\u201d using the malignancy score was 95.7%/ 60.3% and that at the threshold of the risk \u201cHigh\u201d was 82.6%/ 70.2%.\n\nResult of the control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between before and after reviewing photographs were not significant (Control-Before, 46.1%/ 64.5%/ 66.7%; Control-After, 51.8%/ 66.7/ 68.1%; P = .19/ = .42/ = .35). For determining malignancy, the sensitivity/specificity derived from the Top-1 diagnosis was 65.5%/ 81.3% before reviewing and 65.5%/ 86.6% after reviewing (Table 2, P = 1.00/ = .09). The sensitivity/specificity of the attending dermatologists was 79.3%/ 90.2% and that of the patients was 48.1%/ 44.5%.\n\nAI group versus control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between the AI group and the Control were not significant (AI Group = 58.3%/ 70.1%/ 71.5%; Control Group = 51.8%/ 66.7%/ 68.1%; P = .27/ = .53/ = .53). Summarized key results were described in S4 Table.",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmenting the accuracy of trainee doctors in diagnosing skin lesions suspected of skin neoplasms in a real-world setting: A prospective controlled before-and-after study.",
        "abstract": "## BACKGROUND\nAlthough deep neural networks have shown promising results in the diagnosis of skin cancer, a prospective evaluation in a real-world setting could confirm these results. This study aimed to evaluate whether an algorithm (http://b2019.modelderm.com) improves the accuracy of nondermatologists in diagnosing skin neoplasms.\n## METHODS\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n## RESULTS\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n## CONCLUSIONS\nIn real-world settings, AI augmented the diagnostic accuracy of trainee doctors. The limitation of this study is that the algorithm was tested only for Asians recruited from a single region. Additional international randomized controlled trials involving various ethnicities are required.\n",
        "methods": "##### Methods\n\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n\nTraining of the algorithm\n\nThe training history of our algorithm (Model Dermatology; http://modelderm.com) was described in previous studies [5, 9, 12, 15, 17]. Image crops of normal and benign disorders were annotated based on the image findings and these image crops were used for the training to reduce false positives for common benign disorders. The classifier of the algorithm was trained with 721,749 image crops of 178 disease classes. With NVIDIA Caffe (https://github.com/nvidia/caffe; version 0.17.2, CUDA 10.0, cuDNN 7.6.2), we trained our CNN models using a transfer learning method using ImageNet pretrained models. Histogram normalization was performed as a preprocessing step before training the models. The output values of SE-Net [18] and SE-ResNeXt-50 were arithmetically averaged to obtain a final model output. Along with three potential diagnoses, the algorithm reports a malignancy score (range: 0~100) using the following formula: Malignancy score = (basal cell carcinoma output + squamous cell carcinoma (SCC) output + SCC in situ output + keratoacanthoma output + malignant melanoma output) \u00d7 100 + (actinic keratosis output + ulcer output) \u00d7 20. The algorithm reports an overall risk of malignancy as \u201cLow\u201d, \u201cMedium,\u201d or \u201cHigh\u201d. The algorithm reports the risk of malignancy as \u201cLow\u201d when the malignancy score is below 10, \u201cMedium\u201d when the score is between 10 and 20, and \u201cHigh\u201d when the score is over 20.\n\nValidation of the algorithm\n\nAfter obtaining approval from the institutional review board of Asan Medical Center (2018\u20131130), a prospective study was performed at two tertiary care centers in Korea (230 cases from Department of Dermatology, Asan Medical Center, and 55 cases from Seoul National University, Bundang Hospital) between February 1, 2020, and November 7, 2020. The algorithm (Model Dermatology, build 2019; https://b2019.modelderm.com) developed in our previous study [5, 15] was used. The algorithm suggests the three most probable diagnosis of uploaded photographs and also reports a malignancy score (range: 0\u2013100) (Supplementary Methods). After obtaining informed consent, all patients (age > 19 years) who had skin neoplasms suspected of malignancy by either patient or physician were recruited. Exclusion criteria were patient refusal, broken blindness, the wrong version of the algorithm, non-real-time analysis, and exposure of the biopsy results in the referral note (Fig 1). If first impressions were recorded at >24 h after patients\u2019 visits, they were classified as non-real-time. There were no inconclusive cases in the prediction of the algorithm. Ultimately, 270 pathologically diagnosed cases and 15 clinically diagnosed cases were used in the final analysis (Table 1 and S1 Table). A total of 139 and 131 cases were pathologically diagnosed in the AI group and the control group, respectively. A total of 15 cases (5 cases = AI group, 10 cases = Control group) were clinically diagnosed because the attending physicians concluded that they were definitely benign cases and do not to be biopsied. A total of 10 attending physicians (11.4 \u00b1 8.8 years\u2019 experience after board certification), 11 dermatology trainees, and 7 intern doctors participated in this study (S2 Table). Attending physicians routinely recorded their diagnoses after thorough examinations. The trainees who were blinded to attending physicians\u2019 diagnoses evaluated the patients. After quasirandomization using odd/even patient ID, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and provided their diagnoses up to three predictions. In the AI group, trainees selected one photograph and uploaded on http://b2019.modelderm.com. After referring to the algorithm\u2019s three diagnoses and the malignancy score, they were given an opportunity to modify their initial diagnoses. In the control group, trainees just reviewed the photographs once again then provided the after-diagnoses. Top accuracy was calculated as an evaluating metric. Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u201ccorrect.\u201d Only an exact diagnosis was recorded as correct. For evaluating the sensitivity and specificity of malignancy prediction, the physicians\u2019 diagnoses were transformed into either malignant or benign. Top accuracies were compared using two-tailed paired Wilcoxon signed-rank tests (R version 3.5.3), and a P value of < .05 was considered statistically significant.",
        "results": "##### Results\n\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n\nResult of the AI group\n\nAfter analyzing the accuracies before and after assistance, it was noted that the Top-1/Top-2/Top-3 accuracies after assistance were significantly higher than those before assistance (before = 46.5%/ 54.2%/ 54.9%; after = 58.3%/ 70.1%/ 71.5%; P = .008/ < .001/ < .001). The Top-1/Top-2/Top-3 accuracies of the attending dermatologists were 61.8%/ 69.4%/ 71.5%, respectively, and those of the standalone algorithm were 53.5%/ 66.0%/ 70.8%, respectively. In 42.4% (61/144) cases, the Top-1 diagnosis of the algorithm was coherent with that of the trainees, and in 50.0% (72/144) cases, the Top-1 of the algorithm was coherent with that of the attending physicians. The Top-1 of the trainees was coherent with that of the attending physicians in 52.8% (76/144) cases. The trainees revised 28.5% (41/144) of their Top-1 diagnosis after reviewing three diagnoses of the algorithm. A total of 70% (29/41) of their revised answers were correct, whereas 29% (12/41) of their revised answers were incorrect. For determining malignancy, the sensitivity/specificity derived from the Top-1 was 78.3%/ 88.4% before the assistance and 73.9%/ 94.2% after the assistance (Table 2, P = .77/ = .06). The sensitivity/specificity of the attending dermatologists was 82.6%/ 91.7% and that of the patients were 56.5%/ 42.6%. The sensitivity/specificity derived from the Top-1 diagnosis of the algorithm was 52.2%/ 93.4%. The sensitivity/specificity at the threshold of the risk \u201cMedium\u201d using the malignancy score was 95.7%/ 60.3% and that at the threshold of the risk \u201cHigh\u201d was 82.6%/ 70.2%.\n\nResult of the control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between before and after reviewing photographs were not significant (Control-Before, 46.1%/ 64.5%/ 66.7%; Control-After, 51.8%/ 66.7/ 68.1%; P = .19/ = .42/ = .35). For determining malignancy, the sensitivity/specificity derived from the Top-1 diagnosis was 65.5%/ 81.3% before reviewing and 65.5%/ 86.6% after reviewing (Table 2, P = 1.00/ = .09). The sensitivity/specificity of the attending dermatologists was 79.3%/ 90.2% and that of the patients was 48.1%/ 44.5%.\n\nAI group versus control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between the AI group and the Control were not significant (AI Group = 58.3%/ 70.1%/ 71.5%; Control Group = 51.8%/ 66.7%/ 68.1%; P = .27/ = .53/ = .53). Summarized key results were described in S4 Table.",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmenting the accuracy of trainee doctors in diagnosing skin lesions suspected of skin neoplasms in a real-world setting: A prospective controlled before-and-after study.",
        "abstract": "## BACKGROUND\nAlthough deep neural networks have shown promising results in the diagnosis of skin cancer, a prospective evaluation in a real-world setting could confirm these results. This study aimed to evaluate whether an algorithm (http://b2019.modelderm.com) improves the accuracy of nondermatologists in diagnosing skin neoplasms.\n## METHODS\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n## RESULTS\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n## CONCLUSIONS\nIn real-world settings, AI augmented the diagnostic accuracy of trainee doctors. The limitation of this study is that the algorithm was tested only for Asians recruited from a single region. Additional international randomized controlled trials involving various ethnicities are required.\n",
        "methods": "##### Methods\n\nA total of 285 cases (random series) with skin neoplasms suspected of malignancy by either physicians or patients were recruited in two tertiary care centers located in South Korea. An artificial intelligence (AI) group (144 cases, mean [SD] age, 57.0 [17.7] years; 62 [43.1%] men) was diagnosed via routine examination with photographic review and assistance by the algorithm, whereas the control group (141 cases, mean [SD] age, 61.0 [15.3] years; 52 [36.9%] men) was diagnosed only via routine examination with a photographic review. The accuracy of the nondermatologists before and after the interventions was compared.\n\nTraining of the algorithm\n\nThe training history of our algorithm (Model Dermatology; http://modelderm.com) was described in previous studies [5, 9, 12, 15, 17]. Image crops of normal and benign disorders were annotated based on the image findings and these image crops were used for the training to reduce false positives for common benign disorders. The classifier of the algorithm was trained with 721,749 image crops of 178 disease classes. With NVIDIA Caffe (https://github.com/nvidia/caffe; version 0.17.2, CUDA 10.0, cuDNN 7.6.2), we trained our CNN models using a transfer learning method using ImageNet pretrained models. Histogram normalization was performed as a preprocessing step before training the models. The output values of SE-Net [18] and SE-ResNeXt-50 were arithmetically averaged to obtain a final model output. Along with three potential diagnoses, the algorithm reports a malignancy score (range: 0~100) using the following formula: Malignancy score = (basal cell carcinoma output + squamous cell carcinoma (SCC) output + SCC in situ output + keratoacanthoma output + malignant melanoma output) \u00d7 100 + (actinic keratosis output + ulcer output) \u00d7 20. The algorithm reports an overall risk of malignancy as \u201cLow\u201d, \u201cMedium,\u201d or \u201cHigh\u201d. The algorithm reports the risk of malignancy as \u201cLow\u201d when the malignancy score is below 10, \u201cMedium\u201d when the score is between 10 and 20, and \u201cHigh\u201d when the score is over 20.\n\nValidation of the algorithm\n\nAfter obtaining approval from the institutional review board of Asan Medical Center (2018\u20131130), a prospective study was performed at two tertiary care centers in Korea (230 cases from Department of Dermatology, Asan Medical Center, and 55 cases from Seoul National University, Bundang Hospital) between February 1, 2020, and November 7, 2020. The algorithm (Model Dermatology, build 2019; https://b2019.modelderm.com) developed in our previous study [5, 15] was used. The algorithm suggests the three most probable diagnosis of uploaded photographs and also reports a malignancy score (range: 0\u2013100) (Supplementary Methods). After obtaining informed consent, all patients (age > 19 years) who had skin neoplasms suspected of malignancy by either patient or physician were recruited. Exclusion criteria were patient refusal, broken blindness, the wrong version of the algorithm, non-real-time analysis, and exposure of the biopsy results in the referral note (Fig 1). If first impressions were recorded at >24 h after patients\u2019 visits, they were classified as non-real-time. There were no inconclusive cases in the prediction of the algorithm. Ultimately, 270 pathologically diagnosed cases and 15 clinically diagnosed cases were used in the final analysis (Table 1 and S1 Table). A total of 139 and 131 cases were pathologically diagnosed in the AI group and the control group, respectively. A total of 15 cases (5 cases = AI group, 10 cases = Control group) were clinically diagnosed because the attending physicians concluded that they were definitely benign cases and do not to be biopsied. A total of 10 attending physicians (11.4 \u00b1 8.8 years\u2019 experience after board certification), 11 dermatology trainees, and 7 intern doctors participated in this study (S2 Table). Attending physicians routinely recorded their diagnoses after thorough examinations. The trainees who were blinded to attending physicians\u2019 diagnoses evaluated the patients. After quasirandomization using odd/even patient ID, the trainee took the patient\u2019s medical history, performed physical examinations, took photographs, and provided their diagnoses up to three predictions. In the AI group, trainees selected one photograph and uploaded on http://b2019.modelderm.com. After referring to the algorithm\u2019s three diagnoses and the malignancy score, they were given an opportunity to modify their initial diagnoses. In the control group, trainees just reviewed the photographs once again then provided the after-diagnoses. Top accuracy was calculated as an evaluating metric. Top-(n) accuracy is the accuracy of the Top-(n) diagnoses. If any one of the Top-(n) diagnoses is correct, it counts as \u201ccorrect.\u201d Only an exact diagnosis was recorded as correct. For evaluating the sensitivity and specificity of malignancy prediction, the physicians\u2019 diagnoses were transformed into either malignant or benign. Top accuracies were compared using two-tailed paired Wilcoxon signed-rank tests (R version 3.5.3), and a P value of < .05 was considered statistically significant.",
        "results": "##### Results\n\nAmong the AI group, the accuracy of the first impression (Top-1 accuracy; 58.3%) after the assistance of AI was higher than that before the assistance (46.5%, P = .008). The number of differential diagnoses of the participants increased from 1.9 \u00b1 0.5 to 2.2 \u00b1 0.6 after the assistance (P < .001). In the control group, the difference in the Top-1 accuracy between before and after reviewing photographs was not significant (before, 46.1%; after, 51.8%; P = .19), and the number of differential diagnoses did not significantly increase (before, 2.0 \u00b1 0.4; after, 2.1 \u00b1 0.5; P = .57).\n\nResult of the AI group\n\nAfter analyzing the accuracies before and after assistance, it was noted that the Top-1/Top-2/Top-3 accuracies after assistance were significantly higher than those before assistance (before = 46.5%/ 54.2%/ 54.9%; after = 58.3%/ 70.1%/ 71.5%; P = .008/ < .001/ < .001). The Top-1/Top-2/Top-3 accuracies of the attending dermatologists were 61.8%/ 69.4%/ 71.5%, respectively, and those of the standalone algorithm were 53.5%/ 66.0%/ 70.8%, respectively. In 42.4% (61/144) cases, the Top-1 diagnosis of the algorithm was coherent with that of the trainees, and in 50.0% (72/144) cases, the Top-1 of the algorithm was coherent with that of the attending physicians. The Top-1 of the trainees was coherent with that of the attending physicians in 52.8% (76/144) cases. The trainees revised 28.5% (41/144) of their Top-1 diagnosis after reviewing three diagnoses of the algorithm. A total of 70% (29/41) of their revised answers were correct, whereas 29% (12/41) of their revised answers were incorrect. For determining malignancy, the sensitivity/specificity derived from the Top-1 was 78.3%/ 88.4% before the assistance and 73.9%/ 94.2% after the assistance (Table 2, P = .77/ = .06). The sensitivity/specificity of the attending dermatologists was 82.6%/ 91.7% and that of the patients were 56.5%/ 42.6%. The sensitivity/specificity derived from the Top-1 diagnosis of the algorithm was 52.2%/ 93.4%. The sensitivity/specificity at the threshold of the risk \u201cMedium\u201d using the malignancy score was 95.7%/ 60.3% and that at the threshold of the risk \u201cHigh\u201d was 82.6%/ 70.2%.\n\nResult of the control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between before and after reviewing photographs were not significant (Control-Before, 46.1%/ 64.5%/ 66.7%; Control-After, 51.8%/ 66.7/ 68.1%; P = .19/ = .42/ = .35). For determining malignancy, the sensitivity/specificity derived from the Top-1 diagnosis was 65.5%/ 81.3% before reviewing and 65.5%/ 86.6% after reviewing (Table 2, P = 1.00/ = .09). The sensitivity/specificity of the attending dermatologists was 79.3%/ 90.2% and that of the patients was 48.1%/ 44.5%.\n\nAI group versus control group\n\nThe differences of the Top-1/Top-2/Top-3 accuracies between the AI group and the Control were not significant (AI Group = 58.3%/ 70.1%/ 71.5%; Control Group = 51.8%/ 66.7%/ 68.1%; P = .27/ = .53/ = .53). Summarized key results were described in S4 Table.",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented decision-making for acral lentiginous melanoma detection using deep convolutional neural networks.",
        "abstract": "## BACKGROUND\nSeveral studies have achieved high-level performance of melanoma detection using convolutional neural networks (CNNs). However, few have described the extent to which the implementation of CNNs improves the diagnostic performance of the physicians.\n## OBJECTIVE\nThis study is aimed at developing a CNN for detecting acral lentiginous melanoma (ALM) and investigating whether its implementation can improve the initial decision for ALM detection made by the physicians.\n## METHODS\nA CNN was trained using 1072 dermoscopic images of acral benign nevi, ALM and intermediate tumours. To investigate whether the implementation of CNN can improve the initial decision for ALM detection, 60 physicians completed a three-stage survey. In Stage I, they were asked for their decisions solely on the basis of dermoscopic images provided to them. In Stage II, they were also provided with clinical information. In Stage III, they were provided with the additional diagnosis and probability predicted by the CNN.\n## RESULTS\nThe accuracy of ALM detection in the participants was 74.7% (95% confidence interval [CI], 72.6-76.8%) in Stage I and 79.0% (95% CI, 76.7-81.2%) in Stage II. In Stage III, it was 86.9% (95% CI, 85.3-88.4%), which exceeds the accuracy delivered in Stage I by 12.2%p (95% CI, 10.1-14.3%p) and Stage II by 7.9%p (95% CI, 6.0-9.9%p). Moreover, the concordance between the participants considerably increased (Fleiss-\u03ba of 0.436 [95% CI, 0.437-0.573] in Stage I, 0.506 [95% CI, 0.621-0.749] in Stage II and 0.684 [95% CI, 0.621-0.749] in Stage III).\n## CONCLUSIONS\nAugmented decision-making improved the performance of and concordance between the clinical decisions of a diverse group of experts. This study demonstrates the potential use of CNNs as an adjoining, decision-supporting system for physicians' decisions.\n",
        "methods": "**Materials and Methods**\n\n\nDataset\nThis retrospective analysis included dermoscopic images of acral pigmentary lesions collected from 2014 to 2019 at the Department of Dermatology, Severance Hospital, Seoul, Korea. As a means of ensuring quality control, from the dermoscopic image repository of our institution, we constructed a balanced dataset of 500 acral benign nevi (BN) images without atypical or dysplastic features and 500 ALM images. An additional 72 images of acral nevi with atypical or dysplastic features in their histopathological findings were separately collected and labelled as intermediate tumours.16 All dermoscopic images were captured using a DermLite Cam\u00ae (3 Gen, California) in a cross polarized or non-polarized mode, and with or without the use of contact media. Although dCNNs generally require extensive amounts of training data, the relatively higher homogeneity of dermoscopic images allowed us to considerably reduce the amount of data required to train the network.17Because the data were acquired from a population-based clinical setting, many of the BN images were provided alongside their respective clinical diagnoses from follow-up examinations. To ensure accurate diagnoses, these were reviewed by three board-certified dermatologists during the dataset-construction stages. The other samples were obtained following the complete excision of acral lesions by us (K.Y.C. and B.O.), enabling histopathologic assessment. The haematoxylin and eosin, and immunohistochemical (if available) stains were reviewed by board-certified pathologists and dermatologists, and the results served as the diagnostic gold standard throughout the study. In total, 800 (80%) randomly selected BN and ALM images, and all 72 intermediate tumour images, were used to train the dCNN. The remaining 200 (20%) were used for validation (test-set 200, n=200). Two datasets were reviewed to ensure lesion-stratified partitioning. Clinical information was obtained by reviewing patients\u2019 electronic medical charts (Table S1). This study was approved by the institutional review board of Yonsei University College of Medicine (Approval No. 4-2019-0282) and was carried out in accordance with the Declaration of Helsinki. A waiver of written informed consent was granted for the de-identified data used.\nDeep Convolutional Neural Network\nWe utilized ResNet with 50 residual layers (Fig. S1)18. Given the characteristics of the acral site images, in which unnecessary background and colour oversaturation were more likely, we preprocessed images and implemented a modified version of the method used in our previous study, for training and inference (Fig. S2).19,20 To investigate an optimal method of training for ALM detection of intermediate tumours, we trained four dCNNs (Fig. S3). To identify influential pixels for tumour detection, class activation maps were generated using global average pooling (Fig. S4).21 The model with the best ALM detection performance was adopted as the final model, and is referred to as ALMnet. All procedures were repeated 5 times to ensure reproducibility and comparability. No overlap between the training and test data was allowed. Details on the training and inferential methods of the dCNN, including image pre-processing and architecture configuration, are described in Supplementary Method in Supplementary Materials.\nDecision Study\nWe recruited a total of 60 physicians (20 board-certified dermatologists, 20 dermatology residents, and 20 general physicians) to investigate how well physicians could make an initial diagnosis of ALM by interpreting dermoscopic images, and how much their performance could be improved with the aid of ALMnet. For ease of response, 100 images from test-set-200 were extracted for the decision study (human-set-100) (Fig. 1). Using Google Survey, an anonymous online-based diagnostic questionnaire was conducted in three stages over 2-week intervals (Fig. S5). In Stage I, participants were shown 100 dermoscopic images only, and asked for their decision in each case (either \u2018It is more likely a benign nevus. Follow-up is needed.\u2019 or \u2018It is more likely a malignant melanoma. Immediate biopsy or intervention is required.\u2019). In Stage II, they were provided with additional clinical information for the same cases and were asked to answer in the same manner. In most previous studies, human evaluators were asked for their diagnoses without being given any clinical information (equivalent to Stage I). However, this does not occur in clinical settings,  which is why we prepared this intermediate stage. In Stage III, they were further provided with the ALMnet diagnosis (either BN or ALM) and the confidence score for each case, and then asked for their answer. The suggested diagnosis was determined by a dichotomous classification (BN versus ALM), produced by interpreting the prediction output with a fixed cut-off (> 0.5 favours ALM). The performance of participants was assessed by comparing their predictions and associated decisions to the diagnostic gold standard. At each stage, the sequences of cases were shuffled to ensure unbiased responses at each stage. The reference diagnosis for each case and participants\u2019 scores were not disclosed until the end of the study.\nOutcome Measurement and Statistical Analysis\nThe area  under curve (AUC) of the receiver operating characteristics (ROC) curve was taken as the primary measurement for comparing the performances of multiple dCNNs. ROC curves were depicted using the prediction output for dichotomous classification of BN and ALM. Two-tailed independent t-tests were used to compare the AUC values, which measured 5 repeats for each of the 4 models. For the model with the highest performance, a repeat with the median AUC was adopted as ALMnet. The primary outcome of the decision study was the difference in accuracy across all human evaluators between Stages I, II, and III. The secondary outcomes include the differences in sensitivity and specificity. For human-set-100, participant accuracies, sensitivities, and specificities were calculated for each stage. Sensitivity was calculated as the proportion of correctly classified ALM images, and specificity was calculated as the proportion of correctly  classified BN images. Repeated measure analysis of variance was used to compare the performance of participants across the physician groups and study stages.The exact Clopper-Pearson method22 was used for calculating the 95% confidence intervals (CIs) of the dCNN performance parameters. Fleiss-\u03ba statistics23 were calculated to assess concordance among participants\u2019 responses. A heatmap was used to visualize inter-participant agreement rate, with clustering via a hierarchical agglomerative method. All statistical analyses were conducted in R version 3.5.0 (R Foundation for Statistical Computing) at a significance level of 5%. Four R packages were used: ROCit for plotting ROC curves,24 IRR25 and raters26 for calculating \u03ba-statistics, and pheatmap27 to produce the heatmap.",
        "results": "**Results**\n\n**Performance of Deep Convolutional Neural Networks**\n\nThe performances of four dCNNs for test-set-200 in ALM detection are summarized in Table S2 and Fig. S6. The ensembled model (Model 4) showed the best performance, with a highest AUC value of 0.976 (95% CI, 0.974\u20130.978), significantly higher than those of the other models. Model 2 in which intermediate tumours were trained as BN showed higher specificity compared to Model 4. However, a loss in the sensitivity resulted in a sub-optimal AUC and accuracy. On the contrary, Model 3 where these intermediate tumours were trained as ALM showed a higher sensitivity compared to Model 4, but the overall AUC and accuracy was sub-optimal owing to the loss of specificity. From the 5 repeats for Model 4, the median performance (AUC of 0.976) was adopted as ALMnet. With a cut-off value of 0.5 for predictive output of ALM classification, ALMnet demonstrated an accuracy of 92.5% (95% CI, 87.9%\u201395.7%), a sensitivity of 90.0% (95% CI, 82.4%\u201395.1%), and a specificity of 95.0% (95% CI, 88.7%\u201398.4%). For human-set-100, it demonstrated an accuracy of 94.0% (95% CI, 87.4%\u201397.8%), a sensitivity of 92.0% (95% CI, 80.8%\u201397.8%), and a specificity of 96.0% (95% CI, 86.3%\u201399.5%).\n\n**Performance of Human Evaluators in Stages I and II**\n\nA total of 60 participants were provided with 100 dermoscopic images (human-set-100) and made initial decisions for each case. The performances of the participants and ALMnet are summarized in Table 1 and Figs 2 and 3. The interrater agreement rates between participant responses, ALMnet predictions, and the diagnostic gold standard are presented in Fig. 4.\n\nIn Stage I, participants showed an accuracy of 74.7% (95% CI, 72.6%\u201376.8%), a sensitivity of 79.9% (95% CI, 76.2%\u201383.5%), and a specificity of 69.5% (95% CI, 65.1%\u201373.8%) (Table 1I). Despite some exceptions, most participants were outperformed by ALMnet (Fig. 2). Moreover, a prominent disagreement (Fig. 4A, Fleiss-\u03ba of 0.436 [95% CI, 0.370\u20130.503]) and performance gap (Fig. S7A) were observed across participants.\n\nIn Stage II, participants were given dermoscopic images along with each patient\u2019s age, sex, and site of lesion. The accuracy, sensitivity, and specificity were 79.0% (95% CI, 76.7%\u201381.2%), 81.5% (95% CI, 77.7%\u201385.2%), and 76.4% (95% CI, 72.5%\u201380.4%), respectively. The differences compared to Stage I were 4.3%p (95% CI, 2.5%p\u20136.1%p), 1.6%p (95% CI, -1.0%p to 4.2%p), and 7.0%p (95% CI, 3.2%p\u201312.4%p), respectively. Although the availability of clinical information increased participants\u2019 overall performances, many were still outperformed by ALMnet. Regarding interrater agreement, we see that despite a slightly improved concordance between participants, a considerable disagreement was still observed (Fig. S7B and Fig. 4B, Fleiss-\u03ba of 0.506 [95% CI, 0.437\u20130.573]).\n\n**Augmented Decision Making of Human Evaluators in Stage III**\n\nIn Stage III, participants were further provided with the diagnosis and its probability, as predicted by ALMnet for each case. Compared to Stage II, 811/6000 (13.5%) of participant responses were changed as a result of augmented decision making, and 643/811 (79.3%) of these resulted in increased accuracy (Fig. S8). The accuracy, sensitivity, and specificity were 86.9% (95% CI, 85.3%\u201388.4%), 88.7% (95% CI, 86.0%\u201391.5%), and 85.0% (95% CI, 82.7%\u201387.3%), respectively. The differences compared to Stage I were 12.2%p (95% CI, 10.1%p\u201314.3%p), 8.9%p (95% CI, 6.1%p\u201311.7%p), and 15.5%p (95% CI, 11.4%p\u201319.7%p), respectively. Similarly, those compared to Stage II were 7.9%p (95% CI, 6.0%p\u20139.9%p), 7.3%p (95% CI, 4.5%p\u201310.0%p), and 8.6%p (95% CI, 5.3%p\u201311.8%p), respectively, demonstrating a significant improvement in participants\u2019 performances. This improvement was particularly emphasized in the relatively inexperienced groups of dermatology residents and general physicians. Moreover, the performance gap across participants was considerably diminished, and the overall agreement across participants greatly improved (Fig. S7C and Fig. 4C, Fleiss-\u03ba of 0.684 [95% CI, 0.621\u20130.749]).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented decision-making for acral lentiginous melanoma detection using deep convolutional neural networks.",
        "abstract": "## BACKGROUND\nSeveral studies have achieved high-level performance of melanoma detection using convolutional neural networks (CNNs). However, few have described the extent to which the implementation of CNNs improves the diagnostic performance of the physicians.\n## OBJECTIVE\nThis study is aimed at developing a CNN for detecting acral lentiginous melanoma (ALM) and investigating whether its implementation can improve the initial decision for ALM detection made by the physicians.\n## METHODS\nA CNN was trained using 1072 dermoscopic images of acral benign nevi, ALM and intermediate tumours. To investigate whether the implementation of CNN can improve the initial decision for ALM detection, 60 physicians completed a three-stage survey. In Stage I, they were asked for their decisions solely on the basis of dermoscopic images provided to them. In Stage II, they were also provided with clinical information. In Stage III, they were provided with the additional diagnosis and probability predicted by the CNN.\n## RESULTS\nThe accuracy of ALM detection in the participants was 74.7% (95% confidence interval [CI], 72.6-76.8%) in Stage I and 79.0% (95% CI, 76.7-81.2%) in Stage II. In Stage III, it was 86.9% (95% CI, 85.3-88.4%), which exceeds the accuracy delivered in Stage I by 12.2%p (95% CI, 10.1-14.3%p) and Stage II by 7.9%p (95% CI, 6.0-9.9%p). Moreover, the concordance between the participants considerably increased (Fleiss-\u03ba of 0.436 [95% CI, 0.437-0.573] in Stage I, 0.506 [95% CI, 0.621-0.749] in Stage II and 0.684 [95% CI, 0.621-0.749] in Stage III).\n## CONCLUSIONS\nAugmented decision-making improved the performance of and concordance between the clinical decisions of a diverse group of experts. This study demonstrates the potential use of CNNs as an adjoining, decision-supporting system for physicians' decisions.\n",
        "methods": "**Materials and Methods**\n\n\nDataset\nThis retrospective analysis included dermoscopic images of acral pigmentary lesions collected from 2014 to 2019 at the Department of Dermatology, Severance Hospital, Seoul, Korea. As a means of ensuring quality control, from the dermoscopic image repository of our institution, we constructed a balanced dataset of 500 acral benign nevi (BN) images without atypical or dysplastic features and 500 ALM images. An additional 72 images of acral nevi with atypical or dysplastic features in their histopathological findings were separately collected and labelled as intermediate tumours.16 All dermoscopic images were captured using a DermLite Cam\u00ae (3 Gen, California) in a cross polarized or non-polarized mode, and with or without the use of contact media. Although dCNNs generally require extensive amounts of training data, the relatively higher homogeneity of dermoscopic images allowed us to considerably reduce the amount of data required to train the network.17Because the data were acquired from a population-based clinical setting, many of the BN images were provided alongside their respective clinical diagnoses from follow-up examinations. To ensure accurate diagnoses, these were reviewed by three board-certified dermatologists during the dataset-construction stages. The other samples were obtained following the complete excision of acral lesions by us (K.Y.C. and B.O.), enabling histopathologic assessment. The haematoxylin and eosin, and immunohistochemical (if available) stains were reviewed by board-certified pathologists and dermatologists, and the results served as the diagnostic gold standard throughout the study. In total, 800 (80%) randomly selected BN and ALM images, and all 72 intermediate tumour images, were used to train the dCNN. The remaining 200 (20%) were used for validation (test-set 200, n=200). Two datasets were reviewed to ensure lesion-stratified partitioning. Clinical information was obtained by reviewing patients\u2019 electronic medical charts (Table S1). This study was approved by the institutional review board of Yonsei University College of Medicine (Approval No. 4-2019-0282) and was carried out in accordance with the Declaration of Helsinki. A waiver of written informed consent was granted for the de-identified data used.\nDeep Convolutional Neural Network\nWe utilized ResNet with 50 residual layers (Fig. S1)18. Given the characteristics of the acral site images, in which unnecessary background and colour oversaturation were more likely, we preprocessed images and implemented a modified version of the method used in our previous study, for training and inference (Fig. S2).19,20 To investigate an optimal method of training for ALM detection of intermediate tumours, we trained four dCNNs (Fig. S3). To identify influential pixels for tumour detection, class activation maps were generated using global average pooling (Fig. S4).21 The model with the best ALM detection performance was adopted as the final model, and is referred to as ALMnet. All procedures were repeated 5 times to ensure reproducibility and comparability. No overlap between the training and test data was allowed. Details on the training and inferential methods of the dCNN, including image pre-processing and architecture configuration, are described in Supplementary Method in Supplementary Materials.\nDecision Study\nWe recruited a total of 60 physicians (20 board-certified dermatologists, 20 dermatology residents, and 20 general physicians) to investigate how well physicians could make an initial diagnosis of ALM by interpreting dermoscopic images, and how much their performance could be improved with the aid of ALMnet. For ease of response, 100 images from test-set-200 were extracted for the decision study (human-set-100) (Fig. 1). Using Google Survey, an anonymous online-based diagnostic questionnaire was conducted in three stages over 2-week intervals (Fig. S5). In Stage I, participants were shown 100 dermoscopic images only, and asked for their decision in each case (either \u2018It is more likely a benign nevus. Follow-up is needed.\u2019 or \u2018It is more likely a malignant melanoma. Immediate biopsy or intervention is required.\u2019). In Stage II, they were provided with additional clinical information for the same cases and were asked to answer in the same manner. In most previous studies, human evaluators were asked for their diagnoses without being given any clinical information (equivalent to Stage I). However, this does not occur in clinical settings,  which is why we prepared this intermediate stage. In Stage III, they were further provided with the ALMnet diagnosis (either BN or ALM) and the confidence score for each case, and then asked for their answer. The suggested diagnosis was determined by a dichotomous classification (BN versus ALM), produced by interpreting the prediction output with a fixed cut-off (> 0.5 favours ALM). The performance of participants was assessed by comparing their predictions and associated decisions to the diagnostic gold standard. At each stage, the sequences of cases were shuffled to ensure unbiased responses at each stage. The reference diagnosis for each case and participants\u2019 scores were not disclosed until the end of the study.\nOutcome Measurement and Statistical Analysis\nThe area  under curve (AUC) of the receiver operating characteristics (ROC) curve was taken as the primary measurement for comparing the performances of multiple dCNNs. ROC curves were depicted using the prediction output for dichotomous classification of BN and ALM. Two-tailed independent t-tests were used to compare the AUC values, which measured 5 repeats for each of the 4 models. For the model with the highest performance, a repeat with the median AUC was adopted as ALMnet. The primary outcome of the decision study was the difference in accuracy across all human evaluators between Stages I, II, and III. The secondary outcomes include the differences in sensitivity and specificity. For human-set-100, participant accuracies, sensitivities, and specificities were calculated for each stage. Sensitivity was calculated as the proportion of correctly classified ALM images, and specificity was calculated as the proportion of correctly  classified BN images. Repeated measure analysis of variance was used to compare the performance of participants across the physician groups and study stages.The exact Clopper-Pearson method22 was used for calculating the 95% confidence intervals (CIs) of the dCNN performance parameters. Fleiss-\u03ba statistics23 were calculated to assess concordance among participants\u2019 responses. A heatmap was used to visualize inter-participant agreement rate, with clustering via a hierarchical agglomerative method. All statistical analyses were conducted in R version 3.5.0 (R Foundation for Statistical Computing) at a significance level of 5%. Four R packages were used: ROCit for plotting ROC curves,24 IRR25 and raters26 for calculating \u03ba-statistics, and pheatmap27 to produce the heatmap.",
        "results": "**Results**\n\n**Performance of Deep Convolutional Neural Networks**\n\nThe performances of four dCNNs for test-set-200 in ALM detection are summarized in Table S2 and Fig. S6. The ensembled model (Model 4) showed the best performance, with a highest AUC value of 0.976 (95% CI, 0.974\u20130.978), significantly higher than those of the other models. Model 2 in which intermediate tumours were trained as BN showed higher specificity compared to Model 4. However, a loss in the sensitivity resulted in a sub-optimal AUC and accuracy. On the contrary, Model 3 where these intermediate tumours were trained as ALM showed a higher sensitivity compared to Model 4, but the overall AUC and accuracy was sub-optimal owing to the loss of specificity. From the 5 repeats for Model 4, the median performance (AUC of 0.976) was adopted as ALMnet. With a cut-off value of 0.5 for predictive output of ALM classification, ALMnet demonstrated an accuracy of 92.5% (95% CI, 87.9%\u201395.7%), a sensitivity of 90.0% (95% CI, 82.4%\u201395.1%), and a specificity of 95.0% (95% CI, 88.7%\u201398.4%). For human-set-100, it demonstrated an accuracy of 94.0% (95% CI, 87.4%\u201397.8%), a sensitivity of 92.0% (95% CI, 80.8%\u201397.8%), and a specificity of 96.0% (95% CI, 86.3%\u201399.5%).\n\n**Performance of Human Evaluators in Stages I and II**\n\nA total of 60 participants were provided with 100 dermoscopic images (human-set-100) and made initial decisions for each case. The performances of the participants and ALMnet are summarized in Table 1 and Figs 2 and 3. The interrater agreement rates between participant responses, ALMnet predictions, and the diagnostic gold standard are presented in Fig. 4.\n\nIn Stage I, participants showed an accuracy of 74.7% (95% CI, 72.6%\u201376.8%), a sensitivity of 79.9% (95% CI, 76.2%\u201383.5%), and a specificity of 69.5% (95% CI, 65.1%\u201373.8%) (Table 1I). Despite some exceptions, most participants were outperformed by ALMnet (Fig. 2). Moreover, a prominent disagreement (Fig. 4A, Fleiss-\u03ba of 0.436 [95% CI, 0.370\u20130.503]) and performance gap (Fig. S7A) were observed across participants.\n\nIn Stage II, participants were given dermoscopic images along with each patient\u2019s age, sex, and site of lesion. The accuracy, sensitivity, and specificity were 79.0% (95% CI, 76.7%\u201381.2%), 81.5% (95% CI, 77.7%\u201385.2%), and 76.4% (95% CI, 72.5%\u201380.4%), respectively. The differences compared to Stage I were 4.3%p (95% CI, 2.5%p\u20136.1%p), 1.6%p (95% CI, -1.0%p to 4.2%p), and 7.0%p (95% CI, 3.2%p\u201312.4%p), respectively. Although the availability of clinical information increased participants\u2019 overall performances, many were still outperformed by ALMnet. Regarding interrater agreement, we see that despite a slightly improved concordance between participants, a considerable disagreement was still observed (Fig. S7B and Fig. 4B, Fleiss-\u03ba of 0.506 [95% CI, 0.437\u20130.573]).\n\n**Augmented Decision Making of Human Evaluators in Stage III**\n\nIn Stage III, participants were further provided with the diagnosis and its probability, as predicted by ALMnet for each case. Compared to Stage II, 811/6000 (13.5%) of participant responses were changed as a result of augmented decision making, and 643/811 (79.3%) of these resulted in increased accuracy (Fig. S8). The accuracy, sensitivity, and specificity were 86.9% (95% CI, 85.3%\u201388.4%), 88.7% (95% CI, 86.0%\u201391.5%), and 85.0% (95% CI, 82.7%\u201387.3%), respectively. The differences compared to Stage I were 12.2%p (95% CI, 10.1%p\u201314.3%p), 8.9%p (95% CI, 6.1%p\u201311.7%p), and 15.5%p (95% CI, 11.4%p\u201319.7%p), respectively. Similarly, those compared to Stage II were 7.9%p (95% CI, 6.0%p\u20139.9%p), 7.3%p (95% CI, 4.5%p\u201310.0%p), and 8.6%p (95% CI, 5.3%p\u201311.8%p), respectively, demonstrating a significant improvement in participants\u2019 performances. This improvement was particularly emphasized in the relatively inexperienced groups of dermatology residents and general physicians. Moreover, the performance gap across participants was considerably diminished, and the overall agreement across participants greatly improved (Fig. S7C and Fig. 4C, Fleiss-\u03ba of 0.684 [95% CI, 0.621\u20130.749]).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented decision-making for acral lentiginous melanoma detection using deep convolutional neural networks.",
        "abstract": "## BACKGROUND\nSeveral studies have achieved high-level performance of melanoma detection using convolutional neural networks (CNNs). However, few have described the extent to which the implementation of CNNs improves the diagnostic performance of the physicians.\n## OBJECTIVE\nThis study is aimed at developing a CNN for detecting acral lentiginous melanoma (ALM) and investigating whether its implementation can improve the initial decision for ALM detection made by the physicians.\n## METHODS\nA CNN was trained using 1072 dermoscopic images of acral benign nevi, ALM and intermediate tumours. To investigate whether the implementation of CNN can improve the initial decision for ALM detection, 60 physicians completed a three-stage survey. In Stage I, they were asked for their decisions solely on the basis of dermoscopic images provided to them. In Stage II, they were also provided with clinical information. In Stage III, they were provided with the additional diagnosis and probability predicted by the CNN.\n## RESULTS\nThe accuracy of ALM detection in the participants was 74.7% (95% confidence interval [CI], 72.6-76.8%) in Stage I and 79.0% (95% CI, 76.7-81.2%) in Stage II. In Stage III, it was 86.9% (95% CI, 85.3-88.4%), which exceeds the accuracy delivered in Stage I by 12.2%p (95% CI, 10.1-14.3%p) and Stage II by 7.9%p (95% CI, 6.0-9.9%p). Moreover, the concordance between the participants considerably increased (Fleiss-\u03ba of 0.436 [95% CI, 0.437-0.573] in Stage I, 0.506 [95% CI, 0.621-0.749] in Stage II and 0.684 [95% CI, 0.621-0.749] in Stage III).\n## CONCLUSIONS\nAugmented decision-making improved the performance of and concordance between the clinical decisions of a diverse group of experts. This study demonstrates the potential use of CNNs as an adjoining, decision-supporting system for physicians' decisions.\n",
        "methods": "**Materials and Methods**\n\n\nDataset\nThis retrospective analysis included dermoscopic images of acral pigmentary lesions collected from 2014 to 2019 at the Department of Dermatology, Severance Hospital, Seoul, Korea. As a means of ensuring quality control, from the dermoscopic image repository of our institution, we constructed a balanced dataset of 500 acral benign nevi (BN) images without atypical or dysplastic features and 500 ALM images. An additional 72 images of acral nevi with atypical or dysplastic features in their histopathological findings were separately collected and labelled as intermediate tumours.16 All dermoscopic images were captured using a DermLite Cam\u00ae (3 Gen, California) in a cross polarized or non-polarized mode, and with or without the use of contact media. Although dCNNs generally require extensive amounts of training data, the relatively higher homogeneity of dermoscopic images allowed us to considerably reduce the amount of data required to train the network.17Because the data were acquired from a population-based clinical setting, many of the BN images were provided alongside their respective clinical diagnoses from follow-up examinations. To ensure accurate diagnoses, these were reviewed by three board-certified dermatologists during the dataset-construction stages. The other samples were obtained following the complete excision of acral lesions by us (K.Y.C. and B.O.), enabling histopathologic assessment. The haematoxylin and eosin, and immunohistochemical (if available) stains were reviewed by board-certified pathologists and dermatologists, and the results served as the diagnostic gold standard throughout the study. In total, 800 (80%) randomly selected BN and ALM images, and all 72 intermediate tumour images, were used to train the dCNN. The remaining 200 (20%) were used for validation (test-set 200, n=200). Two datasets were reviewed to ensure lesion-stratified partitioning. Clinical information was obtained by reviewing patients\u2019 electronic medical charts (Table S1). This study was approved by the institutional review board of Yonsei University College of Medicine (Approval No. 4-2019-0282) and was carried out in accordance with the Declaration of Helsinki. A waiver of written informed consent was granted for the de-identified data used.\nDeep Convolutional Neural Network\nWe utilized ResNet with 50 residual layers (Fig. S1)18. Given the characteristics of the acral site images, in which unnecessary background and colour oversaturation were more likely, we preprocessed images and implemented a modified version of the method used in our previous study, for training and inference (Fig. S2).19,20 To investigate an optimal method of training for ALM detection of intermediate tumours, we trained four dCNNs (Fig. S3). To identify influential pixels for tumour detection, class activation maps were generated using global average pooling (Fig. S4).21 The model with the best ALM detection performance was adopted as the final model, and is referred to as ALMnet. All procedures were repeated 5 times to ensure reproducibility and comparability. No overlap between the training and test data was allowed. Details on the training and inferential methods of the dCNN, including image pre-processing and architecture configuration, are described in Supplementary Method in Supplementary Materials.\nDecision Study\nWe recruited a total of 60 physicians (20 board-certified dermatologists, 20 dermatology residents, and 20 general physicians) to investigate how well physicians could make an initial diagnosis of ALM by interpreting dermoscopic images, and how much their performance could be improved with the aid of ALMnet. For ease of response, 100 images from test-set-200 were extracted for the decision study (human-set-100) (Fig. 1). Using Google Survey, an anonymous online-based diagnostic questionnaire was conducted in three stages over 2-week intervals (Fig. S5). In Stage I, participants were shown 100 dermoscopic images only, and asked for their decision in each case (either \u2018It is more likely a benign nevus. Follow-up is needed.\u2019 or \u2018It is more likely a malignant melanoma. Immediate biopsy or intervention is required.\u2019). In Stage II, they were provided with additional clinical information for the same cases and were asked to answer in the same manner. In most previous studies, human evaluators were asked for their diagnoses without being given any clinical information (equivalent to Stage I). However, this does not occur in clinical settings,  which is why we prepared this intermediate stage. In Stage III, they were further provided with the ALMnet diagnosis (either BN or ALM) and the confidence score for each case, and then asked for their answer. The suggested diagnosis was determined by a dichotomous classification (BN versus ALM), produced by interpreting the prediction output with a fixed cut-off (> 0.5 favours ALM). The performance of participants was assessed by comparing their predictions and associated decisions to the diagnostic gold standard. At each stage, the sequences of cases were shuffled to ensure unbiased responses at each stage. The reference diagnosis for each case and participants\u2019 scores were not disclosed until the end of the study.\nOutcome Measurement and Statistical Analysis\nThe area  under curve (AUC) of the receiver operating characteristics (ROC) curve was taken as the primary measurement for comparing the performances of multiple dCNNs. ROC curves were depicted using the prediction output for dichotomous classification of BN and ALM. Two-tailed independent t-tests were used to compare the AUC values, which measured 5 repeats for each of the 4 models. For the model with the highest performance, a repeat with the median AUC was adopted as ALMnet. The primary outcome of the decision study was the difference in accuracy across all human evaluators between Stages I, II, and III. The secondary outcomes include the differences in sensitivity and specificity. For human-set-100, participant accuracies, sensitivities, and specificities were calculated for each stage. Sensitivity was calculated as the proportion of correctly classified ALM images, and specificity was calculated as the proportion of correctly  classified BN images. Repeated measure analysis of variance was used to compare the performance of participants across the physician groups and study stages.The exact Clopper-Pearson method22 was used for calculating the 95% confidence intervals (CIs) of the dCNN performance parameters. Fleiss-\u03ba statistics23 were calculated to assess concordance among participants\u2019 responses. A heatmap was used to visualize inter-participant agreement rate, with clustering via a hierarchical agglomerative method. All statistical analyses were conducted in R version 3.5.0 (R Foundation for Statistical Computing) at a significance level of 5%. Four R packages were used: ROCit for plotting ROC curves,24 IRR25 and raters26 for calculating \u03ba-statistics, and pheatmap27 to produce the heatmap.",
        "results": "**Results**\n\n**Performance of Deep Convolutional Neural Networks**\n\nThe performances of four dCNNs for test-set-200 in ALM detection are summarized in Table S2 and Fig. S6. The ensembled model (Model 4) showed the best performance, with a highest AUC value of 0.976 (95% CI, 0.974\u20130.978), significantly higher than those of the other models. Model 2 in which intermediate tumours were trained as BN showed higher specificity compared to Model 4. However, a loss in the sensitivity resulted in a sub-optimal AUC and accuracy. On the contrary, Model 3 where these intermediate tumours were trained as ALM showed a higher sensitivity compared to Model 4, but the overall AUC and accuracy was sub-optimal owing to the loss of specificity. From the 5 repeats for Model 4, the median performance (AUC of 0.976) was adopted as ALMnet. With a cut-off value of 0.5 for predictive output of ALM classification, ALMnet demonstrated an accuracy of 92.5% (95% CI, 87.9%\u201395.7%), a sensitivity of 90.0% (95% CI, 82.4%\u201395.1%), and a specificity of 95.0% (95% CI, 88.7%\u201398.4%). For human-set-100, it demonstrated an accuracy of 94.0% (95% CI, 87.4%\u201397.8%), a sensitivity of 92.0% (95% CI, 80.8%\u201397.8%), and a specificity of 96.0% (95% CI, 86.3%\u201399.5%).\n\n**Performance of Human Evaluators in Stages I and II**\n\nA total of 60 participants were provided with 100 dermoscopic images (human-set-100) and made initial decisions for each case. The performances of the participants and ALMnet are summarized in Table 1 and Figs 2 and 3. The interrater agreement rates between participant responses, ALMnet predictions, and the diagnostic gold standard are presented in Fig. 4.\n\nIn Stage I, participants showed an accuracy of 74.7% (95% CI, 72.6%\u201376.8%), a sensitivity of 79.9% (95% CI, 76.2%\u201383.5%), and a specificity of 69.5% (95% CI, 65.1%\u201373.8%) (Table 1I). Despite some exceptions, most participants were outperformed by ALMnet (Fig. 2). Moreover, a prominent disagreement (Fig. 4A, Fleiss-\u03ba of 0.436 [95% CI, 0.370\u20130.503]) and performance gap (Fig. S7A) were observed across participants.\n\nIn Stage II, participants were given dermoscopic images along with each patient\u2019s age, sex, and site of lesion. The accuracy, sensitivity, and specificity were 79.0% (95% CI, 76.7%\u201381.2%), 81.5% (95% CI, 77.7%\u201385.2%), and 76.4% (95% CI, 72.5%\u201380.4%), respectively. The differences compared to Stage I were 4.3%p (95% CI, 2.5%p\u20136.1%p), 1.6%p (95% CI, -1.0%p to 4.2%p), and 7.0%p (95% CI, 3.2%p\u201312.4%p), respectively. Although the availability of clinical information increased participants\u2019 overall performances, many were still outperformed by ALMnet. Regarding interrater agreement, we see that despite a slightly improved concordance between participants, a considerable disagreement was still observed (Fig. S7B and Fig. 4B, Fleiss-\u03ba of 0.506 [95% CI, 0.437\u20130.573]).\n\n**Augmented Decision Making of Human Evaluators in Stage III**\n\nIn Stage III, participants were further provided with the diagnosis and its probability, as predicted by ALMnet for each case. Compared to Stage II, 811/6000 (13.5%) of participant responses were changed as a result of augmented decision making, and 643/811 (79.3%) of these resulted in increased accuracy (Fig. S8). The accuracy, sensitivity, and specificity were 86.9% (95% CI, 85.3%\u201388.4%), 88.7% (95% CI, 86.0%\u201391.5%), and 85.0% (95% CI, 82.7%\u201387.3%), respectively. The differences compared to Stage I were 12.2%p (95% CI, 10.1%p\u201314.3%p), 8.9%p (95% CI, 6.1%p\u201311.7%p), and 15.5%p (95% CI, 11.4%p\u201319.7%p), respectively. Similarly, those compared to Stage II were 7.9%p (95% CI, 6.0%p\u20139.9%p), 7.3%p (95% CI, 4.5%p\u201310.0%p), and 8.6%p (95% CI, 5.3%p\u201311.8%p), respectively, demonstrating a significant improvement in participants\u2019 performances. This improvement was particularly emphasized in the relatively inexperienced groups of dermatology residents and general physicians. Moreover, the performance gap across participants was considerably diminished, and the overall agreement across participants greatly improved (Fig. S7C and Fig. 4C, Fleiss-\u03ba of 0.684 [95% CI, 0.621\u20130.749]).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented decision-making for acral lentiginous melanoma detection using deep convolutional neural networks.",
        "abstract": "## BACKGROUND\nSeveral studies have achieved high-level performance of melanoma detection using convolutional neural networks (CNNs). However, few have described the extent to which the implementation of CNNs improves the diagnostic performance of the physicians.\n## OBJECTIVE\nThis study is aimed at developing a CNN for detecting acral lentiginous melanoma (ALM) and investigating whether its implementation can improve the initial decision for ALM detection made by the physicians.\n## METHODS\nA CNN was trained using 1072 dermoscopic images of acral benign nevi, ALM and intermediate tumours. To investigate whether the implementation of CNN can improve the initial decision for ALM detection, 60 physicians completed a three-stage survey. In Stage I, they were asked for their decisions solely on the basis of dermoscopic images provided to them. In Stage II, they were also provided with clinical information. In Stage III, they were provided with the additional diagnosis and probability predicted by the CNN.\n## RESULTS\nThe accuracy of ALM detection in the participants was 74.7% (95% confidence interval [CI], 72.6-76.8%) in Stage I and 79.0% (95% CI, 76.7-81.2%) in Stage II. In Stage III, it was 86.9% (95% CI, 85.3-88.4%), which exceeds the accuracy delivered in Stage I by 12.2%p (95% CI, 10.1-14.3%p) and Stage II by 7.9%p (95% CI, 6.0-9.9%p). Moreover, the concordance between the participants considerably increased (Fleiss-\u03ba of 0.436 [95% CI, 0.437-0.573] in Stage I, 0.506 [95% CI, 0.621-0.749] in Stage II and 0.684 [95% CI, 0.621-0.749] in Stage III).\n## CONCLUSIONS\nAugmented decision-making improved the performance of and concordance between the clinical decisions of a diverse group of experts. This study demonstrates the potential use of CNNs as an adjoining, decision-supporting system for physicians' decisions.\n",
        "methods": "**Materials and Methods**\n\n\nDataset\nThis retrospective analysis included dermoscopic images of acral pigmentary lesions collected from 2014 to 2019 at the Department of Dermatology, Severance Hospital, Seoul, Korea. As a means of ensuring quality control, from the dermoscopic image repository of our institution, we constructed a balanced dataset of 500 acral benign nevi (BN) images without atypical or dysplastic features and 500 ALM images. An additional 72 images of acral nevi with atypical or dysplastic features in their histopathological findings were separately collected and labelled as intermediate tumours.16 All dermoscopic images were captured using a DermLite Cam\u00ae (3 Gen, California) in a cross polarized or non-polarized mode, and with or without the use of contact media. Although dCNNs generally require extensive amounts of training data, the relatively higher homogeneity of dermoscopic images allowed us to considerably reduce the amount of data required to train the network.17Because the data were acquired from a population-based clinical setting, many of the BN images were provided alongside their respective clinical diagnoses from follow-up examinations. To ensure accurate diagnoses, these were reviewed by three board-certified dermatologists during the dataset-construction stages. The other samples were obtained following the complete excision of acral lesions by us (K.Y.C. and B.O.), enabling histopathologic assessment. The haematoxylin and eosin, and immunohistochemical (if available) stains were reviewed by board-certified pathologists and dermatologists, and the results served as the diagnostic gold standard throughout the study. In total, 800 (80%) randomly selected BN and ALM images, and all 72 intermediate tumour images, were used to train the dCNN. The remaining 200 (20%) were used for validation (test-set 200, n=200). Two datasets were reviewed to ensure lesion-stratified partitioning. Clinical information was obtained by reviewing patients\u2019 electronic medical charts (Table S1). This study was approved by the institutional review board of Yonsei University College of Medicine (Approval No. 4-2019-0282) and was carried out in accordance with the Declaration of Helsinki. A waiver of written informed consent was granted for the de-identified data used.\nDeep Convolutional Neural Network\nWe utilized ResNet with 50 residual layers (Fig. S1)18. Given the characteristics of the acral site images, in which unnecessary background and colour oversaturation were more likely, we preprocessed images and implemented a modified version of the method used in our previous study, for training and inference (Fig. S2).19,20 To investigate an optimal method of training for ALM detection of intermediate tumours, we trained four dCNNs (Fig. S3). To identify influential pixels for tumour detection, class activation maps were generated using global average pooling (Fig. S4).21 The model with the best ALM detection performance was adopted as the final model, and is referred to as ALMnet. All procedures were repeated 5 times to ensure reproducibility and comparability. No overlap between the training and test data was allowed. Details on the training and inferential methods of the dCNN, including image pre-processing and architecture configuration, are described in Supplementary Method in Supplementary Materials.\nDecision Study\nWe recruited a total of 60 physicians (20 board-certified dermatologists, 20 dermatology residents, and 20 general physicians) to investigate how well physicians could make an initial diagnosis of ALM by interpreting dermoscopic images, and how much their performance could be improved with the aid of ALMnet. For ease of response, 100 images from test-set-200 were extracted for the decision study (human-set-100) (Fig. 1). Using Google Survey, an anonymous online-based diagnostic questionnaire was conducted in three stages over 2-week intervals (Fig. S5). In Stage I, participants were shown 100 dermoscopic images only, and asked for their decision in each case (either \u2018It is more likely a benign nevus. Follow-up is needed.\u2019 or \u2018It is more likely a malignant melanoma. Immediate biopsy or intervention is required.\u2019). In Stage II, they were provided with additional clinical information for the same cases and were asked to answer in the same manner. In most previous studies, human evaluators were asked for their diagnoses without being given any clinical information (equivalent to Stage I). However, this does not occur in clinical settings,  which is why we prepared this intermediate stage. In Stage III, they were further provided with the ALMnet diagnosis (either BN or ALM) and the confidence score for each case, and then asked for their answer. The suggested diagnosis was determined by a dichotomous classification (BN versus ALM), produced by interpreting the prediction output with a fixed cut-off (> 0.5 favours ALM). The performance of participants was assessed by comparing their predictions and associated decisions to the diagnostic gold standard. At each stage, the sequences of cases were shuffled to ensure unbiased responses at each stage. The reference diagnosis for each case and participants\u2019 scores were not disclosed until the end of the study.\nOutcome Measurement and Statistical Analysis\nThe area  under curve (AUC) of the receiver operating characteristics (ROC) curve was taken as the primary measurement for comparing the performances of multiple dCNNs. ROC curves were depicted using the prediction output for dichotomous classification of BN and ALM. Two-tailed independent t-tests were used to compare the AUC values, which measured 5 repeats for each of the 4 models. For the model with the highest performance, a repeat with the median AUC was adopted as ALMnet. The primary outcome of the decision study was the difference in accuracy across all human evaluators between Stages I, II, and III. The secondary outcomes include the differences in sensitivity and specificity. For human-set-100, participant accuracies, sensitivities, and specificities were calculated for each stage. Sensitivity was calculated as the proportion of correctly classified ALM images, and specificity was calculated as the proportion of correctly  classified BN images. Repeated measure analysis of variance was used to compare the performance of participants across the physician groups and study stages.The exact Clopper-Pearson method22 was used for calculating the 95% confidence intervals (CIs) of the dCNN performance parameters. Fleiss-\u03ba statistics23 were calculated to assess concordance among participants\u2019 responses. A heatmap was used to visualize inter-participant agreement rate, with clustering via a hierarchical agglomerative method. All statistical analyses were conducted in R version 3.5.0 (R Foundation for Statistical Computing) at a significance level of 5%. Four R packages were used: ROCit for plotting ROC curves,24 IRR25 and raters26 for calculating \u03ba-statistics, and pheatmap27 to produce the heatmap.",
        "results": "**Results**\n\n**Performance of Deep Convolutional Neural Networks**\n\nThe performances of four dCNNs for test-set-200 in ALM detection are summarized in Table S2 and Fig. S6. The ensembled model (Model 4) showed the best performance, with a highest AUC value of 0.976 (95% CI, 0.974\u20130.978), significantly higher than those of the other models. Model 2 in which intermediate tumours were trained as BN showed higher specificity compared to Model 4. However, a loss in the sensitivity resulted in a sub-optimal AUC and accuracy. On the contrary, Model 3 where these intermediate tumours were trained as ALM showed a higher sensitivity compared to Model 4, but the overall AUC and accuracy was sub-optimal owing to the loss of specificity. From the 5 repeats for Model 4, the median performance (AUC of 0.976) was adopted as ALMnet. With a cut-off value of 0.5 for predictive output of ALM classification, ALMnet demonstrated an accuracy of 92.5% (95% CI, 87.9%\u201395.7%), a sensitivity of 90.0% (95% CI, 82.4%\u201395.1%), and a specificity of 95.0% (95% CI, 88.7%\u201398.4%). For human-set-100, it demonstrated an accuracy of 94.0% (95% CI, 87.4%\u201397.8%), a sensitivity of 92.0% (95% CI, 80.8%\u201397.8%), and a specificity of 96.0% (95% CI, 86.3%\u201399.5%).\n\n**Performance of Human Evaluators in Stages I and II**\n\nA total of 60 participants were provided with 100 dermoscopic images (human-set-100) and made initial decisions for each case. The performances of the participants and ALMnet are summarized in Table 1 and Figs 2 and 3. The interrater agreement rates between participant responses, ALMnet predictions, and the diagnostic gold standard are presented in Fig. 4.\n\nIn Stage I, participants showed an accuracy of 74.7% (95% CI, 72.6%\u201376.8%), a sensitivity of 79.9% (95% CI, 76.2%\u201383.5%), and a specificity of 69.5% (95% CI, 65.1%\u201373.8%) (Table 1I). Despite some exceptions, most participants were outperformed by ALMnet (Fig. 2). Moreover, a prominent disagreement (Fig. 4A, Fleiss-\u03ba of 0.436 [95% CI, 0.370\u20130.503]) and performance gap (Fig. S7A) were observed across participants.\n\nIn Stage II, participants were given dermoscopic images along with each patient\u2019s age, sex, and site of lesion. The accuracy, sensitivity, and specificity were 79.0% (95% CI, 76.7%\u201381.2%), 81.5% (95% CI, 77.7%\u201385.2%), and 76.4% (95% CI, 72.5%\u201380.4%), respectively. The differences compared to Stage I were 4.3%p (95% CI, 2.5%p\u20136.1%p), 1.6%p (95% CI, -1.0%p to 4.2%p), and 7.0%p (95% CI, 3.2%p\u201312.4%p), respectively. Although the availability of clinical information increased participants\u2019 overall performances, many were still outperformed by ALMnet. Regarding interrater agreement, we see that despite a slightly improved concordance between participants, a considerable disagreement was still observed (Fig. S7B and Fig. 4B, Fleiss-\u03ba of 0.506 [95% CI, 0.437\u20130.573]).\n\n**Augmented Decision Making of Human Evaluators in Stage III**\n\nIn Stage III, participants were further provided with the diagnosis and its probability, as predicted by ALMnet for each case. Compared to Stage II, 811/6000 (13.5%) of participant responses were changed as a result of augmented decision making, and 643/811 (79.3%) of these resulted in increased accuracy (Fig. S8). The accuracy, sensitivity, and specificity were 86.9% (95% CI, 85.3%\u201388.4%), 88.7% (95% CI, 86.0%\u201391.5%), and 85.0% (95% CI, 82.7%\u201387.3%), respectively. The differences compared to Stage I were 12.2%p (95% CI, 10.1%p\u201314.3%p), 8.9%p (95% CI, 6.1%p\u201311.7%p), and 15.5%p (95% CI, 11.4%p\u201319.7%p), respectively. Similarly, those compared to Stage II were 7.9%p (95% CI, 6.0%p\u20139.9%p), 7.3%p (95% CI, 4.5%p\u201310.0%p), and 8.6%p (95% CI, 5.3%p\u201311.8%p), respectively, demonstrating a significant improvement in participants\u2019 performances. This improvement was particularly emphasized in the relatively inexperienced groups of dermatology residents and general physicians. Moreover, the performance gap across participants was considerably diminished, and the overall agreement across participants greatly improved (Fig. S7C and Fig. 4C, Fleiss-\u03ba of 0.684 [95% CI, 0.621\u20130.749]).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented decision-making for acral lentiginous melanoma detection using deep convolutional neural networks.",
        "abstract": "## BACKGROUND\nSeveral studies have achieved high-level performance of melanoma detection using convolutional neural networks (CNNs). However, few have described the extent to which the implementation of CNNs improves the diagnostic performance of the physicians.\n## OBJECTIVE\nThis study is aimed at developing a CNN for detecting acral lentiginous melanoma (ALM) and investigating whether its implementation can improve the initial decision for ALM detection made by the physicians.\n## METHODS\nA CNN was trained using 1072 dermoscopic images of acral benign nevi, ALM and intermediate tumours. To investigate whether the implementation of CNN can improve the initial decision for ALM detection, 60 physicians completed a three-stage survey. In Stage I, they were asked for their decisions solely on the basis of dermoscopic images provided to them. In Stage II, they were also provided with clinical information. In Stage III, they were provided with the additional diagnosis and probability predicted by the CNN.\n## RESULTS\nThe accuracy of ALM detection in the participants was 74.7% (95% confidence interval [CI], 72.6-76.8%) in Stage I and 79.0% (95% CI, 76.7-81.2%) in Stage II. In Stage III, it was 86.9% (95% CI, 85.3-88.4%), which exceeds the accuracy delivered in Stage I by 12.2%p (95% CI, 10.1-14.3%p) and Stage II by 7.9%p (95% CI, 6.0-9.9%p). Moreover, the concordance between the participants considerably increased (Fleiss-\u03ba of 0.436 [95% CI, 0.437-0.573] in Stage I, 0.506 [95% CI, 0.621-0.749] in Stage II and 0.684 [95% CI, 0.621-0.749] in Stage III).\n## CONCLUSIONS\nAugmented decision-making improved the performance of and concordance between the clinical decisions of a diverse group of experts. This study demonstrates the potential use of CNNs as an adjoining, decision-supporting system for physicians' decisions.\n",
        "methods": "**Materials and Methods**\n\n\nDataset\nThis retrospective analysis included dermoscopic images of acral pigmentary lesions collected from 2014 to 2019 at the Department of Dermatology, Severance Hospital, Seoul, Korea. As a means of ensuring quality control, from the dermoscopic image repository of our institution, we constructed a balanced dataset of 500 acral benign nevi (BN) images without atypical or dysplastic features and 500 ALM images. An additional 72 images of acral nevi with atypical or dysplastic features in their histopathological findings were separately collected and labelled as intermediate tumours.16 All dermoscopic images were captured using a DermLite Cam\u00ae (3 Gen, California) in a cross polarized or non-polarized mode, and with or without the use of contact media. Although dCNNs generally require extensive amounts of training data, the relatively higher homogeneity of dermoscopic images allowed us to considerably reduce the amount of data required to train the network.17Because the data were acquired from a population-based clinical setting, many of the BN images were provided alongside their respective clinical diagnoses from follow-up examinations. To ensure accurate diagnoses, these were reviewed by three board-certified dermatologists during the dataset-construction stages. The other samples were obtained following the complete excision of acral lesions by us (K.Y.C. and B.O.), enabling histopathologic assessment. The haematoxylin and eosin, and immunohistochemical (if available) stains were reviewed by board-certified pathologists and dermatologists, and the results served as the diagnostic gold standard throughout the study. In total, 800 (80%) randomly selected BN and ALM images, and all 72 intermediate tumour images, were used to train the dCNN. The remaining 200 (20%) were used for validation (test-set 200, n=200). Two datasets were reviewed to ensure lesion-stratified partitioning. Clinical information was obtained by reviewing patients\u2019 electronic medical charts (Table S1). This study was approved by the institutional review board of Yonsei University College of Medicine (Approval No. 4-2019-0282) and was carried out in accordance with the Declaration of Helsinki. A waiver of written informed consent was granted for the de-identified data used.\nDeep Convolutional Neural Network\nWe utilized ResNet with 50 residual layers (Fig. S1)18. Given the characteristics of the acral site images, in which unnecessary background and colour oversaturation were more likely, we preprocessed images and implemented a modified version of the method used in our previous study, for training and inference (Fig. S2).19,20 To investigate an optimal method of training for ALM detection of intermediate tumours, we trained four dCNNs (Fig. S3). To identify influential pixels for tumour detection, class activation maps were generated using global average pooling (Fig. S4).21 The model with the best ALM detection performance was adopted as the final model, and is referred to as ALMnet. All procedures were repeated 5 times to ensure reproducibility and comparability. No overlap between the training and test data was allowed. Details on the training and inferential methods of the dCNN, including image pre-processing and architecture configuration, are described in Supplementary Method in Supplementary Materials.\nDecision Study\nWe recruited a total of 60 physicians (20 board-certified dermatologists, 20 dermatology residents, and 20 general physicians) to investigate how well physicians could make an initial diagnosis of ALM by interpreting dermoscopic images, and how much their performance could be improved with the aid of ALMnet. For ease of response, 100 images from test-set-200 were extracted for the decision study (human-set-100) (Fig. 1). Using Google Survey, an anonymous online-based diagnostic questionnaire was conducted in three stages over 2-week intervals (Fig. S5). In Stage I, participants were shown 100 dermoscopic images only, and asked for their decision in each case (either \u2018It is more likely a benign nevus. Follow-up is needed.\u2019 or \u2018It is more likely a malignant melanoma. Immediate biopsy or intervention is required.\u2019). In Stage II, they were provided with additional clinical information for the same cases and were asked to answer in the same manner. In most previous studies, human evaluators were asked for their diagnoses without being given any clinical information (equivalent to Stage I). However, this does not occur in clinical settings,  which is why we prepared this intermediate stage. In Stage III, they were further provided with the ALMnet diagnosis (either BN or ALM) and the confidence score for each case, and then asked for their answer. The suggested diagnosis was determined by a dichotomous classification (BN versus ALM), produced by interpreting the prediction output with a fixed cut-off (> 0.5 favours ALM). The performance of participants was assessed by comparing their predictions and associated decisions to the diagnostic gold standard. At each stage, the sequences of cases were shuffled to ensure unbiased responses at each stage. The reference diagnosis for each case and participants\u2019 scores were not disclosed until the end of the study.\nOutcome Measurement and Statistical Analysis\nThe area  under curve (AUC) of the receiver operating characteristics (ROC) curve was taken as the primary measurement for comparing the performances of multiple dCNNs. ROC curves were depicted using the prediction output for dichotomous classification of BN and ALM. Two-tailed independent t-tests were used to compare the AUC values, which measured 5 repeats for each of the 4 models. For the model with the highest performance, a repeat with the median AUC was adopted as ALMnet. The primary outcome of the decision study was the difference in accuracy across all human evaluators between Stages I, II, and III. The secondary outcomes include the differences in sensitivity and specificity. For human-set-100, participant accuracies, sensitivities, and specificities were calculated for each stage. Sensitivity was calculated as the proportion of correctly classified ALM images, and specificity was calculated as the proportion of correctly  classified BN images. Repeated measure analysis of variance was used to compare the performance of participants across the physician groups and study stages.The exact Clopper-Pearson method22 was used for calculating the 95% confidence intervals (CIs) of the dCNN performance parameters. Fleiss-\u03ba statistics23 were calculated to assess concordance among participants\u2019 responses. A heatmap was used to visualize inter-participant agreement rate, with clustering via a hierarchical agglomerative method. All statistical analyses were conducted in R version 3.5.0 (R Foundation for Statistical Computing) at a significance level of 5%. Four R packages were used: ROCit for plotting ROC curves,24 IRR25 and raters26 for calculating \u03ba-statistics, and pheatmap27 to produce the heatmap.",
        "results": "**Results**\n\n**Performance of Deep Convolutional Neural Networks**\n\nThe performances of four dCNNs for test-set-200 in ALM detection are summarized in Table S2 and Fig. S6. The ensembled model (Model 4) showed the best performance, with a highest AUC value of 0.976 (95% CI, 0.974\u20130.978), significantly higher than those of the other models. Model 2 in which intermediate tumours were trained as BN showed higher specificity compared to Model 4. However, a loss in the sensitivity resulted in a sub-optimal AUC and accuracy. On the contrary, Model 3 where these intermediate tumours were trained as ALM showed a higher sensitivity compared to Model 4, but the overall AUC and accuracy was sub-optimal owing to the loss of specificity. From the 5 repeats for Model 4, the median performance (AUC of 0.976) was adopted as ALMnet. With a cut-off value of 0.5 for predictive output of ALM classification, ALMnet demonstrated an accuracy of 92.5% (95% CI, 87.9%\u201395.7%), a sensitivity of 90.0% (95% CI, 82.4%\u201395.1%), and a specificity of 95.0% (95% CI, 88.7%\u201398.4%). For human-set-100, it demonstrated an accuracy of 94.0% (95% CI, 87.4%\u201397.8%), a sensitivity of 92.0% (95% CI, 80.8%\u201397.8%), and a specificity of 96.0% (95% CI, 86.3%\u201399.5%).\n\n**Performance of Human Evaluators in Stages I and II**\n\nA total of 60 participants were provided with 100 dermoscopic images (human-set-100) and made initial decisions for each case. The performances of the participants and ALMnet are summarized in Table 1 and Figs 2 and 3. The interrater agreement rates between participant responses, ALMnet predictions, and the diagnostic gold standard are presented in Fig. 4.\n\nIn Stage I, participants showed an accuracy of 74.7% (95% CI, 72.6%\u201376.8%), a sensitivity of 79.9% (95% CI, 76.2%\u201383.5%), and a specificity of 69.5% (95% CI, 65.1%\u201373.8%) (Table 1I). Despite some exceptions, most participants were outperformed by ALMnet (Fig. 2). Moreover, a prominent disagreement (Fig. 4A, Fleiss-\u03ba of 0.436 [95% CI, 0.370\u20130.503]) and performance gap (Fig. S7A) were observed across participants.\n\nIn Stage II, participants were given dermoscopic images along with each patient\u2019s age, sex, and site of lesion. The accuracy, sensitivity, and specificity were 79.0% (95% CI, 76.7%\u201381.2%), 81.5% (95% CI, 77.7%\u201385.2%), and 76.4% (95% CI, 72.5%\u201380.4%), respectively. The differences compared to Stage I were 4.3%p (95% CI, 2.5%p\u20136.1%p), 1.6%p (95% CI, -1.0%p to 4.2%p), and 7.0%p (95% CI, 3.2%p\u201312.4%p), respectively. Although the availability of clinical information increased participants\u2019 overall performances, many were still outperformed by ALMnet. Regarding interrater agreement, we see that despite a slightly improved concordance between participants, a considerable disagreement was still observed (Fig. S7B and Fig. 4B, Fleiss-\u03ba of 0.506 [95% CI, 0.437\u20130.573]).\n\n**Augmented Decision Making of Human Evaluators in Stage III**\n\nIn Stage III, participants were further provided with the diagnosis and its probability, as predicted by ALMnet for each case. Compared to Stage II, 811/6000 (13.5%) of participant responses were changed as a result of augmented decision making, and 643/811 (79.3%) of these resulted in increased accuracy (Fig. S8). The accuracy, sensitivity, and specificity were 86.9% (95% CI, 85.3%\u201388.4%), 88.7% (95% CI, 86.0%\u201391.5%), and 85.0% (95% CI, 82.7%\u201387.3%), respectively. The differences compared to Stage I were 12.2%p (95% CI, 10.1%p\u201314.3%p), 8.9%p (95% CI, 6.1%p\u201311.7%p), and 15.5%p (95% CI, 11.4%p\u201319.7%p), respectively. Similarly, those compared to Stage II were 7.9%p (95% CI, 6.0%p\u20139.9%p), 7.3%p (95% CI, 4.5%p\u201310.0%p), and 8.6%p (95% CI, 5.3%p\u201311.8%p), respectively, demonstrating a significant improvement in participants\u2019 performances. This improvement was particularly emphasized in the relatively inexperienced groups of dermatology residents and general physicians. Moreover, the performance gap across participants was considerably diminished, and the overall agreement across participants greatly improved (Fig. S7C and Fig. 4C, Fleiss-\u03ba of 0.684 [95% CI, 0.621\u20130.749]).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented decision-making for acral lentiginous melanoma detection using deep convolutional neural networks.",
        "abstract": "## BACKGROUND\nSeveral studies have achieved high-level performance of melanoma detection using convolutional neural networks (CNNs). However, few have described the extent to which the implementation of CNNs improves the diagnostic performance of the physicians.\n## OBJECTIVE\nThis study is aimed at developing a CNN for detecting acral lentiginous melanoma (ALM) and investigating whether its implementation can improve the initial decision for ALM detection made by the physicians.\n## METHODS\nA CNN was trained using 1072 dermoscopic images of acral benign nevi, ALM and intermediate tumours. To investigate whether the implementation of CNN can improve the initial decision for ALM detection, 60 physicians completed a three-stage survey. In Stage I, they were asked for their decisions solely on the basis of dermoscopic images provided to them. In Stage II, they were also provided with clinical information. In Stage III, they were provided with the additional diagnosis and probability predicted by the CNN.\n## RESULTS\nThe accuracy of ALM detection in the participants was 74.7% (95% confidence interval [CI], 72.6-76.8%) in Stage I and 79.0% (95% CI, 76.7-81.2%) in Stage II. In Stage III, it was 86.9% (95% CI, 85.3-88.4%), which exceeds the accuracy delivered in Stage I by 12.2%p (95% CI, 10.1-14.3%p) and Stage II by 7.9%p (95% CI, 6.0-9.9%p). Moreover, the concordance between the participants considerably increased (Fleiss-\u03ba of 0.436 [95% CI, 0.437-0.573] in Stage I, 0.506 [95% CI, 0.621-0.749] in Stage II and 0.684 [95% CI, 0.621-0.749] in Stage III).\n## CONCLUSIONS\nAugmented decision-making improved the performance of and concordance between the clinical decisions of a diverse group of experts. This study demonstrates the potential use of CNNs as an adjoining, decision-supporting system for physicians' decisions.\n",
        "methods": "**Materials and Methods**\n\n\nDataset\nThis retrospective analysis included dermoscopic images of acral pigmentary lesions collected from 2014 to 2019 at the Department of Dermatology, Severance Hospital, Seoul, Korea. As a means of ensuring quality control, from the dermoscopic image repository of our institution, we constructed a balanced dataset of 500 acral benign nevi (BN) images without atypical or dysplastic features and 500 ALM images. An additional 72 images of acral nevi with atypical or dysplastic features in their histopathological findings were separately collected and labelled as intermediate tumours.16 All dermoscopic images were captured using a DermLite Cam\u00ae (3 Gen, California) in a cross polarized or non-polarized mode, and with or without the use of contact media. Although dCNNs generally require extensive amounts of training data, the relatively higher homogeneity of dermoscopic images allowed us to considerably reduce the amount of data required to train the network.17Because the data were acquired from a population-based clinical setting, many of the BN images were provided alongside their respective clinical diagnoses from follow-up examinations. To ensure accurate diagnoses, these were reviewed by three board-certified dermatologists during the dataset-construction stages. The other samples were obtained following the complete excision of acral lesions by us (K.Y.C. and B.O.), enabling histopathologic assessment. The haematoxylin and eosin, and immunohistochemical (if available) stains were reviewed by board-certified pathologists and dermatologists, and the results served as the diagnostic gold standard throughout the study. In total, 800 (80%) randomly selected BN and ALM images, and all 72 intermediate tumour images, were used to train the dCNN. The remaining 200 (20%) were used for validation (test-set 200, n=200). Two datasets were reviewed to ensure lesion-stratified partitioning. Clinical information was obtained by reviewing patients\u2019 electronic medical charts (Table S1). This study was approved by the institutional review board of Yonsei University College of Medicine (Approval No. 4-2019-0282) and was carried out in accordance with the Declaration of Helsinki. A waiver of written informed consent was granted for the de-identified data used.\nDeep Convolutional Neural Network\nWe utilized ResNet with 50 residual layers (Fig. S1)18. Given the characteristics of the acral site images, in which unnecessary background and colour oversaturation were more likely, we preprocessed images and implemented a modified version of the method used in our previous study, for training and inference (Fig. S2).19,20 To investigate an optimal method of training for ALM detection of intermediate tumours, we trained four dCNNs (Fig. S3). To identify influential pixels for tumour detection, class activation maps were generated using global average pooling (Fig. S4).21 The model with the best ALM detection performance was adopted as the final model, and is referred to as ALMnet. All procedures were repeated 5 times to ensure reproducibility and comparability. No overlap between the training and test data was allowed. Details on the training and inferential methods of the dCNN, including image pre-processing and architecture configuration, are described in Supplementary Method in Supplementary Materials.\nDecision Study\nWe recruited a total of 60 physicians (20 board-certified dermatologists, 20 dermatology residents, and 20 general physicians) to investigate how well physicians could make an initial diagnosis of ALM by interpreting dermoscopic images, and how much their performance could be improved with the aid of ALMnet. For ease of response, 100 images from test-set-200 were extracted for the decision study (human-set-100) (Fig. 1). Using Google Survey, an anonymous online-based diagnostic questionnaire was conducted in three stages over 2-week intervals (Fig. S5). In Stage I, participants were shown 100 dermoscopic images only, and asked for their decision in each case (either \u2018It is more likely a benign nevus. Follow-up is needed.\u2019 or \u2018It is more likely a malignant melanoma. Immediate biopsy or intervention is required.\u2019). In Stage II, they were provided with additional clinical information for the same cases and were asked to answer in the same manner. In most previous studies, human evaluators were asked for their diagnoses without being given any clinical information (equivalent to Stage I). However, this does not occur in clinical settings,  which is why we prepared this intermediate stage. In Stage III, they were further provided with the ALMnet diagnosis (either BN or ALM) and the confidence score for each case, and then asked for their answer. The suggested diagnosis was determined by a dichotomous classification (BN versus ALM), produced by interpreting the prediction output with a fixed cut-off (> 0.5 favours ALM). The performance of participants was assessed by comparing their predictions and associated decisions to the diagnostic gold standard. At each stage, the sequences of cases were shuffled to ensure unbiased responses at each stage. The reference diagnosis for each case and participants\u2019 scores were not disclosed until the end of the study.\nOutcome Measurement and Statistical Analysis\nThe area  under curve (AUC) of the receiver operating characteristics (ROC) curve was taken as the primary measurement for comparing the performances of multiple dCNNs. ROC curves were depicted using the prediction output for dichotomous classification of BN and ALM. Two-tailed independent t-tests were used to compare the AUC values, which measured 5 repeats for each of the 4 models. For the model with the highest performance, a repeat with the median AUC was adopted as ALMnet. The primary outcome of the decision study was the difference in accuracy across all human evaluators between Stages I, II, and III. The secondary outcomes include the differences in sensitivity and specificity. For human-set-100, participant accuracies, sensitivities, and specificities were calculated for each stage. Sensitivity was calculated as the proportion of correctly classified ALM images, and specificity was calculated as the proportion of correctly  classified BN images. Repeated measure analysis of variance was used to compare the performance of participants across the physician groups and study stages.The exact Clopper-Pearson method22 was used for calculating the 95% confidence intervals (CIs) of the dCNN performance parameters. Fleiss-\u03ba statistics23 were calculated to assess concordance among participants\u2019 responses. A heatmap was used to visualize inter-participant agreement rate, with clustering via a hierarchical agglomerative method. All statistical analyses were conducted in R version 3.5.0 (R Foundation for Statistical Computing) at a significance level of 5%. Four R packages were used: ROCit for plotting ROC curves,24 IRR25 and raters26 for calculating \u03ba-statistics, and pheatmap27 to produce the heatmap.",
        "results": "**Results**\n\n**Performance of Deep Convolutional Neural Networks**\n\nThe performances of four dCNNs for test-set-200 in ALM detection are summarized in Table S2 and Fig. S6. The ensembled model (Model 4) showed the best performance, with a highest AUC value of 0.976 (95% CI, 0.974\u20130.978), significantly higher than those of the other models. Model 2 in which intermediate tumours were trained as BN showed higher specificity compared to Model 4. However, a loss in the sensitivity resulted in a sub-optimal AUC and accuracy. On the contrary, Model 3 where these intermediate tumours were trained as ALM showed a higher sensitivity compared to Model 4, but the overall AUC and accuracy was sub-optimal owing to the loss of specificity. From the 5 repeats for Model 4, the median performance (AUC of 0.976) was adopted as ALMnet. With a cut-off value of 0.5 for predictive output of ALM classification, ALMnet demonstrated an accuracy of 92.5% (95% CI, 87.9%\u201395.7%), a sensitivity of 90.0% (95% CI, 82.4%\u201395.1%), and a specificity of 95.0% (95% CI, 88.7%\u201398.4%). For human-set-100, it demonstrated an accuracy of 94.0% (95% CI, 87.4%\u201397.8%), a sensitivity of 92.0% (95% CI, 80.8%\u201397.8%), and a specificity of 96.0% (95% CI, 86.3%\u201399.5%).\n\n**Performance of Human Evaluators in Stages I and II**\n\nA total of 60 participants were provided with 100 dermoscopic images (human-set-100) and made initial decisions for each case. The performances of the participants and ALMnet are summarized in Table 1 and Figs 2 and 3. The interrater agreement rates between participant responses, ALMnet predictions, and the diagnostic gold standard are presented in Fig. 4.\n\nIn Stage I, participants showed an accuracy of 74.7% (95% CI, 72.6%\u201376.8%), a sensitivity of 79.9% (95% CI, 76.2%\u201383.5%), and a specificity of 69.5% (95% CI, 65.1%\u201373.8%) (Table 1I). Despite some exceptions, most participants were outperformed by ALMnet (Fig. 2). Moreover, a prominent disagreement (Fig. 4A, Fleiss-\u03ba of 0.436 [95% CI, 0.370\u20130.503]) and performance gap (Fig. S7A) were observed across participants.\n\nIn Stage II, participants were given dermoscopic images along with each patient\u2019s age, sex, and site of lesion. The accuracy, sensitivity, and specificity were 79.0% (95% CI, 76.7%\u201381.2%), 81.5% (95% CI, 77.7%\u201385.2%), and 76.4% (95% CI, 72.5%\u201380.4%), respectively. The differences compared to Stage I were 4.3%p (95% CI, 2.5%p\u20136.1%p), 1.6%p (95% CI, -1.0%p to 4.2%p), and 7.0%p (95% CI, 3.2%p\u201312.4%p), respectively. Although the availability of clinical information increased participants\u2019 overall performances, many were still outperformed by ALMnet. Regarding interrater agreement, we see that despite a slightly improved concordance between participants, a considerable disagreement was still observed (Fig. S7B and Fig. 4B, Fleiss-\u03ba of 0.506 [95% CI, 0.437\u20130.573]).\n\n**Augmented Decision Making of Human Evaluators in Stage III**\n\nIn Stage III, participants were further provided with the diagnosis and its probability, as predicted by ALMnet for each case. Compared to Stage II, 811/6000 (13.5%) of participant responses were changed as a result of augmented decision making, and 643/811 (79.3%) of these resulted in increased accuracy (Fig. S8). The accuracy, sensitivity, and specificity were 86.9% (95% CI, 85.3%\u201388.4%), 88.7% (95% CI, 86.0%\u201391.5%), and 85.0% (95% CI, 82.7%\u201387.3%), respectively. The differences compared to Stage I were 12.2%p (95% CI, 10.1%p\u201314.3%p), 8.9%p (95% CI, 6.1%p\u201311.7%p), and 15.5%p (95% CI, 11.4%p\u201319.7%p), respectively. Similarly, those compared to Stage II were 7.9%p (95% CI, 6.0%p\u20139.9%p), 7.3%p (95% CI, 4.5%p\u201310.0%p), and 8.6%p (95% CI, 5.3%p\u201311.8%p), respectively, demonstrating a significant improvement in participants\u2019 performances. This improvement was particularly emphasized in the relatively inexperienced groups of dermatology residents and general physicians. Moreover, the performance gap across participants was considerably diminished, and the overall agreement across participants greatly improved (Fig. S7C and Fig. 4C, Fleiss-\u03ba of 0.684 [95% CI, 0.621\u20130.749]).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Augmented decision-making for acral lentiginous melanoma detection using deep convolutional neural networks.",
        "abstract": "## BACKGROUND\nSeveral studies have achieved high-level performance of melanoma detection using convolutional neural networks (CNNs). However, few have described the extent to which the implementation of CNNs improves the diagnostic performance of the physicians.\n## OBJECTIVE\nThis study is aimed at developing a CNN for detecting acral lentiginous melanoma (ALM) and investigating whether its implementation can improve the initial decision for ALM detection made by the physicians.\n## METHODS\nA CNN was trained using 1072 dermoscopic images of acral benign nevi, ALM and intermediate tumours. To investigate whether the implementation of CNN can improve the initial decision for ALM detection, 60 physicians completed a three-stage survey. In Stage I, they were asked for their decisions solely on the basis of dermoscopic images provided to them. In Stage II, they were also provided with clinical information. In Stage III, they were provided with the additional diagnosis and probability predicted by the CNN.\n## RESULTS\nThe accuracy of ALM detection in the participants was 74.7% (95% confidence interval [CI], 72.6-76.8%) in Stage I and 79.0% (95% CI, 76.7-81.2%) in Stage II. In Stage III, it was 86.9% (95% CI, 85.3-88.4%), which exceeds the accuracy delivered in Stage I by 12.2%p (95% CI, 10.1-14.3%p) and Stage II by 7.9%p (95% CI, 6.0-9.9%p). Moreover, the concordance between the participants considerably increased (Fleiss-\u03ba of 0.436 [95% CI, 0.437-0.573] in Stage I, 0.506 [95% CI, 0.621-0.749] in Stage II and 0.684 [95% CI, 0.621-0.749] in Stage III).\n## CONCLUSIONS\nAugmented decision-making improved the performance of and concordance between the clinical decisions of a diverse group of experts. This study demonstrates the potential use of CNNs as an adjoining, decision-supporting system for physicians' decisions.\n",
        "methods": "**Materials and Methods**\n\n\nDataset\nThis retrospective analysis included dermoscopic images of acral pigmentary lesions collected from 2014 to 2019 at the Department of Dermatology, Severance Hospital, Seoul, Korea. As a means of ensuring quality control, from the dermoscopic image repository of our institution, we constructed a balanced dataset of 500 acral benign nevi (BN) images without atypical or dysplastic features and 500 ALM images. An additional 72 images of acral nevi with atypical or dysplastic features in their histopathological findings were separately collected and labelled as intermediate tumours.16 All dermoscopic images were captured using a DermLite Cam\u00ae (3 Gen, California) in a cross polarized or non-polarized mode, and with or without the use of contact media. Although dCNNs generally require extensive amounts of training data, the relatively higher homogeneity of dermoscopic images allowed us to considerably reduce the amount of data required to train the network.17Because the data were acquired from a population-based clinical setting, many of the BN images were provided alongside their respective clinical diagnoses from follow-up examinations. To ensure accurate diagnoses, these were reviewed by three board-certified dermatologists during the dataset-construction stages. The other samples were obtained following the complete excision of acral lesions by us (K.Y.C. and B.O.), enabling histopathologic assessment. The haematoxylin and eosin, and immunohistochemical (if available) stains were reviewed by board-certified pathologists and dermatologists, and the results served as the diagnostic gold standard throughout the study. In total, 800 (80%) randomly selected BN and ALM images, and all 72 intermediate tumour images, were used to train the dCNN. The remaining 200 (20%) were used for validation (test-set 200, n=200). Two datasets were reviewed to ensure lesion-stratified partitioning. Clinical information was obtained by reviewing patients\u2019 electronic medical charts (Table S1). This study was approved by the institutional review board of Yonsei University College of Medicine (Approval No. 4-2019-0282) and was carried out in accordance with the Declaration of Helsinki. A waiver of written informed consent was granted for the de-identified data used.\nDeep Convolutional Neural Network\nWe utilized ResNet with 50 residual layers (Fig. S1)18. Given the characteristics of the acral site images, in which unnecessary background and colour oversaturation were more likely, we preprocessed images and implemented a modified version of the method used in our previous study, for training and inference (Fig. S2).19,20 To investigate an optimal method of training for ALM detection of intermediate tumours, we trained four dCNNs (Fig. S3). To identify influential pixels for tumour detection, class activation maps were generated using global average pooling (Fig. S4).21 The model with the best ALM detection performance was adopted as the final model, and is referred to as ALMnet. All procedures were repeated 5 times to ensure reproducibility and comparability. No overlap between the training and test data was allowed. Details on the training and inferential methods of the dCNN, including image pre-processing and architecture configuration, are described in Supplementary Method in Supplementary Materials.\nDecision Study\nWe recruited a total of 60 physicians (20 board-certified dermatologists, 20 dermatology residents, and 20 general physicians) to investigate how well physicians could make an initial diagnosis of ALM by interpreting dermoscopic images, and how much their performance could be improved with the aid of ALMnet. For ease of response, 100 images from test-set-200 were extracted for the decision study (human-set-100) (Fig. 1). Using Google Survey, an anonymous online-based diagnostic questionnaire was conducted in three stages over 2-week intervals (Fig. S5). In Stage I, participants were shown 100 dermoscopic images only, and asked for their decision in each case (either \u2018It is more likely a benign nevus. Follow-up is needed.\u2019 or \u2018It is more likely a malignant melanoma. Immediate biopsy or intervention is required.\u2019). In Stage II, they were provided with additional clinical information for the same cases and were asked to answer in the same manner. In most previous studies, human evaluators were asked for their diagnoses without being given any clinical information (equivalent to Stage I). However, this does not occur in clinical settings,  which is why we prepared this intermediate stage. In Stage III, they were further provided with the ALMnet diagnosis (either BN or ALM) and the confidence score for each case, and then asked for their answer. The suggested diagnosis was determined by a dichotomous classification (BN versus ALM), produced by interpreting the prediction output with a fixed cut-off (> 0.5 favours ALM). The performance of participants was assessed by comparing their predictions and associated decisions to the diagnostic gold standard. At each stage, the sequences of cases were shuffled to ensure unbiased responses at each stage. The reference diagnosis for each case and participants\u2019 scores were not disclosed until the end of the study.\nOutcome Measurement and Statistical Analysis\nThe area  under curve (AUC) of the receiver operating characteristics (ROC) curve was taken as the primary measurement for comparing the performances of multiple dCNNs. ROC curves were depicted using the prediction output for dichotomous classification of BN and ALM. Two-tailed independent t-tests were used to compare the AUC values, which measured 5 repeats for each of the 4 models. For the model with the highest performance, a repeat with the median AUC was adopted as ALMnet. The primary outcome of the decision study was the difference in accuracy across all human evaluators between Stages I, II, and III. The secondary outcomes include the differences in sensitivity and specificity. For human-set-100, participant accuracies, sensitivities, and specificities were calculated for each stage. Sensitivity was calculated as the proportion of correctly classified ALM images, and specificity was calculated as the proportion of correctly  classified BN images. Repeated measure analysis of variance was used to compare the performance of participants across the physician groups and study stages.The exact Clopper-Pearson method22 was used for calculating the 95% confidence intervals (CIs) of the dCNN performance parameters. Fleiss-\u03ba statistics23 were calculated to assess concordance among participants\u2019 responses. A heatmap was used to visualize inter-participant agreement rate, with clustering via a hierarchical agglomerative method. All statistical analyses were conducted in R version 3.5.0 (R Foundation for Statistical Computing) at a significance level of 5%. Four R packages were used: ROCit for plotting ROC curves,24 IRR25 and raters26 for calculating \u03ba-statistics, and pheatmap27 to produce the heatmap.",
        "results": "**Results**\n\n**Performance of Deep Convolutional Neural Networks**\n\nThe performances of four dCNNs for test-set-200 in ALM detection are summarized in Table S2 and Fig. S6. The ensembled model (Model 4) showed the best performance, with a highest AUC value of 0.976 (95% CI, 0.974\u20130.978), significantly higher than those of the other models. Model 2 in which intermediate tumours were trained as BN showed higher specificity compared to Model 4. However, a loss in the sensitivity resulted in a sub-optimal AUC and accuracy. On the contrary, Model 3 where these intermediate tumours were trained as ALM showed a higher sensitivity compared to Model 4, but the overall AUC and accuracy was sub-optimal owing to the loss of specificity. From the 5 repeats for Model 4, the median performance (AUC of 0.976) was adopted as ALMnet. With a cut-off value of 0.5 for predictive output of ALM classification, ALMnet demonstrated an accuracy of 92.5% (95% CI, 87.9%\u201395.7%), a sensitivity of 90.0% (95% CI, 82.4%\u201395.1%), and a specificity of 95.0% (95% CI, 88.7%\u201398.4%). For human-set-100, it demonstrated an accuracy of 94.0% (95% CI, 87.4%\u201397.8%), a sensitivity of 92.0% (95% CI, 80.8%\u201397.8%), and a specificity of 96.0% (95% CI, 86.3%\u201399.5%).\n\n**Performance of Human Evaluators in Stages I and II**\n\nA total of 60 participants were provided with 100 dermoscopic images (human-set-100) and made initial decisions for each case. The performances of the participants and ALMnet are summarized in Table 1 and Figs 2 and 3. The interrater agreement rates between participant responses, ALMnet predictions, and the diagnostic gold standard are presented in Fig. 4.\n\nIn Stage I, participants showed an accuracy of 74.7% (95% CI, 72.6%\u201376.8%), a sensitivity of 79.9% (95% CI, 76.2%\u201383.5%), and a specificity of 69.5% (95% CI, 65.1%\u201373.8%) (Table 1I). Despite some exceptions, most participants were outperformed by ALMnet (Fig. 2). Moreover, a prominent disagreement (Fig. 4A, Fleiss-\u03ba of 0.436 [95% CI, 0.370\u20130.503]) and performance gap (Fig. S7A) were observed across participants.\n\nIn Stage II, participants were given dermoscopic images along with each patient\u2019s age, sex, and site of lesion. The accuracy, sensitivity, and specificity were 79.0% (95% CI, 76.7%\u201381.2%), 81.5% (95% CI, 77.7%\u201385.2%), and 76.4% (95% CI, 72.5%\u201380.4%), respectively. The differences compared to Stage I were 4.3%p (95% CI, 2.5%p\u20136.1%p), 1.6%p (95% CI, -1.0%p to 4.2%p), and 7.0%p (95% CI, 3.2%p\u201312.4%p), respectively. Although the availability of clinical information increased participants\u2019 overall performances, many were still outperformed by ALMnet. Regarding interrater agreement, we see that despite a slightly improved concordance between participants, a considerable disagreement was still observed (Fig. S7B and Fig. 4B, Fleiss-\u03ba of 0.506 [95% CI, 0.437\u20130.573]).\n\n**Augmented Decision Making of Human Evaluators in Stage III**\n\nIn Stage III, participants were further provided with the diagnosis and its probability, as predicted by ALMnet for each case. Compared to Stage II, 811/6000 (13.5%) of participant responses were changed as a result of augmented decision making, and 643/811 (79.3%) of these resulted in increased accuracy (Fig. S8). The accuracy, sensitivity, and specificity were 86.9% (95% CI, 85.3%\u201388.4%), 88.7% (95% CI, 86.0%\u201391.5%), and 85.0% (95% CI, 82.7%\u201387.3%), respectively. The differences compared to Stage I were 12.2%p (95% CI, 10.1%p\u201314.3%p), 8.9%p (95% CI, 6.1%p\u201311.7%p), and 15.5%p (95% CI, 11.4%p\u201319.7%p), respectively. Similarly, those compared to Stage II were 7.9%p (95% CI, 6.0%p\u20139.9%p), 7.3%p (95% CI, 4.5%p\u201310.0%p), and 8.6%p (95% CI, 5.3%p\u201311.8%p), respectively, demonstrating a significant improvement in participants\u2019 performances. This improvement was particularly emphasized in the relatively inexperienced groups of dermatology residents and general physicians. Moreover, the performance gap across participants was considerably diminished, and the overall agreement across participants greatly improved (Fig. S7C and Fig. 4C, Fleiss-\u03ba of 0.684 [95% CI, 0.621\u20130.749]).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Human-computer collaboration for skin cancer recognition.",
        "abstract": "The rapid increase in telemedicine coupled with recent advances in diagnostic artificial intelligence (AI) create the imperative to consider the opportunities and risks of inserting AI-based support into new paradigms of care. Here we build on recent achievements in the accuracy of image-based AI for skin cancer diagnosis to address the effects of varied representations of AI-based support across different levels of clinical expertise and multiple clinical workflows. We find that good quality AI-based support of clinical decision-making improves diagnostic accuracy over that of either AI or physicians alone, and that the least experienced clinicians gain the most from AI-based support. We further find that AI-based multiclass probabilities outperformed content-based image retrieval (CBIR) representations of AI in the mobile technology environment, and AI-based support had utility in simulations of second opinions and of telemedicine triage. In addition to demonstrating the potential benefits associated with good quality AI in the hands of non-expert clinicians, we find that faulty AI can mislead the entire spectrum of clinicians, including experts. Lastly, we show that insights derived from AI class-activation maps can inform improvements in human diagnosis. Together, our approach and findings offer a framework for future studies across the spectrum of image-based diagnostics to improve human-computer collaboration in clinical practice.\n",
        "methods": "**Methods**\n\n**Network training.** We fine-tuned a CNN for classification of seven different categories of the HAM10000 dataset[17]. We performed training on NVIDIA graphics processing units (GPUs) using the Pytorch[34] framework and chose a ResNet34 (ref. [35]) architecture, with weights initiated by pretraining on ImageNet[36] data. Cross-entropy served as the loss function, with weighting dependent on the frequency of classes within the dataset. The learning rate was initialized at 0.0001, with a tenfold reduction in case of no validation loss improvement for more than three epochs, but a minimum of 1 \u00d7 10[\u2212][9]. We used adaptive moment estimation (Adam[37]) as the optimizer and performed a maximum of 100 training epochs with early stopping. Images were presented in batches of 32, randomly cropped and resized to 224 \u00d7 224 pixels without normalization of a mean pixel, randomly rotated by 90 degrees and flipped with minor jitter in color, contrast, saturation and hue. The publicly available HAM10000 dataset, which corresponds to the training set of the ISIC 2018 challenge[18], was the source of images used for training and fivefold cross-validation. We selected the single best performing network on the hold-out validation set for further interaction with raters. For inference, images were cropped to 80% and resized to 224 \u00d7 224 pixels, with minor test-time augmentation consisting of horizontal flipping and rotation by 0 or 90 degrees. For the telemedicine dataset, we also applied color normalization via Shades of Gray[38] with a Minkowski norm of 6. The multiclass probabilities presented to the raters were obtained by applying a softmax function to contain all class probabilities between 0\u2013100%. To find similar images, we used the same CNN to extract the feature vector of the target image and compared it to feature vectors of images in the HAM10000 dataset via cosine similarity[20]. We stored the four closest images of each class and presented them in the AI-based CBIR decision support.\n\n**Interaction platform and raters.** _Online interaction platform._ The web-based platform DermaChallenge, which was developed at the Medical University of Vienna, served as the interface through which the performance of human raters and AI for the diagnostic task was evaluated and quantified. The platform is split into a back end and a front end, and both are deployed on a stack of well-known web technologies (Linux, Apache, MySQL and PHP). Please refer to the Nature Research Reporting Summary for details of the specific software versions used. The back end offers a representational state transfer interface to load and persist data, as well as JavaScript Object Notation web tokens to authenticate participants. The transport layer security and secure sockets layer protocol are used to encrypt all communications. The front end is optimized for mobile devices (mobile phones and tablets) but can also be used on any other platform via a JavaScript-enabled web browser. Before public deployment, five users tested the platform.\n\n_Recruitment and characteristics of raters._ We used mailing lists and social media posts of the International Society of Dermoscopy to recruit online raters. To participate in the study, raters had to register with a username, valid email address and password. In addition, we asked raters for details on their age (age groups spanning 10 years), gender, country, profession and years of experience in dermatoscopy ((1) less than 1 year, (2) opportunistic use for more than 1 year, (3) regular use for 1 to 5 years, (4) regular use for more than 5 years or (5) more than 10 years of experience). Each rater had to perform multiple screening tests to ensure that the self-reported experience matched actual skills. Screening tests consisted of simple domain-specific tasks, for example, to assign one of the seven possible diagnoses to ten cases, to separate melanomas from non-melanomas and to separate seborrheic keratoses from other lesions. We recruited 302 raters for the first interaction study that screened different forms of AI-based support, and 155 raters were recruited for the extended interaction study (inclusion of images with faulty AI-based support) and the telemedicine study (Supplementary Table 4). The distribution of raters according to task is presented in Supplementary Table 4. Second-opinion raters consisted of eight board-certified dermatologists and three dermatology residents, who were recruited because they diagnosed and managed more than two suspicious skin lesions on a face-to-face basis between April and September 2019. For the knowledge transfer study, we invited fourth-year medical students to participate; of the 650 medical students invited, 200 agreed to participate and 189 answered more than 50% of the test questions.\n\n_Characteristics of images and patients._ The benchmark test set of the ISIC 2018 challenge served as the sample for the interaction studies[9]. Of the 1,511 dermoscopic images in this set, 928 images were collected in the Department of Dermatology at the Medical University of Vienna, 267 images were collected in the skin cancer practice of Cliff Rosendahl in Queensland and the remaining 316 images were collected in other centers in Turkey (n = 117), New Zealand (n = 87), Sweden (n = 92) and Argentina (n = 20), to ensure diversity of skin types. The mean age of patients was 50.8 years (s.d. 17.4 years), and 46.2% of patients were female. The Austrian image set consists of lesions from patients referred to a tertiary European center specializing in the early detection of melanoma in high-risk groups. This group of patients is mainly of European ancestry and have a large number of nevi and skin types I\u2013III. The Australian image set includes lesions from patients of a primary-care facility in an area with a high incidence of skin cancer. Patients are typified by Celtic complexion, skin type I or II and chronic sun damage. Routine pathology evaluation (n = 786), biology (that is, >1.5 years of sequential dermoscopic imaging without changes; n = 458), expert consensus in common, straightforward, non-melanocytic cases that were not excised (n = 260) and in vivo confocal images (n = 7) served as the ground truth. Controversial cases with ambiguous histopathologic reports were excluded. Due to random sampling, only 1,412 of 1,511 images were finally used and evaluated by the raters. The 1,412 used cases consisted of 43 AKIECs, 93 BCCs, 217 BKLs, 44 DFs, 171 MELs, 809 NVs and 35 VASCs. For the telemedicine study, we included 93 of 98 participants (mean age 41.1 years (s.d. 12.2 years); 71% female) from the intervention arm of a recently conducted prospective randomized study[27] on mobile teledermoscopy for skin self-examinations. All 93 patients permitted reuse of their images. The participants had at least two skin cancer risk factors (light skin complexion and fair hair; skin that never or rarely tans and always or mostly burns; a family history of melanoma or a personal history of skin cancer, or many nevi; and residing in Queensland) as self-reported in the eligibility survey. A teledermoscopic evaluation was performed for all lesions. Face-to-face examination by an experienced board-certified dermatologist (H.P.S.) or the histopathologic report, in cases where the lesion was removed, served as the ground truth. The set of lesions consisted of 1,521 images of 596 lesions, including 29 AKIECs, 6 BCCs, 102 BKLs, 410 NVs, 2 squamous cell carcinomas (SCCs) and 9 VASCs. For calculation of diagnostic values, ground-truth data were mapped to classes of the HAM10000 dataset, if possible. We excluded nonspecific categories (n = 38 lesions) such as \u2018other\u2019, \u2018no lesion\u2019 or \u2018previously removed\u2019, because they could be mapped to neither the \u2018benign\u2019 nor \u2018malignant\u2019 category. The sample also included images that were not represented in the training data (non-pigmented variants of keratinocyte cancers, mucosal lesions and low-quality images), which were excluded from the telemedicine support study but not from the triage study, to better simulate a realistic scenario. For the second-opinion study, we searched the database of the Department of Dermatology at the Medical University of Vienna for dermoscopy images taken between April and September 2019. We included images if the lesion was excised and had a definite histopathologic diagnosis and if lesions were examined by a physician who was responsible for the face-to-face diagnosis of at least two other cases in this time period. The final sample set (n = 79) included 3 AKIECs, 23 BCCs, 13 BKLs, 2 DFs, 15 MELs, 21 NVs, 1 \u2018other\u2019 (scar) and 1 SCC. The mean age of patients was 64.6 years (s.d. 19.8 years), and 34.5% of patients were female. Patients were mainly of European ancestry and had skin type II (41.7%), III (57.1%) or IV (1.2%). As in the telemedicine scenario, we did not exclude images of categories that were not present in the training data or images of low quality. For the knowledge transfer study, the sample cases (n = 25) were randomly selected from the ISIC 2018 test set and stratified by diagnosis (6 AKIECs, 3 BCCs, 3 BKLs, 3 DFs, 3 MELs, 4 NVs and 3 VASCs).\n\n_Design of diagnostic studies._ To test the interaction of raters with different forms of AI-based decision support, we generated batches of 28 images. Each batch contained four randomly selected examples of every class. The raters\u2019 task was to diagnose the 28 unknown test images, first without and then with one type of decision support. We created a stratified randomization procedure to ensure a balanced distribution of the four types of decision support over all disease categories. The interaction study was online from 29 May 2019 to 15 January 2020. We excluded tests if the number of correct answers was lower than expected by chance to avoid noisy random data. We included only the first five tests for each rater to avoid biasing the results toward raters with high repetitions. The extended interaction study was open for participation between 15 January 2020 and 18 February 2020, presenting only multiclass probabilities as decision support. It included one image for every diagnosis from the ISIC 2018 test set with unaltered AI-based multiclass probabilities, two images with shuffled (resulting as incorrect) AI-based multiclass probabilities and eight images from the telemedicine study (see \u2018Characteristics of images and patients\u2019). The image sources were not disclosed to the raters. The second-opinion study was performed on a local web interface. Physicians who examined the patient face to face in real life were asked to reconsider their diagnosis and decisions with AI-based support. The case presentations included metadata (age, gender and localization), overview and close-up images (if available) and dermoscopic images. Physicians were not made aware that they had treated the patient before or of their previous decision on the case. Physicians were asked to provide their best diagnosis out of the seven predefined disease categories, as well as an extra category termed \u2018other\u2019, followed by their management decision (\u2018no intervention\u2019, \u2018monitor\u2019 or \u2018excise\u2019). No time constraints were set for this task. For the knowledge transfer study, we first examined the gradient-weighted class-activation maps[29], which were created for all images of the training set. We observed that the background attention of the CNN was significantly higher for predictions of the \u2018pigmented actinic keratosis\u2019 class than for other classes (P = 4.6 \u00d7 10[\u2212][12], two-sided unpaired t-test; Extended Data Fig. 2). We interpreted this finding as a diagnostic clue that points to the severely sun-damaged skin in the background of actinic keratoses, which is usually absent or not as severe in other disease categories. To test the hypothesis that teaching this clue to humans will improve their diagnostic skills, fourth-year medical students without previous knowledge of skin cancer detection received a 30-min introductory lecture about dermoscopy, and immediately thereafter students had to diagnose 25 test images (single best diagnosis). Answers were collected with a wireless audience response and voting system. Next, the lecturer presented an additional clue of \u2018sun-damaged skin in the surrounding skin of actinic keratoses\u2019 and the students repeated the test.\n\n**Statistics.** To simulate collective ratings of realistically small human groups (Fig. 3a), we confined the dataset to images with at least three distinct ratings (resulting range of ratings per image: 3\u201369). For each image, we created 30 bootstraps of three to five randomly selected ratings, whichever was the maximum available without replacement, and determined the most common rating as the prediction of the collective (first past the post). Ties were broken randomly. Next, we calculated the proportion of correct bootstrapped predictions to obtain the mean accuracy for each image as published previously[39]. To combine human collectives with CNN-based predictions, we took the arithmetic mean of the sum of the human multiclass probabilities, which were derived from the frequencies of bootstrapped human ratings, and the corresponding CNN-based multiclass probabilities. For analyses of diagnoses, we averaged the results for each image before comparisons; for analyses of raters with and without decision support, we calculated the arithmetic mean for each user before comparisons. The mean answering time for each user in every interaction modality served as a surrogate marker for confidence; answers that were faster or slower than the individual mean were regarded as \u2018confident\u2019 or \u2018non-confident\u2019, respectively. For the filtering procedure in the telemedicine study, we used a predefined cutoff of \u22650.17 to indicate malignancy, because this cutoff was selected by human raters in the interaction study (Extended Data Fig. 2). If patients photographed a lesion more than once, a single image above the cutoff was sufficient to label the lesion as \u2018probably malignant\u2019 and likewise on the patient level. We used a one-sample t-test to distinguish whether continuous data with normal distributions deviated from zero. Comparisons of continuous data between groups were performed with paired or unpaired t-tests or Wilcoxon signed-rank test, as appropriate. A chi-squared test was used to compare proportions. All reported P values were corrected for multiple testing (Holm\u2013Bonferroni[40]), and a two-sided P value < 0.05 was regarded as statistically significant. All analyses were performed using R v3.6.2 (ref. [41]), and plots were created with ggplot v3.2.1 (ref. [42]) and ggalluvial v0.11.1.\n\n**Reporting Summary.** Further information on research design is available in the Nature Research Reporting Summary linked to this article.",
        "results": "**Results**\n\nOur results show that decision support with AI-based multiclass probabilities improves the accuracy of human raters from 63.6% to 77.0% (increase of 13.3%, 95% CI 11.5% to 15.2%; P = 4.9 \u00d7 10[\u2212][35], two-sided paired t-test, t = 14.5, d.f. = 301; n = 302 raters), but no improvement was observed for decision support with AI-based prediction of malignancy or with our representation of AI-based CBIR. This suggests that the form of decision support should be in accordance with the given task. The probability of malignancy may be useful for simple binary management decisions such as whether to perform a biopsy or not, but not for a multiclass diagnostic problem. The studied form of AI-based CBIR is neither simple nor concrete; it needs more extensive cognitive engagement in terms of time and decision-making, because the rater needs to extrapolate the diagnosis from similarities between the test image and images with known diagnoses. The raters needed significantly more time to interact with AI-based CBIR decision support than with other types of support. Over time, human raters also tended to ignore the AI-based CBIR decision support. However, given that a large spectrum of CBIR approaches are described in the literature, another form of CBIR may still provide benefit. \n\nAfter we established that multiclass probabilities were the best form of CNN output for the given task, we focused on this form to explore the impact of AI-based support on human performance in more detail. We show an inverse relationship between the net gain from AI-based support and rater experience (Pearson\u2019s r = \u22120.18, 95% CI \u22120.28 to \u22120.07, P = 1.5 \u00d7 10[\u2212][2]; n = 302 raters). Raters in the least experienced group changed their initial diagnosis more often than experts (mean 26.0%, 95% CI 21.3% to 30.7% versus mean 14.7%, 95% CI 9.9% to 19.6%). Expert raters benefited only marginally (net gain 13.4%, 95% CI 6.3% to 20.6%) and only if they were not confident with their initial diagnosis, but not if they were confident. If experts were confident, they were usually correct and did not need support. \n\nThis finding suggests that, if experts have high confidence in their initial diagnosis, they should ignore AI-based support or not use it at all. The least experienced raters tended to accept AI-based support that contradicted their initial diagnosis even if they were confident. In general, raters changed their initial diagnosis less often if they were confident than if they were not confident in their decision (14.7%, 95% CI 12.6% to 16.8% versus 37.5%, 95% CI 34.0% to 41.0%; P = 1.9 \u00d7 10[\u2212][25], two-sided paired t-test; n = 302 raters). \n\nHaving established a positive impact of good quality AI-based support on diagnostic accuracy, we tested the impact of \u2018faulty\u2019 AI on diagnostic accuracy. Faulty AI could result from the application of AI algorithms on examples beyond the domain of images on which the AI was trained or the more remote possibility of adversarial attacks. To represent faulty AI, we intentionally generated misleading AI-based multiclass probabilities. If the top class probability of the CNN favored the correct diagnosis, we switched the probabilities in such a way that the CNN output favored a random incorrect diagnosis. We demonstrate that any previously observed gains in accuracy with AI-based support turn into a loss when that AI support is faulty. Our results suggest that, if raters build up the trust that is necessary to benefit from AI-based support, they are also vulnerable to perform below their expected ability if there is a fault with the AI. \n\nAnother finding of importance is that the benefit of human\u2013computer collaboration is asymmetrically distributed across disease categories. Our data showed that the net gain was higher for the class of pigmented actinic keratoses and intraepithelial carcinoma (increase of 31.5%, 95% CI 22.9% to 40.1%; n = 43 images) than for other categories. This suggests that the benefit of AI-based support needs to be adapted to the given task and the expected prevalence of target conditions. \n\nWe further demonstrate that AI-based multiclass ranking and probabilities have an impact on the raters\u2019 tendency to change their initial diagnosis. Most changes occurred in favor of the AI-predicted top category. Raters typically maintained their decisions that were in disagreement with the AI prediction only if that decision was ranked by AI prediction as at least the second or third option. Furthermore, raters tended to change their assessments more frequently when the difference in the AI-predicted probability between the initially selected category and the AI top category was high. \n\nAdditionally, we demonstrate that aggregated AI-based multiclass probabilities and crowd wisdom significantly increased the number of correct diagnoses in comparison to individual raters or AI in isolation. \n\nNext, we analyzed the impact of AI-based support in clinically relevant scenarios. To examine the potential of AI-based support in telemedicine, we reused prospectively collected images of a randomized controlled trial on self-examinations in high-risk patients. While the CNN was trained only on curated images of pigmented lesions, this sample also included non-pigmented variants of keratinocyte cancer, mucosal lesions and low-quality images. Although the proportion of correct specific diagnoses was significantly lower for these images (53.9% versus 76.2%; P = 8.9 \u00d7 10[\u2212][14], chi-squared test; n = 1,430 images), the CNN was able to recognize 95.2% of patients with skin cancer at a specificity of 59.2%. \n\nSimilarly to recent findings in AI-based breast cancer screening, our results indicate that AI-based skin cancer screening could triage high-risk cases and extend the intervals between face-to-face visits in low-risk cases. The optimal operating points to balance the potential benefits of AI-based triage with the risk of filtering out patients with skin cancer remain to be determined. \n\nIn another scenario, we asked dermatologists to rethink their face-to-face decisions in suspicious cases after providing them with AI-based multiclass probabilities, but without making them aware that they had previously managed the patient. With AI-based support, dermatologists switched from \u2018excision\u2019 to \u2018monitor\u2019 in 15.5% (7 of 45) of decisions for benign lesions, without increasing the number of malignant lesions that switched contrariwise. This result illustrates how human\u2013computer collaboration could decrease the number of unwarranted interventions and costs. AI-based support in this setting increased the frequency of correct specific diagnoses from 55.6% to 75.0%. \n\nFinally, we demonstrate that explanations for AI-based predictions can be translated into a human-understandable visual concept. By analyzing gradient-weighted class activation mapping (Grad-CAM), we show that attention of the CNN outside the object is higher for the prediction of actinic keratoses than for other categories. We hypothesize that, due to visual entrenchment, humans focus on the lesion and not on the background and frequently miss this clue. Here we show that teaching medical students to pay attention to chronic sun damage in the background improved the frequency of correct diagnoses of pigmented actinic keratoses from 32.5% to 47.3%. The overall frequency of correct diagnoses in all categories combined increased from 55.2% to 59.1%. \n\nThis study examines human\u2013computer collaboration from multiple angles and under varying conditions. We used the domain of skin cancer recognition for simplicity, but our study could serve as a framework for similar research in image-based diagnostic medicine. In contrast to the current narrative, our findings suggest that the primary focus should shift from human\u2013computer competition to human\u2013computer collaboration. From a regulatory perspective, the performance of AI-based systems should be tested under real-world conditions in the hands of the intended users and not as stand-alone devices. Only then can we expect to rationally adopt and improve AI-based decision support and to accelerate its evolution.",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Human-computer collaboration for skin cancer recognition.",
        "abstract": "The rapid increase in telemedicine coupled with recent advances in diagnostic artificial intelligence (AI) create the imperative to consider the opportunities and risks of inserting AI-based support into new paradigms of care. Here we build on recent achievements in the accuracy of image-based AI for skin cancer diagnosis to address the effects of varied representations of AI-based support across different levels of clinical expertise and multiple clinical workflows. We find that good quality AI-based support of clinical decision-making improves diagnostic accuracy over that of either AI or physicians alone, and that the least experienced clinicians gain the most from AI-based support. We further find that AI-based multiclass probabilities outperformed content-based image retrieval (CBIR) representations of AI in the mobile technology environment, and AI-based support had utility in simulations of second opinions and of telemedicine triage. In addition to demonstrating the potential benefits associated with good quality AI in the hands of non-expert clinicians, we find that faulty AI can mislead the entire spectrum of clinicians, including experts. Lastly, we show that insights derived from AI class-activation maps can inform improvements in human diagnosis. Together, our approach and findings offer a framework for future studies across the spectrum of image-based diagnostics to improve human-computer collaboration in clinical practice.\n",
        "methods": "**Methods**\n\n**Network training.** We fine-tuned a CNN for classification of seven different categories of the HAM10000 dataset[17]. We performed training on NVIDIA graphics processing units (GPUs) using the Pytorch[34] framework and chose a ResNet34 (ref. [35]) architecture, with weights initiated by pretraining on ImageNet[36] data. Cross-entropy served as the loss function, with weighting dependent on the frequency of classes within the dataset. The learning rate was initialized at 0.0001, with a tenfold reduction in case of no validation loss improvement for more than three epochs, but a minimum of 1 \u00d7 10[\u2212][9]. We used adaptive moment estimation (Adam[37]) as the optimizer and performed a maximum of 100 training epochs with early stopping. Images were presented in batches of 32, randomly cropped and resized to 224 \u00d7 224 pixels without normalization of a mean pixel, randomly rotated by 90 degrees and flipped with minor jitter in color, contrast, saturation and hue. The publicly available HAM10000 dataset, which corresponds to the training set of the ISIC 2018 challenge[18], was the source of images used for training and fivefold cross-validation. We selected the single best performing network on the hold-out validation set for further interaction with raters. For inference, images were cropped to 80% and resized to 224 \u00d7 224 pixels, with minor test-time augmentation consisting of horizontal flipping and rotation by 0 or 90 degrees. For the telemedicine dataset, we also applied color normalization via Shades of Gray[38] with a Minkowski norm of 6. The multiclass probabilities presented to the raters were obtained by applying a softmax function to contain all class probabilities between 0\u2013100%. To find similar images, we used the same CNN to extract the feature vector of the target image and compared it to feature vectors of images in the HAM10000 dataset via cosine similarity[20]. We stored the four closest images of each class and presented them in the AI-based CBIR decision support.\n\n**Interaction platform and raters.** _Online interaction platform._ The web-based platform DermaChallenge, which was developed at the Medical University of Vienna, served as the interface through which the performance of human raters and AI for the diagnostic task was evaluated and quantified. The platform is split into a back end and a front end, and both are deployed on a stack of well-known web technologies (Linux, Apache, MySQL and PHP). Please refer to the Nature Research Reporting Summary for details of the specific software versions used. The back end offers a representational state transfer interface to load and persist data, as well as JavaScript Object Notation web tokens to authenticate participants. The transport layer security and secure sockets layer protocol are used to encrypt all communications. The front end is optimized for mobile devices (mobile phones and tablets) but can also be used on any other platform via a JavaScript-enabled web browser. Before public deployment, five users tested the platform.\n\n_Recruitment and characteristics of raters._ We used mailing lists and social media posts of the International Society of Dermoscopy to recruit online raters. To participate in the study, raters had to register with a username, valid email address and password. In addition, we asked raters for details on their age (age groups spanning 10 years), gender, country, profession and years of experience in dermatoscopy ((1) less than 1 year, (2) opportunistic use for more than 1 year, (3) regular use for 1 to 5 years, (4) regular use for more than 5 years or (5) more than 10 years of experience). Each rater had to perform multiple screening tests to ensure that the self-reported experience matched actual skills. Screening tests consisted of simple domain-specific tasks, for example, to assign one of the seven possible diagnoses to ten cases, to separate melanomas from non-melanomas and to separate seborrheic keratoses from other lesions. We recruited 302 raters for the first interaction study that screened different forms of AI-based support, and 155 raters were recruited for the extended interaction study (inclusion of images with faulty AI-based support) and the telemedicine study (Supplementary Table 4). The distribution of raters according to task is presented in Supplementary Table 4. Second-opinion raters consisted of eight board-certified dermatologists and three dermatology residents, who were recruited because they diagnosed and managed more than two suspicious skin lesions on a face-to-face basis between April and September 2019. For the knowledge transfer study, we invited fourth-year medical students to participate; of the 650 medical students invited, 200 agreed to participate and 189 answered more than 50% of the test questions.\n\n_Characteristics of images and patients._ The benchmark test set of the ISIC 2018 challenge served as the sample for the interaction studies[9]. Of the 1,511 dermoscopic images in this set, 928 images were collected in the Department of Dermatology at the Medical University of Vienna, 267 images were collected in the skin cancer practice of Cliff Rosendahl in Queensland and the remaining 316 images were collected in other centers in Turkey (n = 117), New Zealand (n = 87), Sweden (n = 92) and Argentina (n = 20), to ensure diversity of skin types. The mean age of patients was 50.8 years (s.d. 17.4 years), and 46.2% of patients were female. The Austrian image set consists of lesions from patients referred to a tertiary European center specializing in the early detection of melanoma in high-risk groups. This group of patients is mainly of European ancestry and have a large number of nevi and skin types I\u2013III. The Australian image set includes lesions from patients of a primary-care facility in an area with a high incidence of skin cancer. Patients are typified by Celtic complexion, skin type I or II and chronic sun damage. Routine pathology evaluation (n = 786), biology (that is, >1.5 years of sequential dermoscopic imaging without changes; n = 458), expert consensus in common, straightforward, non-melanocytic cases that were not excised (n = 260) and in vivo confocal images (n = 7) served as the ground truth. Controversial cases with ambiguous histopathologic reports were excluded. Due to random sampling, only 1,412 of 1,511 images were finally used and evaluated by the raters. The 1,412 used cases consisted of 43 AKIECs, 93 BCCs, 217 BKLs, 44 DFs, 171 MELs, 809 NVs and 35 VASCs. For the telemedicine study, we included 93 of 98 participants (mean age 41.1 years (s.d. 12.2 years); 71% female) from the intervention arm of a recently conducted prospective randomized study[27] on mobile teledermoscopy for skin self-examinations. All 93 patients permitted reuse of their images. The participants had at least two skin cancer risk factors (light skin complexion and fair hair; skin that never or rarely tans and always or mostly burns; a family history of melanoma or a personal history of skin cancer, or many nevi; and residing in Queensland) as self-reported in the eligibility survey. A teledermoscopic evaluation was performed for all lesions. Face-to-face examination by an experienced board-certified dermatologist (H.P.S.) or the histopathologic report, in cases where the lesion was removed, served as the ground truth. The set of lesions consisted of 1,521 images of 596 lesions, including 29 AKIECs, 6 BCCs, 102 BKLs, 410 NVs, 2 squamous cell carcinomas (SCCs) and 9 VASCs. For calculation of diagnostic values, ground-truth data were mapped to classes of the HAM10000 dataset, if possible. We excluded nonspecific categories (n = 38 lesions) such as \u2018other\u2019, \u2018no lesion\u2019 or \u2018previously removed\u2019, because they could be mapped to neither the \u2018benign\u2019 nor \u2018malignant\u2019 category. The sample also included images that were not represented in the training data (non-pigmented variants of keratinocyte cancers, mucosal lesions and low-quality images), which were excluded from the telemedicine support study but not from the triage study, to better simulate a realistic scenario. For the second-opinion study, we searched the database of the Department of Dermatology at the Medical University of Vienna for dermoscopy images taken between April and September 2019. We included images if the lesion was excised and had a definite histopathologic diagnosis and if lesions were examined by a physician who was responsible for the face-to-face diagnosis of at least two other cases in this time period. The final sample set (n = 79) included 3 AKIECs, 23 BCCs, 13 BKLs, 2 DFs, 15 MELs, 21 NVs, 1 \u2018other\u2019 (scar) and 1 SCC. The mean age of patients was 64.6 years (s.d. 19.8 years), and 34.5% of patients were female. Patients were mainly of European ancestry and had skin type II (41.7%), III (57.1%) or IV (1.2%). As in the telemedicine scenario, we did not exclude images of categories that were not present in the training data or images of low quality. For the knowledge transfer study, the sample cases (n = 25) were randomly selected from the ISIC 2018 test set and stratified by diagnosis (6 AKIECs, 3 BCCs, 3 BKLs, 3 DFs, 3 MELs, 4 NVs and 3 VASCs).\n\n_Design of diagnostic studies._ To test the interaction of raters with different forms of AI-based decision support, we generated batches of 28 images. Each batch contained four randomly selected examples of every class. The raters\u2019 task was to diagnose the 28 unknown test images, first without and then with one type of decision support. We created a stratified randomization procedure to ensure a balanced distribution of the four types of decision support over all disease categories. The interaction study was online from 29 May 2019 to 15 January 2020. We excluded tests if the number of correct answers was lower than expected by chance to avoid noisy random data. We included only the first five tests for each rater to avoid biasing the results toward raters with high repetitions. The extended interaction study was open for participation between 15 January 2020 and 18 February 2020, presenting only multiclass probabilities as decision support. It included one image for every diagnosis from the ISIC 2018 test set with unaltered AI-based multiclass probabilities, two images with shuffled (resulting as incorrect) AI-based multiclass probabilities and eight images from the telemedicine study (see \u2018Characteristics of images and patients\u2019). The image sources were not disclosed to the raters. The second-opinion study was performed on a local web interface. Physicians who examined the patient face to face in real life were asked to reconsider their diagnosis and decisions with AI-based support. The case presentations included metadata (age, gender and localization), overview and close-up images (if available) and dermoscopic images. Physicians were not made aware that they had treated the patient before or of their previous decision on the case. Physicians were asked to provide their best diagnosis out of the seven predefined disease categories, as well as an extra category termed \u2018other\u2019, followed by their management decision (\u2018no intervention\u2019, \u2018monitor\u2019 or \u2018excise\u2019). No time constraints were set for this task. For the knowledge transfer study, we first examined the gradient-weighted class-activation maps[29], which were created for all images of the training set. We observed that the background attention of the CNN was significantly higher for predictions of the \u2018pigmented actinic keratosis\u2019 class than for other classes (P = 4.6 \u00d7 10[\u2212][12], two-sided unpaired t-test; Extended Data Fig. 2). We interpreted this finding as a diagnostic clue that points to the severely sun-damaged skin in the background of actinic keratoses, which is usually absent or not as severe in other disease categories. To test the hypothesis that teaching this clue to humans will improve their diagnostic skills, fourth-year medical students without previous knowledge of skin cancer detection received a 30-min introductory lecture about dermoscopy, and immediately thereafter students had to diagnose 25 test images (single best diagnosis). Answers were collected with a wireless audience response and voting system. Next, the lecturer presented an additional clue of \u2018sun-damaged skin in the surrounding skin of actinic keratoses\u2019 and the students repeated the test.\n\n**Statistics.** To simulate collective ratings of realistically small human groups (Fig. 3a), we confined the dataset to images with at least three distinct ratings (resulting range of ratings per image: 3\u201369). For each image, we created 30 bootstraps of three to five randomly selected ratings, whichever was the maximum available without replacement, and determined the most common rating as the prediction of the collective (first past the post). Ties were broken randomly. Next, we calculated the proportion of correct bootstrapped predictions to obtain the mean accuracy for each image as published previously[39]. To combine human collectives with CNN-based predictions, we took the arithmetic mean of the sum of the human multiclass probabilities, which were derived from the frequencies of bootstrapped human ratings, and the corresponding CNN-based multiclass probabilities. For analyses of diagnoses, we averaged the results for each image before comparisons; for analyses of raters with and without decision support, we calculated the arithmetic mean for each user before comparisons. The mean answering time for each user in every interaction modality served as a surrogate marker for confidence; answers that were faster or slower than the individual mean were regarded as \u2018confident\u2019 or \u2018non-confident\u2019, respectively. For the filtering procedure in the telemedicine study, we used a predefined cutoff of \u22650.17 to indicate malignancy, because this cutoff was selected by human raters in the interaction study (Extended Data Fig. 2). If patients photographed a lesion more than once, a single image above the cutoff was sufficient to label the lesion as \u2018probably malignant\u2019 and likewise on the patient level. We used a one-sample t-test to distinguish whether continuous data with normal distributions deviated from zero. Comparisons of continuous data between groups were performed with paired or unpaired t-tests or Wilcoxon signed-rank test, as appropriate. A chi-squared test was used to compare proportions. All reported P values were corrected for multiple testing (Holm\u2013Bonferroni[40]), and a two-sided P value < 0.05 was regarded as statistically significant. All analyses were performed using R v3.6.2 (ref. [41]), and plots were created with ggplot v3.2.1 (ref. [42]) and ggalluvial v0.11.1.\n\n**Reporting Summary.** Further information on research design is available in the Nature Research Reporting Summary linked to this article.",
        "results": "**Results**\n\nOur results show that decision support with AI-based multiclass probabilities improves the accuracy of human raters from 63.6% to 77.0% (increase of 13.3%, 95% CI 11.5% to 15.2%; P = 4.9 \u00d7 10[\u2212][35], two-sided paired t-test, t = 14.5, d.f. = 301; n = 302 raters), but no improvement was observed for decision support with AI-based prediction of malignancy or with our representation of AI-based CBIR. This suggests that the form of decision support should be in accordance with the given task. The probability of malignancy may be useful for simple binary management decisions such as whether to perform a biopsy or not, but not for a multiclass diagnostic problem. The studied form of AI-based CBIR is neither simple nor concrete; it needs more extensive cognitive engagement in terms of time and decision-making, because the rater needs to extrapolate the diagnosis from similarities between the test image and images with known diagnoses. The raters needed significantly more time to interact with AI-based CBIR decision support than with other types of support. Over time, human raters also tended to ignore the AI-based CBIR decision support. However, given that a large spectrum of CBIR approaches are described in the literature, another form of CBIR may still provide benefit. \n\nAfter we established that multiclass probabilities were the best form of CNN output for the given task, we focused on this form to explore the impact of AI-based support on human performance in more detail. We show an inverse relationship between the net gain from AI-based support and rater experience (Pearson\u2019s r = \u22120.18, 95% CI \u22120.28 to \u22120.07, P = 1.5 \u00d7 10[\u2212][2]; n = 302 raters). Raters in the least experienced group changed their initial diagnosis more often than experts (mean 26.0%, 95% CI 21.3% to 30.7% versus mean 14.7%, 95% CI 9.9% to 19.6%). Expert raters benefited only marginally (net gain 13.4%, 95% CI 6.3% to 20.6%) and only if they were not confident with their initial diagnosis, but not if they were confident. If experts were confident, they were usually correct and did not need support. \n\nThis finding suggests that, if experts have high confidence in their initial diagnosis, they should ignore AI-based support or not use it at all. The least experienced raters tended to accept AI-based support that contradicted their initial diagnosis even if they were confident. In general, raters changed their initial diagnosis less often if they were confident than if they were not confident in their decision (14.7%, 95% CI 12.6% to 16.8% versus 37.5%, 95% CI 34.0% to 41.0%; P = 1.9 \u00d7 10[\u2212][25], two-sided paired t-test; n = 302 raters). \n\nHaving established a positive impact of good quality AI-based support on diagnostic accuracy, we tested the impact of \u2018faulty\u2019 AI on diagnostic accuracy. Faulty AI could result from the application of AI algorithms on examples beyond the domain of images on which the AI was trained or the more remote possibility of adversarial attacks. To represent faulty AI, we intentionally generated misleading AI-based multiclass probabilities. If the top class probability of the CNN favored the correct diagnosis, we switched the probabilities in such a way that the CNN output favored a random incorrect diagnosis. We demonstrate that any previously observed gains in accuracy with AI-based support turn into a loss when that AI support is faulty. Our results suggest that, if raters build up the trust that is necessary to benefit from AI-based support, they are also vulnerable to perform below their expected ability if there is a fault with the AI. \n\nAnother finding of importance is that the benefit of human\u2013computer collaboration is asymmetrically distributed across disease categories. Our data showed that the net gain was higher for the class of pigmented actinic keratoses and intraepithelial carcinoma (increase of 31.5%, 95% CI 22.9% to 40.1%; n = 43 images) than for other categories. This suggests that the benefit of AI-based support needs to be adapted to the given task and the expected prevalence of target conditions. \n\nWe further demonstrate that AI-based multiclass ranking and probabilities have an impact on the raters\u2019 tendency to change their initial diagnosis. Most changes occurred in favor of the AI-predicted top category. Raters typically maintained their decisions that were in disagreement with the AI prediction only if that decision was ranked by AI prediction as at least the second or third option. Furthermore, raters tended to change their assessments more frequently when the difference in the AI-predicted probability between the initially selected category and the AI top category was high. \n\nAdditionally, we demonstrate that aggregated AI-based multiclass probabilities and crowd wisdom significantly increased the number of correct diagnoses in comparison to individual raters or AI in isolation. \n\nNext, we analyzed the impact of AI-based support in clinically relevant scenarios. To examine the potential of AI-based support in telemedicine, we reused prospectively collected images of a randomized controlled trial on self-examinations in high-risk patients. While the CNN was trained only on curated images of pigmented lesions, this sample also included non-pigmented variants of keratinocyte cancer, mucosal lesions and low-quality images. Although the proportion of correct specific diagnoses was significantly lower for these images (53.9% versus 76.2%; P = 8.9 \u00d7 10[\u2212][14], chi-squared test; n = 1,430 images), the CNN was able to recognize 95.2% of patients with skin cancer at a specificity of 59.2%. \n\nSimilarly to recent findings in AI-based breast cancer screening, our results indicate that AI-based skin cancer screening could triage high-risk cases and extend the intervals between face-to-face visits in low-risk cases. The optimal operating points to balance the potential benefits of AI-based triage with the risk of filtering out patients with skin cancer remain to be determined. \n\nIn another scenario, we asked dermatologists to rethink their face-to-face decisions in suspicious cases after providing them with AI-based multiclass probabilities, but without making them aware that they had previously managed the patient. With AI-based support, dermatologists switched from \u2018excision\u2019 to \u2018monitor\u2019 in 15.5% (7 of 45) of decisions for benign lesions, without increasing the number of malignant lesions that switched contrariwise. This result illustrates how human\u2013computer collaboration could decrease the number of unwarranted interventions and costs. AI-based support in this setting increased the frequency of correct specific diagnoses from 55.6% to 75.0%. \n\nFinally, we demonstrate that explanations for AI-based predictions can be translated into a human-understandable visual concept. By analyzing gradient-weighted class activation mapping (Grad-CAM), we show that attention of the CNN outside the object is higher for the prediction of actinic keratoses than for other categories. We hypothesize that, due to visual entrenchment, humans focus on the lesion and not on the background and frequently miss this clue. Here we show that teaching medical students to pay attention to chronic sun damage in the background improved the frequency of correct diagnoses of pigmented actinic keratoses from 32.5% to 47.3%. The overall frequency of correct diagnoses in all categories combined increased from 55.2% to 59.1%. \n\nThis study examines human\u2013computer collaboration from multiple angles and under varying conditions. We used the domain of skin cancer recognition for simplicity, but our study could serve as a framework for similar research in image-based diagnostic medicine. In contrast to the current narrative, our findings suggest that the primary focus should shift from human\u2013computer competition to human\u2013computer collaboration. From a regulatory perspective, the performance of AI-based systems should be tested under real-world conditions in the hands of the intended users and not as stand-alone devices. Only then can we expect to rationally adopt and improve AI-based decision support and to accelerate its evolution.",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Human-computer collaboration for skin cancer recognition.",
        "abstract": "The rapid increase in telemedicine coupled with recent advances in diagnostic artificial intelligence (AI) create the imperative to consider the opportunities and risks of inserting AI-based support into new paradigms of care. Here we build on recent achievements in the accuracy of image-based AI for skin cancer diagnosis to address the effects of varied representations of AI-based support across different levels of clinical expertise and multiple clinical workflows. We find that good quality AI-based support of clinical decision-making improves diagnostic accuracy over that of either AI or physicians alone, and that the least experienced clinicians gain the most from AI-based support. We further find that AI-based multiclass probabilities outperformed content-based image retrieval (CBIR) representations of AI in the mobile technology environment, and AI-based support had utility in simulations of second opinions and of telemedicine triage. In addition to demonstrating the potential benefits associated with good quality AI in the hands of non-expert clinicians, we find that faulty AI can mislead the entire spectrum of clinicians, including experts. Lastly, we show that insights derived from AI class-activation maps can inform improvements in human diagnosis. Together, our approach and findings offer a framework for future studies across the spectrum of image-based diagnostics to improve human-computer collaboration in clinical practice.\n",
        "methods": "**Methods**\n\n**Network training.** We fine-tuned a CNN for classification of seven different categories of the HAM10000 dataset[17]. We performed training on NVIDIA graphics processing units (GPUs) using the Pytorch[34] framework and chose a ResNet34 (ref. [35]) architecture, with weights initiated by pretraining on ImageNet[36] data. Cross-entropy served as the loss function, with weighting dependent on the frequency of classes within the dataset. The learning rate was initialized at 0.0001, with a tenfold reduction in case of no validation loss improvement for more than three epochs, but a minimum of 1 \u00d7 10[\u2212][9]. We used adaptive moment estimation (Adam[37]) as the optimizer and performed a maximum of 100 training epochs with early stopping. Images were presented in batches of 32, randomly cropped and resized to 224 \u00d7 224 pixels without normalization of a mean pixel, randomly rotated by 90 degrees and flipped with minor jitter in color, contrast, saturation and hue. The publicly available HAM10000 dataset, which corresponds to the training set of the ISIC 2018 challenge[18], was the source of images used for training and fivefold cross-validation. We selected the single best performing network on the hold-out validation set for further interaction with raters. For inference, images were cropped to 80% and resized to 224 \u00d7 224 pixels, with minor test-time augmentation consisting of horizontal flipping and rotation by 0 or 90 degrees. For the telemedicine dataset, we also applied color normalization via Shades of Gray[38] with a Minkowski norm of 6. The multiclass probabilities presented to the raters were obtained by applying a softmax function to contain all class probabilities between 0\u2013100%. To find similar images, we used the same CNN to extract the feature vector of the target image and compared it to feature vectors of images in the HAM10000 dataset via cosine similarity[20]. We stored the four closest images of each class and presented them in the AI-based CBIR decision support.\n\n**Interaction platform and raters.** _Online interaction platform._ The web-based platform DermaChallenge, which was developed at the Medical University of Vienna, served as the interface through which the performance of human raters and AI for the diagnostic task was evaluated and quantified. The platform is split into a back end and a front end, and both are deployed on a stack of well-known web technologies (Linux, Apache, MySQL and PHP). Please refer to the Nature Research Reporting Summary for details of the specific software versions used. The back end offers a representational state transfer interface to load and persist data, as well as JavaScript Object Notation web tokens to authenticate participants. The transport layer security and secure sockets layer protocol are used to encrypt all communications. The front end is optimized for mobile devices (mobile phones and tablets) but can also be used on any other platform via a JavaScript-enabled web browser. Before public deployment, five users tested the platform.\n\n_Recruitment and characteristics of raters._ We used mailing lists and social media posts of the International Society of Dermoscopy to recruit online raters. To participate in the study, raters had to register with a username, valid email address and password. In addition, we asked raters for details on their age (age groups spanning 10 years), gender, country, profession and years of experience in dermatoscopy ((1) less than 1 year, (2) opportunistic use for more than 1 year, (3) regular use for 1 to 5 years, (4) regular use for more than 5 years or (5) more than 10 years of experience). Each rater had to perform multiple screening tests to ensure that the self-reported experience matched actual skills. Screening tests consisted of simple domain-specific tasks, for example, to assign one of the seven possible diagnoses to ten cases, to separate melanomas from non-melanomas and to separate seborrheic keratoses from other lesions. We recruited 302 raters for the first interaction study that screened different forms of AI-based support, and 155 raters were recruited for the extended interaction study (inclusion of images with faulty AI-based support) and the telemedicine study (Supplementary Table 4). The distribution of raters according to task is presented in Supplementary Table 4. Second-opinion raters consisted of eight board-certified dermatologists and three dermatology residents, who were recruited because they diagnosed and managed more than two suspicious skin lesions on a face-to-face basis between April and September 2019. For the knowledge transfer study, we invited fourth-year medical students to participate; of the 650 medical students invited, 200 agreed to participate and 189 answered more than 50% of the test questions.\n\n_Characteristics of images and patients._ The benchmark test set of the ISIC 2018 challenge served as the sample for the interaction studies[9]. Of the 1,511 dermoscopic images in this set, 928 images were collected in the Department of Dermatology at the Medical University of Vienna, 267 images were collected in the skin cancer practice of Cliff Rosendahl in Queensland and the remaining 316 images were collected in other centers in Turkey (n = 117), New Zealand (n = 87), Sweden (n = 92) and Argentina (n = 20), to ensure diversity of skin types. The mean age of patients was 50.8 years (s.d. 17.4 years), and 46.2% of patients were female. The Austrian image set consists of lesions from patients referred to a tertiary European center specializing in the early detection of melanoma in high-risk groups. This group of patients is mainly of European ancestry and have a large number of nevi and skin types I\u2013III. The Australian image set includes lesions from patients of a primary-care facility in an area with a high incidence of skin cancer. Patients are typified by Celtic complexion, skin type I or II and chronic sun damage. Routine pathology evaluation (n = 786), biology (that is, >1.5 years of sequential dermoscopic imaging without changes; n = 458), expert consensus in common, straightforward, non-melanocytic cases that were not excised (n = 260) and in vivo confocal images (n = 7) served as the ground truth. Controversial cases with ambiguous histopathologic reports were excluded. Due to random sampling, only 1,412 of 1,511 images were finally used and evaluated by the raters. The 1,412 used cases consisted of 43 AKIECs, 93 BCCs, 217 BKLs, 44 DFs, 171 MELs, 809 NVs and 35 VASCs. For the telemedicine study, we included 93 of 98 participants (mean age 41.1 years (s.d. 12.2 years); 71% female) from the intervention arm of a recently conducted prospective randomized study[27] on mobile teledermoscopy for skin self-examinations. All 93 patients permitted reuse of their images. The participants had at least two skin cancer risk factors (light skin complexion and fair hair; skin that never or rarely tans and always or mostly burns; a family history of melanoma or a personal history of skin cancer, or many nevi; and residing in Queensland) as self-reported in the eligibility survey. A teledermoscopic evaluation was performed for all lesions. Face-to-face examination by an experienced board-certified dermatologist (H.P.S.) or the histopathologic report, in cases where the lesion was removed, served as the ground truth. The set of lesions consisted of 1,521 images of 596 lesions, including 29 AKIECs, 6 BCCs, 102 BKLs, 410 NVs, 2 squamous cell carcinomas (SCCs) and 9 VASCs. For calculation of diagnostic values, ground-truth data were mapped to classes of the HAM10000 dataset, if possible. We excluded nonspecific categories (n = 38 lesions) such as \u2018other\u2019, \u2018no lesion\u2019 or \u2018previously removed\u2019, because they could be mapped to neither the \u2018benign\u2019 nor \u2018malignant\u2019 category. The sample also included images that were not represented in the training data (non-pigmented variants of keratinocyte cancers, mucosal lesions and low-quality images), which were excluded from the telemedicine support study but not from the triage study, to better simulate a realistic scenario. For the second-opinion study, we searched the database of the Department of Dermatology at the Medical University of Vienna for dermoscopy images taken between April and September 2019. We included images if the lesion was excised and had a definite histopathologic diagnosis and if lesions were examined by a physician who was responsible for the face-to-face diagnosis of at least two other cases in this time period. The final sample set (n = 79) included 3 AKIECs, 23 BCCs, 13 BKLs, 2 DFs, 15 MELs, 21 NVs, 1 \u2018other\u2019 (scar) and 1 SCC. The mean age of patients was 64.6 years (s.d. 19.8 years), and 34.5% of patients were female. Patients were mainly of European ancestry and had skin type II (41.7%), III (57.1%) or IV (1.2%). As in the telemedicine scenario, we did not exclude images of categories that were not present in the training data or images of low quality. For the knowledge transfer study, the sample cases (n = 25) were randomly selected from the ISIC 2018 test set and stratified by diagnosis (6 AKIECs, 3 BCCs, 3 BKLs, 3 DFs, 3 MELs, 4 NVs and 3 VASCs).\n\n_Design of diagnostic studies._ To test the interaction of raters with different forms of AI-based decision support, we generated batches of 28 images. Each batch contained four randomly selected examples of every class. The raters\u2019 task was to diagnose the 28 unknown test images, first without and then with one type of decision support. We created a stratified randomization procedure to ensure a balanced distribution of the four types of decision support over all disease categories. The interaction study was online from 29 May 2019 to 15 January 2020. We excluded tests if the number of correct answers was lower than expected by chance to avoid noisy random data. We included only the first five tests for each rater to avoid biasing the results toward raters with high repetitions. The extended interaction study was open for participation between 15 January 2020 and 18 February 2020, presenting only multiclass probabilities as decision support. It included one image for every diagnosis from the ISIC 2018 test set with unaltered AI-based multiclass probabilities, two images with shuffled (resulting as incorrect) AI-based multiclass probabilities and eight images from the telemedicine study (see \u2018Characteristics of images and patients\u2019). The image sources were not disclosed to the raters. The second-opinion study was performed on a local web interface. Physicians who examined the patient face to face in real life were asked to reconsider their diagnosis and decisions with AI-based support. The case presentations included metadata (age, gender and localization), overview and close-up images (if available) and dermoscopic images. Physicians were not made aware that they had treated the patient before or of their previous decision on the case. Physicians were asked to provide their best diagnosis out of the seven predefined disease categories, as well as an extra category termed \u2018other\u2019, followed by their management decision (\u2018no intervention\u2019, \u2018monitor\u2019 or \u2018excise\u2019). No time constraints were set for this task. For the knowledge transfer study, we first examined the gradient-weighted class-activation maps[29], which were created for all images of the training set. We observed that the background attention of the CNN was significantly higher for predictions of the \u2018pigmented actinic keratosis\u2019 class than for other classes (P = 4.6 \u00d7 10[\u2212][12], two-sided unpaired t-test; Extended Data Fig. 2). We interpreted this finding as a diagnostic clue that points to the severely sun-damaged skin in the background of actinic keratoses, which is usually absent or not as severe in other disease categories. To test the hypothesis that teaching this clue to humans will improve their diagnostic skills, fourth-year medical students without previous knowledge of skin cancer detection received a 30-min introductory lecture about dermoscopy, and immediately thereafter students had to diagnose 25 test images (single best diagnosis). Answers were collected with a wireless audience response and voting system. Next, the lecturer presented an additional clue of \u2018sun-damaged skin in the surrounding skin of actinic keratoses\u2019 and the students repeated the test.\n\n**Statistics.** To simulate collective ratings of realistically small human groups (Fig. 3a), we confined the dataset to images with at least three distinct ratings (resulting range of ratings per image: 3\u201369). For each image, we created 30 bootstraps of three to five randomly selected ratings, whichever was the maximum available without replacement, and determined the most common rating as the prediction of the collective (first past the post). Ties were broken randomly. Next, we calculated the proportion of correct bootstrapped predictions to obtain the mean accuracy for each image as published previously[39]. To combine human collectives with CNN-based predictions, we took the arithmetic mean of the sum of the human multiclass probabilities, which were derived from the frequencies of bootstrapped human ratings, and the corresponding CNN-based multiclass probabilities. For analyses of diagnoses, we averaged the results for each image before comparisons; for analyses of raters with and without decision support, we calculated the arithmetic mean for each user before comparisons. The mean answering time for each user in every interaction modality served as a surrogate marker for confidence; answers that were faster or slower than the individual mean were regarded as \u2018confident\u2019 or \u2018non-confident\u2019, respectively. For the filtering procedure in the telemedicine study, we used a predefined cutoff of \u22650.17 to indicate malignancy, because this cutoff was selected by human raters in the interaction study (Extended Data Fig. 2). If patients photographed a lesion more than once, a single image above the cutoff was sufficient to label the lesion as \u2018probably malignant\u2019 and likewise on the patient level. We used a one-sample t-test to distinguish whether continuous data with normal distributions deviated from zero. Comparisons of continuous data between groups were performed with paired or unpaired t-tests or Wilcoxon signed-rank test, as appropriate. A chi-squared test was used to compare proportions. All reported P values were corrected for multiple testing (Holm\u2013Bonferroni[40]), and a two-sided P value < 0.05 was regarded as statistically significant. All analyses were performed using R v3.6.2 (ref. [41]), and plots were created with ggplot v3.2.1 (ref. [42]) and ggalluvial v0.11.1.\n\n**Reporting Summary.** Further information on research design is available in the Nature Research Reporting Summary linked to this article.",
        "results": "**Results**\n\nOur results show that decision support with AI-based multiclass probabilities improves the accuracy of human raters from 63.6% to 77.0% (increase of 13.3%, 95% CI 11.5% to 15.2%; P = 4.9 \u00d7 10[\u2212][35], two-sided paired t-test, t = 14.5, d.f. = 301; n = 302 raters), but no improvement was observed for decision support with AI-based prediction of malignancy or with our representation of AI-based CBIR. This suggests that the form of decision support should be in accordance with the given task. The probability of malignancy may be useful for simple binary management decisions such as whether to perform a biopsy or not, but not for a multiclass diagnostic problem. The studied form of AI-based CBIR is neither simple nor concrete; it needs more extensive cognitive engagement in terms of time and decision-making, because the rater needs to extrapolate the diagnosis from similarities between the test image and images with known diagnoses. The raters needed significantly more time to interact with AI-based CBIR decision support than with other types of support. Over time, human raters also tended to ignore the AI-based CBIR decision support. However, given that a large spectrum of CBIR approaches are described in the literature, another form of CBIR may still provide benefit. \n\nAfter we established that multiclass probabilities were the best form of CNN output for the given task, we focused on this form to explore the impact of AI-based support on human performance in more detail. We show an inverse relationship between the net gain from AI-based support and rater experience (Pearson\u2019s r = \u22120.18, 95% CI \u22120.28 to \u22120.07, P = 1.5 \u00d7 10[\u2212][2]; n = 302 raters). Raters in the least experienced group changed their initial diagnosis more often than experts (mean 26.0%, 95% CI 21.3% to 30.7% versus mean 14.7%, 95% CI 9.9% to 19.6%). Expert raters benefited only marginally (net gain 13.4%, 95% CI 6.3% to 20.6%) and only if they were not confident with their initial diagnosis, but not if they were confident. If experts were confident, they were usually correct and did not need support. \n\nThis finding suggests that, if experts have high confidence in their initial diagnosis, they should ignore AI-based support or not use it at all. The least experienced raters tended to accept AI-based support that contradicted their initial diagnosis even if they were confident. In general, raters changed their initial diagnosis less often if they were confident than if they were not confident in their decision (14.7%, 95% CI 12.6% to 16.8% versus 37.5%, 95% CI 34.0% to 41.0%; P = 1.9 \u00d7 10[\u2212][25], two-sided paired t-test; n = 302 raters). \n\nHaving established a positive impact of good quality AI-based support on diagnostic accuracy, we tested the impact of \u2018faulty\u2019 AI on diagnostic accuracy. Faulty AI could result from the application of AI algorithms on examples beyond the domain of images on which the AI was trained or the more remote possibility of adversarial attacks. To represent faulty AI, we intentionally generated misleading AI-based multiclass probabilities. If the top class probability of the CNN favored the correct diagnosis, we switched the probabilities in such a way that the CNN output favored a random incorrect diagnosis. We demonstrate that any previously observed gains in accuracy with AI-based support turn into a loss when that AI support is faulty. Our results suggest that, if raters build up the trust that is necessary to benefit from AI-based support, they are also vulnerable to perform below their expected ability if there is a fault with the AI. \n\nAnother finding of importance is that the benefit of human\u2013computer collaboration is asymmetrically distributed across disease categories. Our data showed that the net gain was higher for the class of pigmented actinic keratoses and intraepithelial carcinoma (increase of 31.5%, 95% CI 22.9% to 40.1%; n = 43 images) than for other categories. This suggests that the benefit of AI-based support needs to be adapted to the given task and the expected prevalence of target conditions. \n\nWe further demonstrate that AI-based multiclass ranking and probabilities have an impact on the raters\u2019 tendency to change their initial diagnosis. Most changes occurred in favor of the AI-predicted top category. Raters typically maintained their decisions that were in disagreement with the AI prediction only if that decision was ranked by AI prediction as at least the second or third option. Furthermore, raters tended to change their assessments more frequently when the difference in the AI-predicted probability between the initially selected category and the AI top category was high. \n\nAdditionally, we demonstrate that aggregated AI-based multiclass probabilities and crowd wisdom significantly increased the number of correct diagnoses in comparison to individual raters or AI in isolation. \n\nNext, we analyzed the impact of AI-based support in clinically relevant scenarios. To examine the potential of AI-based support in telemedicine, we reused prospectively collected images of a randomized controlled trial on self-examinations in high-risk patients. While the CNN was trained only on curated images of pigmented lesions, this sample also included non-pigmented variants of keratinocyte cancer, mucosal lesions and low-quality images. Although the proportion of correct specific diagnoses was significantly lower for these images (53.9% versus 76.2%; P = 8.9 \u00d7 10[\u2212][14], chi-squared test; n = 1,430 images), the CNN was able to recognize 95.2% of patients with skin cancer at a specificity of 59.2%. \n\nSimilarly to recent findings in AI-based breast cancer screening, our results indicate that AI-based skin cancer screening could triage high-risk cases and extend the intervals between face-to-face visits in low-risk cases. The optimal operating points to balance the potential benefits of AI-based triage with the risk of filtering out patients with skin cancer remain to be determined. \n\nIn another scenario, we asked dermatologists to rethink their face-to-face decisions in suspicious cases after providing them with AI-based multiclass probabilities, but without making them aware that they had previously managed the patient. With AI-based support, dermatologists switched from \u2018excision\u2019 to \u2018monitor\u2019 in 15.5% (7 of 45) of decisions for benign lesions, without increasing the number of malignant lesions that switched contrariwise. This result illustrates how human\u2013computer collaboration could decrease the number of unwarranted interventions and costs. AI-based support in this setting increased the frequency of correct specific diagnoses from 55.6% to 75.0%. \n\nFinally, we demonstrate that explanations for AI-based predictions can be translated into a human-understandable visual concept. By analyzing gradient-weighted class activation mapping (Grad-CAM), we show that attention of the CNN outside the object is higher for the prediction of actinic keratoses than for other categories. We hypothesize that, due to visual entrenchment, humans focus on the lesion and not on the background and frequently miss this clue. Here we show that teaching medical students to pay attention to chronic sun damage in the background improved the frequency of correct diagnoses of pigmented actinic keratoses from 32.5% to 47.3%. The overall frequency of correct diagnoses in all categories combined increased from 55.2% to 59.1%. \n\nThis study examines human\u2013computer collaboration from multiple angles and under varying conditions. We used the domain of skin cancer recognition for simplicity, but our study could serve as a framework for similar research in image-based diagnostic medicine. In contrast to the current narrative, our findings suggest that the primary focus should shift from human\u2013computer competition to human\u2013computer collaboration. From a regulatory perspective, the performance of AI-based systems should be tested under real-world conditions in the hands of the intended users and not as stand-alone devices. Only then can we expect to rationally adopt and improve AI-based decision support and to accelerate its evolution.",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Human-computer collaboration for skin cancer recognition.",
        "abstract": "The rapid increase in telemedicine coupled with recent advances in diagnostic artificial intelligence (AI) create the imperative to consider the opportunities and risks of inserting AI-based support into new paradigms of care. Here we build on recent achievements in the accuracy of image-based AI for skin cancer diagnosis to address the effects of varied representations of AI-based support across different levels of clinical expertise and multiple clinical workflows. We find that good quality AI-based support of clinical decision-making improves diagnostic accuracy over that of either AI or physicians alone, and that the least experienced clinicians gain the most from AI-based support. We further find that AI-based multiclass probabilities outperformed content-based image retrieval (CBIR) representations of AI in the mobile technology environment, and AI-based support had utility in simulations of second opinions and of telemedicine triage. In addition to demonstrating the potential benefits associated with good quality AI in the hands of non-expert clinicians, we find that faulty AI can mislead the entire spectrum of clinicians, including experts. Lastly, we show that insights derived from AI class-activation maps can inform improvements in human diagnosis. Together, our approach and findings offer a framework for future studies across the spectrum of image-based diagnostics to improve human-computer collaboration in clinical practice.\n",
        "methods": "**Methods**\n\n**Network training.** We fine-tuned a CNN for classification of seven different categories of the HAM10000 dataset[17]. We performed training on NVIDIA graphics processing units (GPUs) using the Pytorch[34] framework and chose a ResNet34 (ref. [35]) architecture, with weights initiated by pretraining on ImageNet[36] data. Cross-entropy served as the loss function, with weighting dependent on the frequency of classes within the dataset. The learning rate was initialized at 0.0001, with a tenfold reduction in case of no validation loss improvement for more than three epochs, but a minimum of 1 \u00d7 10[\u2212][9]. We used adaptive moment estimation (Adam[37]) as the optimizer and performed a maximum of 100 training epochs with early stopping. Images were presented in batches of 32, randomly cropped and resized to 224 \u00d7 224 pixels without normalization of a mean pixel, randomly rotated by 90 degrees and flipped with minor jitter in color, contrast, saturation and hue. The publicly available HAM10000 dataset, which corresponds to the training set of the ISIC 2018 challenge[18], was the source of images used for training and fivefold cross-validation. We selected the single best performing network on the hold-out validation set for further interaction with raters. For inference, images were cropped to 80% and resized to 224 \u00d7 224 pixels, with minor test-time augmentation consisting of horizontal flipping and rotation by 0 or 90 degrees. For the telemedicine dataset, we also applied color normalization via Shades of Gray[38] with a Minkowski norm of 6. The multiclass probabilities presented to the raters were obtained by applying a softmax function to contain all class probabilities between 0\u2013100%. To find similar images, we used the same CNN to extract the feature vector of the target image and compared it to feature vectors of images in the HAM10000 dataset via cosine similarity[20]. We stored the four closest images of each class and presented them in the AI-based CBIR decision support.\n\n**Interaction platform and raters.** _Online interaction platform._ The web-based platform DermaChallenge, which was developed at the Medical University of Vienna, served as the interface through which the performance of human raters and AI for the diagnostic task was evaluated and quantified. The platform is split into a back end and a front end, and both are deployed on a stack of well-known web technologies (Linux, Apache, MySQL and PHP). Please refer to the Nature Research Reporting Summary for details of the specific software versions used. The back end offers a representational state transfer interface to load and persist data, as well as JavaScript Object Notation web tokens to authenticate participants. The transport layer security and secure sockets layer protocol are used to encrypt all communications. The front end is optimized for mobile devices (mobile phones and tablets) but can also be used on any other platform via a JavaScript-enabled web browser. Before public deployment, five users tested the platform.\n\n_Recruitment and characteristics of raters._ We used mailing lists and social media posts of the International Society of Dermoscopy to recruit online raters. To participate in the study, raters had to register with a username, valid email address and password. In addition, we asked raters for details on their age (age groups spanning 10 years), gender, country, profession and years of experience in dermatoscopy ((1) less than 1 year, (2) opportunistic use for more than 1 year, (3) regular use for 1 to 5 years, (4) regular use for more than 5 years or (5) more than 10 years of experience). Each rater had to perform multiple screening tests to ensure that the self-reported experience matched actual skills. Screening tests consisted of simple domain-specific tasks, for example, to assign one of the seven possible diagnoses to ten cases, to separate melanomas from non-melanomas and to separate seborrheic keratoses from other lesions. We recruited 302 raters for the first interaction study that screened different forms of AI-based support, and 155 raters were recruited for the extended interaction study (inclusion of images with faulty AI-based support) and the telemedicine study (Supplementary Table 4). The distribution of raters according to task is presented in Supplementary Table 4. Second-opinion raters consisted of eight board-certified dermatologists and three dermatology residents, who were recruited because they diagnosed and managed more than two suspicious skin lesions on a face-to-face basis between April and September 2019. For the knowledge transfer study, we invited fourth-year medical students to participate; of the 650 medical students invited, 200 agreed to participate and 189 answered more than 50% of the test questions.\n\n_Characteristics of images and patients._ The benchmark test set of the ISIC 2018 challenge served as the sample for the interaction studies[9]. Of the 1,511 dermoscopic images in this set, 928 images were collected in the Department of Dermatology at the Medical University of Vienna, 267 images were collected in the skin cancer practice of Cliff Rosendahl in Queensland and the remaining 316 images were collected in other centers in Turkey (n = 117), New Zealand (n = 87), Sweden (n = 92) and Argentina (n = 20), to ensure diversity of skin types. The mean age of patients was 50.8 years (s.d. 17.4 years), and 46.2% of patients were female. The Austrian image set consists of lesions from patients referred to a tertiary European center specializing in the early detection of melanoma in high-risk groups. This group of patients is mainly of European ancestry and have a large number of nevi and skin types I\u2013III. The Australian image set includes lesions from patients of a primary-care facility in an area with a high incidence of skin cancer. Patients are typified by Celtic complexion, skin type I or II and chronic sun damage. Routine pathology evaluation (n = 786), biology (that is, >1.5 years of sequential dermoscopic imaging without changes; n = 458), expert consensus in common, straightforward, non-melanocytic cases that were not excised (n = 260) and in vivo confocal images (n = 7) served as the ground truth. Controversial cases with ambiguous histopathologic reports were excluded. Due to random sampling, only 1,412 of 1,511 images were finally used and evaluated by the raters. The 1,412 used cases consisted of 43 AKIECs, 93 BCCs, 217 BKLs, 44 DFs, 171 MELs, 809 NVs and 35 VASCs. For the telemedicine study, we included 93 of 98 participants (mean age 41.1 years (s.d. 12.2 years); 71% female) from the intervention arm of a recently conducted prospective randomized study[27] on mobile teledermoscopy for skin self-examinations. All 93 patients permitted reuse of their images. The participants had at least two skin cancer risk factors (light skin complexion and fair hair; skin that never or rarely tans and always or mostly burns; a family history of melanoma or a personal history of skin cancer, or many nevi; and residing in Queensland) as self-reported in the eligibility survey. A teledermoscopic evaluation was performed for all lesions. Face-to-face examination by an experienced board-certified dermatologist (H.P.S.) or the histopathologic report, in cases where the lesion was removed, served as the ground truth. The set of lesions consisted of 1,521 images of 596 lesions, including 29 AKIECs, 6 BCCs, 102 BKLs, 410 NVs, 2 squamous cell carcinomas (SCCs) and 9 VASCs. For calculation of diagnostic values, ground-truth data were mapped to classes of the HAM10000 dataset, if possible. We excluded nonspecific categories (n = 38 lesions) such as \u2018other\u2019, \u2018no lesion\u2019 or \u2018previously removed\u2019, because they could be mapped to neither the \u2018benign\u2019 nor \u2018malignant\u2019 category. The sample also included images that were not represented in the training data (non-pigmented variants of keratinocyte cancers, mucosal lesions and low-quality images), which were excluded from the telemedicine support study but not from the triage study, to better simulate a realistic scenario. For the second-opinion study, we searched the database of the Department of Dermatology at the Medical University of Vienna for dermoscopy images taken between April and September 2019. We included images if the lesion was excised and had a definite histopathologic diagnosis and if lesions were examined by a physician who was responsible for the face-to-face diagnosis of at least two other cases in this time period. The final sample set (n = 79) included 3 AKIECs, 23 BCCs, 13 BKLs, 2 DFs, 15 MELs, 21 NVs, 1 \u2018other\u2019 (scar) and 1 SCC. The mean age of patients was 64.6 years (s.d. 19.8 years), and 34.5% of patients were female. Patients were mainly of European ancestry and had skin type II (41.7%), III (57.1%) or IV (1.2%). As in the telemedicine scenario, we did not exclude images of categories that were not present in the training data or images of low quality. For the knowledge transfer study, the sample cases (n = 25) were randomly selected from the ISIC 2018 test set and stratified by diagnosis (6 AKIECs, 3 BCCs, 3 BKLs, 3 DFs, 3 MELs, 4 NVs and 3 VASCs).\n\n_Design of diagnostic studies._ To test the interaction of raters with different forms of AI-based decision support, we generated batches of 28 images. Each batch contained four randomly selected examples of every class. The raters\u2019 task was to diagnose the 28 unknown test images, first without and then with one type of decision support. We created a stratified randomization procedure to ensure a balanced distribution of the four types of decision support over all disease categories. The interaction study was online from 29 May 2019 to 15 January 2020. We excluded tests if the number of correct answers was lower than expected by chance to avoid noisy random data. We included only the first five tests for each rater to avoid biasing the results toward raters with high repetitions. The extended interaction study was open for participation between 15 January 2020 and 18 February 2020, presenting only multiclass probabilities as decision support. It included one image for every diagnosis from the ISIC 2018 test set with unaltered AI-based multiclass probabilities, two images with shuffled (resulting as incorrect) AI-based multiclass probabilities and eight images from the telemedicine study (see \u2018Characteristics of images and patients\u2019). The image sources were not disclosed to the raters. The second-opinion study was performed on a local web interface. Physicians who examined the patient face to face in real life were asked to reconsider their diagnosis and decisions with AI-based support. The case presentations included metadata (age, gender and localization), overview and close-up images (if available) and dermoscopic images. Physicians were not made aware that they had treated the patient before or of their previous decision on the case. Physicians were asked to provide their best diagnosis out of the seven predefined disease categories, as well as an extra category termed \u2018other\u2019, followed by their management decision (\u2018no intervention\u2019, \u2018monitor\u2019 or \u2018excise\u2019). No time constraints were set for this task. For the knowledge transfer study, we first examined the gradient-weighted class-activation maps[29], which were created for all images of the training set. We observed that the background attention of the CNN was significantly higher for predictions of the \u2018pigmented actinic keratosis\u2019 class than for other classes (P = 4.6 \u00d7 10[\u2212][12], two-sided unpaired t-test; Extended Data Fig. 2). We interpreted this finding as a diagnostic clue that points to the severely sun-damaged skin in the background of actinic keratoses, which is usually absent or not as severe in other disease categories. To test the hypothesis that teaching this clue to humans will improve their diagnostic skills, fourth-year medical students without previous knowledge of skin cancer detection received a 30-min introductory lecture about dermoscopy, and immediately thereafter students had to diagnose 25 test images (single best diagnosis). Answers were collected with a wireless audience response and voting system. Next, the lecturer presented an additional clue of \u2018sun-damaged skin in the surrounding skin of actinic keratoses\u2019 and the students repeated the test.\n\n**Statistics.** To simulate collective ratings of realistically small human groups (Fig. 3a), we confined the dataset to images with at least three distinct ratings (resulting range of ratings per image: 3\u201369). For each image, we created 30 bootstraps of three to five randomly selected ratings, whichever was the maximum available without replacement, and determined the most common rating as the prediction of the collective (first past the post). Ties were broken randomly. Next, we calculated the proportion of correct bootstrapped predictions to obtain the mean accuracy for each image as published previously[39]. To combine human collectives with CNN-based predictions, we took the arithmetic mean of the sum of the human multiclass probabilities, which were derived from the frequencies of bootstrapped human ratings, and the corresponding CNN-based multiclass probabilities. For analyses of diagnoses, we averaged the results for each image before comparisons; for analyses of raters with and without decision support, we calculated the arithmetic mean for each user before comparisons. The mean answering time for each user in every interaction modality served as a surrogate marker for confidence; answers that were faster or slower than the individual mean were regarded as \u2018confident\u2019 or \u2018non-confident\u2019, respectively. For the filtering procedure in the telemedicine study, we used a predefined cutoff of \u22650.17 to indicate malignancy, because this cutoff was selected by human raters in the interaction study (Extended Data Fig. 2). If patients photographed a lesion more than once, a single image above the cutoff was sufficient to label the lesion as \u2018probably malignant\u2019 and likewise on the patient level. We used a one-sample t-test to distinguish whether continuous data with normal distributions deviated from zero. Comparisons of continuous data between groups were performed with paired or unpaired t-tests or Wilcoxon signed-rank test, as appropriate. A chi-squared test was used to compare proportions. All reported P values were corrected for multiple testing (Holm\u2013Bonferroni[40]), and a two-sided P value < 0.05 was regarded as statistically significant. All analyses were performed using R v3.6.2 (ref. [41]), and plots were created with ggplot v3.2.1 (ref. [42]) and ggalluvial v0.11.1.\n\n**Reporting Summary.** Further information on research design is available in the Nature Research Reporting Summary linked to this article.",
        "results": "**Results**\n\nOur results show that decision support with AI-based multiclass probabilities improves the accuracy of human raters from 63.6% to 77.0% (increase of 13.3%, 95% CI 11.5% to 15.2%; P = 4.9 \u00d7 10[\u2212][35], two-sided paired t-test, t = 14.5, d.f. = 301; n = 302 raters), but no improvement was observed for decision support with AI-based prediction of malignancy or with our representation of AI-based CBIR. This suggests that the form of decision support should be in accordance with the given task. The probability of malignancy may be useful for simple binary management decisions such as whether to perform a biopsy or not, but not for a multiclass diagnostic problem. The studied form of AI-based CBIR is neither simple nor concrete; it needs more extensive cognitive engagement in terms of time and decision-making, because the rater needs to extrapolate the diagnosis from similarities between the test image and images with known diagnoses. The raters needed significantly more time to interact with AI-based CBIR decision support than with other types of support. Over time, human raters also tended to ignore the AI-based CBIR decision support. However, given that a large spectrum of CBIR approaches are described in the literature, another form of CBIR may still provide benefit. \n\nAfter we established that multiclass probabilities were the best form of CNN output for the given task, we focused on this form to explore the impact of AI-based support on human performance in more detail. We show an inverse relationship between the net gain from AI-based support and rater experience (Pearson\u2019s r = \u22120.18, 95% CI \u22120.28 to \u22120.07, P = 1.5 \u00d7 10[\u2212][2]; n = 302 raters). Raters in the least experienced group changed their initial diagnosis more often than experts (mean 26.0%, 95% CI 21.3% to 30.7% versus mean 14.7%, 95% CI 9.9% to 19.6%). Expert raters benefited only marginally (net gain 13.4%, 95% CI 6.3% to 20.6%) and only if they were not confident with their initial diagnosis, but not if they were confident. If experts were confident, they were usually correct and did not need support. \n\nThis finding suggests that, if experts have high confidence in their initial diagnosis, they should ignore AI-based support or not use it at all. The least experienced raters tended to accept AI-based support that contradicted their initial diagnosis even if they were confident. In general, raters changed their initial diagnosis less often if they were confident than if they were not confident in their decision (14.7%, 95% CI 12.6% to 16.8% versus 37.5%, 95% CI 34.0% to 41.0%; P = 1.9 \u00d7 10[\u2212][25], two-sided paired t-test; n = 302 raters). \n\nHaving established a positive impact of good quality AI-based support on diagnostic accuracy, we tested the impact of \u2018faulty\u2019 AI on diagnostic accuracy. Faulty AI could result from the application of AI algorithms on examples beyond the domain of images on which the AI was trained or the more remote possibility of adversarial attacks. To represent faulty AI, we intentionally generated misleading AI-based multiclass probabilities. If the top class probability of the CNN favored the correct diagnosis, we switched the probabilities in such a way that the CNN output favored a random incorrect diagnosis. We demonstrate that any previously observed gains in accuracy with AI-based support turn into a loss when that AI support is faulty. Our results suggest that, if raters build up the trust that is necessary to benefit from AI-based support, they are also vulnerable to perform below their expected ability if there is a fault with the AI. \n\nAnother finding of importance is that the benefit of human\u2013computer collaboration is asymmetrically distributed across disease categories. Our data showed that the net gain was higher for the class of pigmented actinic keratoses and intraepithelial carcinoma (increase of 31.5%, 95% CI 22.9% to 40.1%; n = 43 images) than for other categories. This suggests that the benefit of AI-based support needs to be adapted to the given task and the expected prevalence of target conditions. \n\nWe further demonstrate that AI-based multiclass ranking and probabilities have an impact on the raters\u2019 tendency to change their initial diagnosis. Most changes occurred in favor of the AI-predicted top category. Raters typically maintained their decisions that were in disagreement with the AI prediction only if that decision was ranked by AI prediction as at least the second or third option. Furthermore, raters tended to change their assessments more frequently when the difference in the AI-predicted probability between the initially selected category and the AI top category was high. \n\nAdditionally, we demonstrate that aggregated AI-based multiclass probabilities and crowd wisdom significantly increased the number of correct diagnoses in comparison to individual raters or AI in isolation. \n\nNext, we analyzed the impact of AI-based support in clinically relevant scenarios. To examine the potential of AI-based support in telemedicine, we reused prospectively collected images of a randomized controlled trial on self-examinations in high-risk patients. While the CNN was trained only on curated images of pigmented lesions, this sample also included non-pigmented variants of keratinocyte cancer, mucosal lesions and low-quality images. Although the proportion of correct specific diagnoses was significantly lower for these images (53.9% versus 76.2%; P = 8.9 \u00d7 10[\u2212][14], chi-squared test; n = 1,430 images), the CNN was able to recognize 95.2% of patients with skin cancer at a specificity of 59.2%. \n\nSimilarly to recent findings in AI-based breast cancer screening, our results indicate that AI-based skin cancer screening could triage high-risk cases and extend the intervals between face-to-face visits in low-risk cases. The optimal operating points to balance the potential benefits of AI-based triage with the risk of filtering out patients with skin cancer remain to be determined. \n\nIn another scenario, we asked dermatologists to rethink their face-to-face decisions in suspicious cases after providing them with AI-based multiclass probabilities, but without making them aware that they had previously managed the patient. With AI-based support, dermatologists switched from \u2018excision\u2019 to \u2018monitor\u2019 in 15.5% (7 of 45) of decisions for benign lesions, without increasing the number of malignant lesions that switched contrariwise. This result illustrates how human\u2013computer collaboration could decrease the number of unwarranted interventions and costs. AI-based support in this setting increased the frequency of correct specific diagnoses from 55.6% to 75.0%. \n\nFinally, we demonstrate that explanations for AI-based predictions can be translated into a human-understandable visual concept. By analyzing gradient-weighted class activation mapping (Grad-CAM), we show that attention of the CNN outside the object is higher for the prediction of actinic keratoses than for other categories. We hypothesize that, due to visual entrenchment, humans focus on the lesion and not on the background and frequently miss this clue. Here we show that teaching medical students to pay attention to chronic sun damage in the background improved the frequency of correct diagnoses of pigmented actinic keratoses from 32.5% to 47.3%. The overall frequency of correct diagnoses in all categories combined increased from 55.2% to 59.1%. \n\nThis study examines human\u2013computer collaboration from multiple angles and under varying conditions. We used the domain of skin cancer recognition for simplicity, but our study could serve as a framework for similar research in image-based diagnostic medicine. In contrast to the current narrative, our findings suggest that the primary focus should shift from human\u2013computer competition to human\u2013computer collaboration. From a regulatory perspective, the performance of AI-based systems should be tested under real-world conditions in the hands of the intended users and not as stand-alone devices. Only then can we expect to rationally adopt and improve AI-based decision support and to accelerate its evolution.",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Human-computer collaboration for skin cancer recognition.",
        "abstract": "The rapid increase in telemedicine coupled with recent advances in diagnostic artificial intelligence (AI) create the imperative to consider the opportunities and risks of inserting AI-based support into new paradigms of care. Here we build on recent achievements in the accuracy of image-based AI for skin cancer diagnosis to address the effects of varied representations of AI-based support across different levels of clinical expertise and multiple clinical workflows. We find that good quality AI-based support of clinical decision-making improves diagnostic accuracy over that of either AI or physicians alone, and that the least experienced clinicians gain the most from AI-based support. We further find that AI-based multiclass probabilities outperformed content-based image retrieval (CBIR) representations of AI in the mobile technology environment, and AI-based support had utility in simulations of second opinions and of telemedicine triage. In addition to demonstrating the potential benefits associated with good quality AI in the hands of non-expert clinicians, we find that faulty AI can mislead the entire spectrum of clinicians, including experts. Lastly, we show that insights derived from AI class-activation maps can inform improvements in human diagnosis. Together, our approach and findings offer a framework for future studies across the spectrum of image-based diagnostics to improve human-computer collaboration in clinical practice.\n",
        "methods": "**Methods**\n\n**Network training.** We fine-tuned a CNN for classification of seven different categories of the HAM10000 dataset[17]. We performed training on NVIDIA graphics processing units (GPUs) using the Pytorch[34] framework and chose a ResNet34 (ref. [35]) architecture, with weights initiated by pretraining on ImageNet[36] data. Cross-entropy served as the loss function, with weighting dependent on the frequency of classes within the dataset. The learning rate was initialized at 0.0001, with a tenfold reduction in case of no validation loss improvement for more than three epochs, but a minimum of 1 \u00d7 10[\u2212][9]. We used adaptive moment estimation (Adam[37]) as the optimizer and performed a maximum of 100 training epochs with early stopping. Images were presented in batches of 32, randomly cropped and resized to 224 \u00d7 224 pixels without normalization of a mean pixel, randomly rotated by 90 degrees and flipped with minor jitter in color, contrast, saturation and hue. The publicly available HAM10000 dataset, which corresponds to the training set of the ISIC 2018 challenge[18], was the source of images used for training and fivefold cross-validation. We selected the single best performing network on the hold-out validation set for further interaction with raters. For inference, images were cropped to 80% and resized to 224 \u00d7 224 pixels, with minor test-time augmentation consisting of horizontal flipping and rotation by 0 or 90 degrees. For the telemedicine dataset, we also applied color normalization via Shades of Gray[38] with a Minkowski norm of 6. The multiclass probabilities presented to the raters were obtained by applying a softmax function to contain all class probabilities between 0\u2013100%. To find similar images, we used the same CNN to extract the feature vector of the target image and compared it to feature vectors of images in the HAM10000 dataset via cosine similarity[20]. We stored the four closest images of each class and presented them in the AI-based CBIR decision support.\n\n**Interaction platform and raters.** _Online interaction platform._ The web-based platform DermaChallenge, which was developed at the Medical University of Vienna, served as the interface through which the performance of human raters and AI for the diagnostic task was evaluated and quantified. The platform is split into a back end and a front end, and both are deployed on a stack of well-known web technologies (Linux, Apache, MySQL and PHP). Please refer to the Nature Research Reporting Summary for details of the specific software versions used. The back end offers a representational state transfer interface to load and persist data, as well as JavaScript Object Notation web tokens to authenticate participants. The transport layer security and secure sockets layer protocol are used to encrypt all communications. The front end is optimized for mobile devices (mobile phones and tablets) but can also be used on any other platform via a JavaScript-enabled web browser. Before public deployment, five users tested the platform.\n\n_Recruitment and characteristics of raters._ We used mailing lists and social media posts of the International Society of Dermoscopy to recruit online raters. To participate in the study, raters had to register with a username, valid email address and password. In addition, we asked raters for details on their age (age groups spanning 10 years), gender, country, profession and years of experience in dermatoscopy ((1) less than 1 year, (2) opportunistic use for more than 1 year, (3) regular use for 1 to 5 years, (4) regular use for more than 5 years or (5) more than 10 years of experience). Each rater had to perform multiple screening tests to ensure that the self-reported experience matched actual skills. Screening tests consisted of simple domain-specific tasks, for example, to assign one of the seven possible diagnoses to ten cases, to separate melanomas from non-melanomas and to separate seborrheic keratoses from other lesions. We recruited 302 raters for the first interaction study that screened different forms of AI-based support, and 155 raters were recruited for the extended interaction study (inclusion of images with faulty AI-based support) and the telemedicine study (Supplementary Table 4). The distribution of raters according to task is presented in Supplementary Table 4. Second-opinion raters consisted of eight board-certified dermatologists and three dermatology residents, who were recruited because they diagnosed and managed more than two suspicious skin lesions on a face-to-face basis between April and September 2019. For the knowledge transfer study, we invited fourth-year medical students to participate; of the 650 medical students invited, 200 agreed to participate and 189 answered more than 50% of the test questions.\n\n_Characteristics of images and patients._ The benchmark test set of the ISIC 2018 challenge served as the sample for the interaction studies[9]. Of the 1,511 dermoscopic images in this set, 928 images were collected in the Department of Dermatology at the Medical University of Vienna, 267 images were collected in the skin cancer practice of Cliff Rosendahl in Queensland and the remaining 316 images were collected in other centers in Turkey (n = 117), New Zealand (n = 87), Sweden (n = 92) and Argentina (n = 20), to ensure diversity of skin types. The mean age of patients was 50.8 years (s.d. 17.4 years), and 46.2% of patients were female. The Austrian image set consists of lesions from patients referred to a tertiary European center specializing in the early detection of melanoma in high-risk groups. This group of patients is mainly of European ancestry and have a large number of nevi and skin types I\u2013III. The Australian image set includes lesions from patients of a primary-care facility in an area with a high incidence of skin cancer. Patients are typified by Celtic complexion, skin type I or II and chronic sun damage. Routine pathology evaluation (n = 786), biology (that is, >1.5 years of sequential dermoscopic imaging without changes; n = 458), expert consensus in common, straightforward, non-melanocytic cases that were not excised (n = 260) and in vivo confocal images (n = 7) served as the ground truth. Controversial cases with ambiguous histopathologic reports were excluded. Due to random sampling, only 1,412 of 1,511 images were finally used and evaluated by the raters. The 1,412 used cases consisted of 43 AKIECs, 93 BCCs, 217 BKLs, 44 DFs, 171 MELs, 809 NVs and 35 VASCs. For the telemedicine study, we included 93 of 98 participants (mean age 41.1 years (s.d. 12.2 years); 71% female) from the intervention arm of a recently conducted prospective randomized study[27] on mobile teledermoscopy for skin self-examinations. All 93 patients permitted reuse of their images. The participants had at least two skin cancer risk factors (light skin complexion and fair hair; skin that never or rarely tans and always or mostly burns; a family history of melanoma or a personal history of skin cancer, or many nevi; and residing in Queensland) as self-reported in the eligibility survey. A teledermoscopic evaluation was performed for all lesions. Face-to-face examination by an experienced board-certified dermatologist (H.P.S.) or the histopathologic report, in cases where the lesion was removed, served as the ground truth. The set of lesions consisted of 1,521 images of 596 lesions, including 29 AKIECs, 6 BCCs, 102 BKLs, 410 NVs, 2 squamous cell carcinomas (SCCs) and 9 VASCs. For calculation of diagnostic values, ground-truth data were mapped to classes of the HAM10000 dataset, if possible. We excluded nonspecific categories (n = 38 lesions) such as \u2018other\u2019, \u2018no lesion\u2019 or \u2018previously removed\u2019, because they could be mapped to neither the \u2018benign\u2019 nor \u2018malignant\u2019 category. The sample also included images that were not represented in the training data (non-pigmented variants of keratinocyte cancers, mucosal lesions and low-quality images), which were excluded from the telemedicine support study but not from the triage study, to better simulate a realistic scenario. For the second-opinion study, we searched the database of the Department of Dermatology at the Medical University of Vienna for dermoscopy images taken between April and September 2019. We included images if the lesion was excised and had a definite histopathologic diagnosis and if lesions were examined by a physician who was responsible for the face-to-face diagnosis of at least two other cases in this time period. The final sample set (n = 79) included 3 AKIECs, 23 BCCs, 13 BKLs, 2 DFs, 15 MELs, 21 NVs, 1 \u2018other\u2019 (scar) and 1 SCC. The mean age of patients was 64.6 years (s.d. 19.8 years), and 34.5% of patients were female. Patients were mainly of European ancestry and had skin type II (41.7%), III (57.1%) or IV (1.2%). As in the telemedicine scenario, we did not exclude images of categories that were not present in the training data or images of low quality. For the knowledge transfer study, the sample cases (n = 25) were randomly selected from the ISIC 2018 test set and stratified by diagnosis (6 AKIECs, 3 BCCs, 3 BKLs, 3 DFs, 3 MELs, 4 NVs and 3 VASCs).\n\n_Design of diagnostic studies._ To test the interaction of raters with different forms of AI-based decision support, we generated batches of 28 images. Each batch contained four randomly selected examples of every class. The raters\u2019 task was to diagnose the 28 unknown test images, first without and then with one type of decision support. We created a stratified randomization procedure to ensure a balanced distribution of the four types of decision support over all disease categories. The interaction study was online from 29 May 2019 to 15 January 2020. We excluded tests if the number of correct answers was lower than expected by chance to avoid noisy random data. We included only the first five tests for each rater to avoid biasing the results toward raters with high repetitions. The extended interaction study was open for participation between 15 January 2020 and 18 February 2020, presenting only multiclass probabilities as decision support. It included one image for every diagnosis from the ISIC 2018 test set with unaltered AI-based multiclass probabilities, two images with shuffled (resulting as incorrect) AI-based multiclass probabilities and eight images from the telemedicine study (see \u2018Characteristics of images and patients\u2019). The image sources were not disclosed to the raters. The second-opinion study was performed on a local web interface. Physicians who examined the patient face to face in real life were asked to reconsider their diagnosis and decisions with AI-based support. The case presentations included metadata (age, gender and localization), overview and close-up images (if available) and dermoscopic images. Physicians were not made aware that they had treated the patient before or of their previous decision on the case. Physicians were asked to provide their best diagnosis out of the seven predefined disease categories, as well as an extra category termed \u2018other\u2019, followed by their management decision (\u2018no intervention\u2019, \u2018monitor\u2019 or \u2018excise\u2019). No time constraints were set for this task. For the knowledge transfer study, we first examined the gradient-weighted class-activation maps[29], which were created for all images of the training set. We observed that the background attention of the CNN was significantly higher for predictions of the \u2018pigmented actinic keratosis\u2019 class than for other classes (P = 4.6 \u00d7 10[\u2212][12], two-sided unpaired t-test; Extended Data Fig. 2). We interpreted this finding as a diagnostic clue that points to the severely sun-damaged skin in the background of actinic keratoses, which is usually absent or not as severe in other disease categories. To test the hypothesis that teaching this clue to humans will improve their diagnostic skills, fourth-year medical students without previous knowledge of skin cancer detection received a 30-min introductory lecture about dermoscopy, and immediately thereafter students had to diagnose 25 test images (single best diagnosis). Answers were collected with a wireless audience response and voting system. Next, the lecturer presented an additional clue of \u2018sun-damaged skin in the surrounding skin of actinic keratoses\u2019 and the students repeated the test.\n\n**Statistics.** To simulate collective ratings of realistically small human groups (Fig. 3a), we confined the dataset to images with at least three distinct ratings (resulting range of ratings per image: 3\u201369). For each image, we created 30 bootstraps of three to five randomly selected ratings, whichever was the maximum available without replacement, and determined the most common rating as the prediction of the collective (first past the post). Ties were broken randomly. Next, we calculated the proportion of correct bootstrapped predictions to obtain the mean accuracy for each image as published previously[39]. To combine human collectives with CNN-based predictions, we took the arithmetic mean of the sum of the human multiclass probabilities, which were derived from the frequencies of bootstrapped human ratings, and the corresponding CNN-based multiclass probabilities. For analyses of diagnoses, we averaged the results for each image before comparisons; for analyses of raters with and without decision support, we calculated the arithmetic mean for each user before comparisons. The mean answering time for each user in every interaction modality served as a surrogate marker for confidence; answers that were faster or slower than the individual mean were regarded as \u2018confident\u2019 or \u2018non-confident\u2019, respectively. For the filtering procedure in the telemedicine study, we used a predefined cutoff of \u22650.17 to indicate malignancy, because this cutoff was selected by human raters in the interaction study (Extended Data Fig. 2). If patients photographed a lesion more than once, a single image above the cutoff was sufficient to label the lesion as \u2018probably malignant\u2019 and likewise on the patient level. We used a one-sample t-test to distinguish whether continuous data with normal distributions deviated from zero. Comparisons of continuous data between groups were performed with paired or unpaired t-tests or Wilcoxon signed-rank test, as appropriate. A chi-squared test was used to compare proportions. All reported P values were corrected for multiple testing (Holm\u2013Bonferroni[40]), and a two-sided P value < 0.05 was regarded as statistically significant. All analyses were performed using R v3.6.2 (ref. [41]), and plots were created with ggplot v3.2.1 (ref. [42]) and ggalluvial v0.11.1.\n\n**Reporting Summary.** Further information on research design is available in the Nature Research Reporting Summary linked to this article.",
        "results": "**Results**\n\nOur results show that decision support with AI-based multiclass probabilities improves the accuracy of human raters from 63.6% to 77.0% (increase of 13.3%, 95% CI 11.5% to 15.2%; P = 4.9 \u00d7 10[\u2212][35], two-sided paired t-test, t = 14.5, d.f. = 301; n = 302 raters), but no improvement was observed for decision support with AI-based prediction of malignancy or with our representation of AI-based CBIR. This suggests that the form of decision support should be in accordance with the given task. The probability of malignancy may be useful for simple binary management decisions such as whether to perform a biopsy or not, but not for a multiclass diagnostic problem. The studied form of AI-based CBIR is neither simple nor concrete; it needs more extensive cognitive engagement in terms of time and decision-making, because the rater needs to extrapolate the diagnosis from similarities between the test image and images with known diagnoses. The raters needed significantly more time to interact with AI-based CBIR decision support than with other types of support. Over time, human raters also tended to ignore the AI-based CBIR decision support. However, given that a large spectrum of CBIR approaches are described in the literature, another form of CBIR may still provide benefit. \n\nAfter we established that multiclass probabilities were the best form of CNN output for the given task, we focused on this form to explore the impact of AI-based support on human performance in more detail. We show an inverse relationship between the net gain from AI-based support and rater experience (Pearson\u2019s r = \u22120.18, 95% CI \u22120.28 to \u22120.07, P = 1.5 \u00d7 10[\u2212][2]; n = 302 raters). Raters in the least experienced group changed their initial diagnosis more often than experts (mean 26.0%, 95% CI 21.3% to 30.7% versus mean 14.7%, 95% CI 9.9% to 19.6%). Expert raters benefited only marginally (net gain 13.4%, 95% CI 6.3% to 20.6%) and only if they were not confident with their initial diagnosis, but not if they were confident. If experts were confident, they were usually correct and did not need support. \n\nThis finding suggests that, if experts have high confidence in their initial diagnosis, they should ignore AI-based support or not use it at all. The least experienced raters tended to accept AI-based support that contradicted their initial diagnosis even if they were confident. In general, raters changed their initial diagnosis less often if they were confident than if they were not confident in their decision (14.7%, 95% CI 12.6% to 16.8% versus 37.5%, 95% CI 34.0% to 41.0%; P = 1.9 \u00d7 10[\u2212][25], two-sided paired t-test; n = 302 raters). \n\nHaving established a positive impact of good quality AI-based support on diagnostic accuracy, we tested the impact of \u2018faulty\u2019 AI on diagnostic accuracy. Faulty AI could result from the application of AI algorithms on examples beyond the domain of images on which the AI was trained or the more remote possibility of adversarial attacks. To represent faulty AI, we intentionally generated misleading AI-based multiclass probabilities. If the top class probability of the CNN favored the correct diagnosis, we switched the probabilities in such a way that the CNN output favored a random incorrect diagnosis. We demonstrate that any previously observed gains in accuracy with AI-based support turn into a loss when that AI support is faulty. Our results suggest that, if raters build up the trust that is necessary to benefit from AI-based support, they are also vulnerable to perform below their expected ability if there is a fault with the AI. \n\nAnother finding of importance is that the benefit of human\u2013computer collaboration is asymmetrically distributed across disease categories. Our data showed that the net gain was higher for the class of pigmented actinic keratoses and intraepithelial carcinoma (increase of 31.5%, 95% CI 22.9% to 40.1%; n = 43 images) than for other categories. This suggests that the benefit of AI-based support needs to be adapted to the given task and the expected prevalence of target conditions. \n\nWe further demonstrate that AI-based multiclass ranking and probabilities have an impact on the raters\u2019 tendency to change their initial diagnosis. Most changes occurred in favor of the AI-predicted top category. Raters typically maintained their decisions that were in disagreement with the AI prediction only if that decision was ranked by AI prediction as at least the second or third option. Furthermore, raters tended to change their assessments more frequently when the difference in the AI-predicted probability between the initially selected category and the AI top category was high. \n\nAdditionally, we demonstrate that aggregated AI-based multiclass probabilities and crowd wisdom significantly increased the number of correct diagnoses in comparison to individual raters or AI in isolation. \n\nNext, we analyzed the impact of AI-based support in clinically relevant scenarios. To examine the potential of AI-based support in telemedicine, we reused prospectively collected images of a randomized controlled trial on self-examinations in high-risk patients. While the CNN was trained only on curated images of pigmented lesions, this sample also included non-pigmented variants of keratinocyte cancer, mucosal lesions and low-quality images. Although the proportion of correct specific diagnoses was significantly lower for these images (53.9% versus 76.2%; P = 8.9 \u00d7 10[\u2212][14], chi-squared test; n = 1,430 images), the CNN was able to recognize 95.2% of patients with skin cancer at a specificity of 59.2%. \n\nSimilarly to recent findings in AI-based breast cancer screening, our results indicate that AI-based skin cancer screening could triage high-risk cases and extend the intervals between face-to-face visits in low-risk cases. The optimal operating points to balance the potential benefits of AI-based triage with the risk of filtering out patients with skin cancer remain to be determined. \n\nIn another scenario, we asked dermatologists to rethink their face-to-face decisions in suspicious cases after providing them with AI-based multiclass probabilities, but without making them aware that they had previously managed the patient. With AI-based support, dermatologists switched from \u2018excision\u2019 to \u2018monitor\u2019 in 15.5% (7 of 45) of decisions for benign lesions, without increasing the number of malignant lesions that switched contrariwise. This result illustrates how human\u2013computer collaboration could decrease the number of unwarranted interventions and costs. AI-based support in this setting increased the frequency of correct specific diagnoses from 55.6% to 75.0%. \n\nFinally, we demonstrate that explanations for AI-based predictions can be translated into a human-understandable visual concept. By analyzing gradient-weighted class activation mapping (Grad-CAM), we show that attention of the CNN outside the object is higher for the prediction of actinic keratoses than for other categories. We hypothesize that, due to visual entrenchment, humans focus on the lesion and not on the background and frequently miss this clue. Here we show that teaching medical students to pay attention to chronic sun damage in the background improved the frequency of correct diagnoses of pigmented actinic keratoses from 32.5% to 47.3%. The overall frequency of correct diagnoses in all categories combined increased from 55.2% to 59.1%. \n\nThis study examines human\u2013computer collaboration from multiple angles and under varying conditions. We used the domain of skin cancer recognition for simplicity, but our study could serve as a framework for similar research in image-based diagnostic medicine. In contrast to the current narrative, our findings suggest that the primary focus should shift from human\u2013computer competition to human\u2013computer collaboration. From a regulatory perspective, the performance of AI-based systems should be tested under real-world conditions in the hands of the intended users and not as stand-alone devices. Only then can we expect to rationally adopt and improve AI-based decision support and to accelerate its evolution.",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Human-computer collaboration for skin cancer recognition.",
        "abstract": "The rapid increase in telemedicine coupled with recent advances in diagnostic artificial intelligence (AI) create the imperative to consider the opportunities and risks of inserting AI-based support into new paradigms of care. Here we build on recent achievements in the accuracy of image-based AI for skin cancer diagnosis to address the effects of varied representations of AI-based support across different levels of clinical expertise and multiple clinical workflows. We find that good quality AI-based support of clinical decision-making improves diagnostic accuracy over that of either AI or physicians alone, and that the least experienced clinicians gain the most from AI-based support. We further find that AI-based multiclass probabilities outperformed content-based image retrieval (CBIR) representations of AI in the mobile technology environment, and AI-based support had utility in simulations of second opinions and of telemedicine triage. In addition to demonstrating the potential benefits associated with good quality AI in the hands of non-expert clinicians, we find that faulty AI can mislead the entire spectrum of clinicians, including experts. Lastly, we show that insights derived from AI class-activation maps can inform improvements in human diagnosis. Together, our approach and findings offer a framework for future studies across the spectrum of image-based diagnostics to improve human-computer collaboration in clinical practice.\n",
        "methods": "**Methods**\n\n**Network training.** We fine-tuned a CNN for classification of seven different categories of the HAM10000 dataset[17]. We performed training on NVIDIA graphics processing units (GPUs) using the Pytorch[34] framework and chose a ResNet34 (ref. [35]) architecture, with weights initiated by pretraining on ImageNet[36] data. Cross-entropy served as the loss function, with weighting dependent on the frequency of classes within the dataset. The learning rate was initialized at 0.0001, with a tenfold reduction in case of no validation loss improvement for more than three epochs, but a minimum of 1 \u00d7 10[\u2212][9]. We used adaptive moment estimation (Adam[37]) as the optimizer and performed a maximum of 100 training epochs with early stopping. Images were presented in batches of 32, randomly cropped and resized to 224 \u00d7 224 pixels without normalization of a mean pixel, randomly rotated by 90 degrees and flipped with minor jitter in color, contrast, saturation and hue. The publicly available HAM10000 dataset, which corresponds to the training set of the ISIC 2018 challenge[18], was the source of images used for training and fivefold cross-validation. We selected the single best performing network on the hold-out validation set for further interaction with raters. For inference, images were cropped to 80% and resized to 224 \u00d7 224 pixels, with minor test-time augmentation consisting of horizontal flipping and rotation by 0 or 90 degrees. For the telemedicine dataset, we also applied color normalization via Shades of Gray[38] with a Minkowski norm of 6. The multiclass probabilities presented to the raters were obtained by applying a softmax function to contain all class probabilities between 0\u2013100%. To find similar images, we used the same CNN to extract the feature vector of the target image and compared it to feature vectors of images in the HAM10000 dataset via cosine similarity[20]. We stored the four closest images of each class and presented them in the AI-based CBIR decision support.\n\n**Interaction platform and raters.** _Online interaction platform._ The web-based platform DermaChallenge, which was developed at the Medical University of Vienna, served as the interface through which the performance of human raters and AI for the diagnostic task was evaluated and quantified. The platform is split into a back end and a front end, and both are deployed on a stack of well-known web technologies (Linux, Apache, MySQL and PHP). Please refer to the Nature Research Reporting Summary for details of the specific software versions used. The back end offers a representational state transfer interface to load and persist data, as well as JavaScript Object Notation web tokens to authenticate participants. The transport layer security and secure sockets layer protocol are used to encrypt all communications. The front end is optimized for mobile devices (mobile phones and tablets) but can also be used on any other platform via a JavaScript-enabled web browser. Before public deployment, five users tested the platform.\n\n_Recruitment and characteristics of raters._ We used mailing lists and social media posts of the International Society of Dermoscopy to recruit online raters. To participate in the study, raters had to register with a username, valid email address and password. In addition, we asked raters for details on their age (age groups spanning 10 years), gender, country, profession and years of experience in dermatoscopy ((1) less than 1 year, (2) opportunistic use for more than 1 year, (3) regular use for 1 to 5 years, (4) regular use for more than 5 years or (5) more than 10 years of experience). Each rater had to perform multiple screening tests to ensure that the self-reported experience matched actual skills. Screening tests consisted of simple domain-specific tasks, for example, to assign one of the seven possible diagnoses to ten cases, to separate melanomas from non-melanomas and to separate seborrheic keratoses from other lesions. We recruited 302 raters for the first interaction study that screened different forms of AI-based support, and 155 raters were recruited for the extended interaction study (inclusion of images with faulty AI-based support) and the telemedicine study (Supplementary Table 4). The distribution of raters according to task is presented in Supplementary Table 4. Second-opinion raters consisted of eight board-certified dermatologists and three dermatology residents, who were recruited because they diagnosed and managed more than two suspicious skin lesions on a face-to-face basis between April and September 2019. For the knowledge transfer study, we invited fourth-year medical students to participate; of the 650 medical students invited, 200 agreed to participate and 189 answered more than 50% of the test questions.\n\n_Characteristics of images and patients._ The benchmark test set of the ISIC 2018 challenge served as the sample for the interaction studies[9]. Of the 1,511 dermoscopic images in this set, 928 images were collected in the Department of Dermatology at the Medical University of Vienna, 267 images were collected in the skin cancer practice of Cliff Rosendahl in Queensland and the remaining 316 images were collected in other centers in Turkey (n = 117), New Zealand (n = 87), Sweden (n = 92) and Argentina (n = 20), to ensure diversity of skin types. The mean age of patients was 50.8 years (s.d. 17.4 years), and 46.2% of patients were female. The Austrian image set consists of lesions from patients referred to a tertiary European center specializing in the early detection of melanoma in high-risk groups. This group of patients is mainly of European ancestry and have a large number of nevi and skin types I\u2013III. The Australian image set includes lesions from patients of a primary-care facility in an area with a high incidence of skin cancer. Patients are typified by Celtic complexion, skin type I or II and chronic sun damage. Routine pathology evaluation (n = 786), biology (that is, >1.5 years of sequential dermoscopic imaging without changes; n = 458), expert consensus in common, straightforward, non-melanocytic cases that were not excised (n = 260) and in vivo confocal images (n = 7) served as the ground truth. Controversial cases with ambiguous histopathologic reports were excluded. Due to random sampling, only 1,412 of 1,511 images were finally used and evaluated by the raters. The 1,412 used cases consisted of 43 AKIECs, 93 BCCs, 217 BKLs, 44 DFs, 171 MELs, 809 NVs and 35 VASCs. For the telemedicine study, we included 93 of 98 participants (mean age 41.1 years (s.d. 12.2 years); 71% female) from the intervention arm of a recently conducted prospective randomized study[27] on mobile teledermoscopy for skin self-examinations. All 93 patients permitted reuse of their images. The participants had at least two skin cancer risk factors (light skin complexion and fair hair; skin that never or rarely tans and always or mostly burns; a family history of melanoma or a personal history of skin cancer, or many nevi; and residing in Queensland) as self-reported in the eligibility survey. A teledermoscopic evaluation was performed for all lesions. Face-to-face examination by an experienced board-certified dermatologist (H.P.S.) or the histopathologic report, in cases where the lesion was removed, served as the ground truth. The set of lesions consisted of 1,521 images of 596 lesions, including 29 AKIECs, 6 BCCs, 102 BKLs, 410 NVs, 2 squamous cell carcinomas (SCCs) and 9 VASCs. For calculation of diagnostic values, ground-truth data were mapped to classes of the HAM10000 dataset, if possible. We excluded nonspecific categories (n = 38 lesions) such as \u2018other\u2019, \u2018no lesion\u2019 or \u2018previously removed\u2019, because they could be mapped to neither the \u2018benign\u2019 nor \u2018malignant\u2019 category. The sample also included images that were not represented in the training data (non-pigmented variants of keratinocyte cancers, mucosal lesions and low-quality images), which were excluded from the telemedicine support study but not from the triage study, to better simulate a realistic scenario. For the second-opinion study, we searched the database of the Department of Dermatology at the Medical University of Vienna for dermoscopy images taken between April and September 2019. We included images if the lesion was excised and had a definite histopathologic diagnosis and if lesions were examined by a physician who was responsible for the face-to-face diagnosis of at least two other cases in this time period. The final sample set (n = 79) included 3 AKIECs, 23 BCCs, 13 BKLs, 2 DFs, 15 MELs, 21 NVs, 1 \u2018other\u2019 (scar) and 1 SCC. The mean age of patients was 64.6 years (s.d. 19.8 years), and 34.5% of patients were female. Patients were mainly of European ancestry and had skin type II (41.7%), III (57.1%) or IV (1.2%). As in the telemedicine scenario, we did not exclude images of categories that were not present in the training data or images of low quality. For the knowledge transfer study, the sample cases (n = 25) were randomly selected from the ISIC 2018 test set and stratified by diagnosis (6 AKIECs, 3 BCCs, 3 BKLs, 3 DFs, 3 MELs, 4 NVs and 3 VASCs).\n\n_Design of diagnostic studies._ To test the interaction of raters with different forms of AI-based decision support, we generated batches of 28 images. Each batch contained four randomly selected examples of every class. The raters\u2019 task was to diagnose the 28 unknown test images, first without and then with one type of decision support. We created a stratified randomization procedure to ensure a balanced distribution of the four types of decision support over all disease categories. The interaction study was online from 29 May 2019 to 15 January 2020. We excluded tests if the number of correct answers was lower than expected by chance to avoid noisy random data. We included only the first five tests for each rater to avoid biasing the results toward raters with high repetitions. The extended interaction study was open for participation between 15 January 2020 and 18 February 2020, presenting only multiclass probabilities as decision support. It included one image for every diagnosis from the ISIC 2018 test set with unaltered AI-based multiclass probabilities, two images with shuffled (resulting as incorrect) AI-based multiclass probabilities and eight images from the telemedicine study (see \u2018Characteristics of images and patients\u2019). The image sources were not disclosed to the raters. The second-opinion study was performed on a local web interface. Physicians who examined the patient face to face in real life were asked to reconsider their diagnosis and decisions with AI-based support. The case presentations included metadata (age, gender and localization), overview and close-up images (if available) and dermoscopic images. Physicians were not made aware that they had treated the patient before or of their previous decision on the case. Physicians were asked to provide their best diagnosis out of the seven predefined disease categories, as well as an extra category termed \u2018other\u2019, followed by their management decision (\u2018no intervention\u2019, \u2018monitor\u2019 or \u2018excise\u2019). No time constraints were set for this task. For the knowledge transfer study, we first examined the gradient-weighted class-activation maps[29], which were created for all images of the training set. We observed that the background attention of the CNN was significantly higher for predictions of the \u2018pigmented actinic keratosis\u2019 class than for other classes (P = 4.6 \u00d7 10[\u2212][12], two-sided unpaired t-test; Extended Data Fig. 2). We interpreted this finding as a diagnostic clue that points to the severely sun-damaged skin in the background of actinic keratoses, which is usually absent or not as severe in other disease categories. To test the hypothesis that teaching this clue to humans will improve their diagnostic skills, fourth-year medical students without previous knowledge of skin cancer detection received a 30-min introductory lecture about dermoscopy, and immediately thereafter students had to diagnose 25 test images (single best diagnosis). Answers were collected with a wireless audience response and voting system. Next, the lecturer presented an additional clue of \u2018sun-damaged skin in the surrounding skin of actinic keratoses\u2019 and the students repeated the test.\n\n**Statistics.** To simulate collective ratings of realistically small human groups (Fig. 3a), we confined the dataset to images with at least three distinct ratings (resulting range of ratings per image: 3\u201369). For each image, we created 30 bootstraps of three to five randomly selected ratings, whichever was the maximum available without replacement, and determined the most common rating as the prediction of the collective (first past the post). Ties were broken randomly. Next, we calculated the proportion of correct bootstrapped predictions to obtain the mean accuracy for each image as published previously[39]. To combine human collectives with CNN-based predictions, we took the arithmetic mean of the sum of the human multiclass probabilities, which were derived from the frequencies of bootstrapped human ratings, and the corresponding CNN-based multiclass probabilities. For analyses of diagnoses, we averaged the results for each image before comparisons; for analyses of raters with and without decision support, we calculated the arithmetic mean for each user before comparisons. The mean answering time for each user in every interaction modality served as a surrogate marker for confidence; answers that were faster or slower than the individual mean were regarded as \u2018confident\u2019 or \u2018non-confident\u2019, respectively. For the filtering procedure in the telemedicine study, we used a predefined cutoff of \u22650.17 to indicate malignancy, because this cutoff was selected by human raters in the interaction study (Extended Data Fig. 2). If patients photographed a lesion more than once, a single image above the cutoff was sufficient to label the lesion as \u2018probably malignant\u2019 and likewise on the patient level. We used a one-sample t-test to distinguish whether continuous data with normal distributions deviated from zero. Comparisons of continuous data between groups were performed with paired or unpaired t-tests or Wilcoxon signed-rank test, as appropriate. A chi-squared test was used to compare proportions. All reported P values were corrected for multiple testing (Holm\u2013Bonferroni[40]), and a two-sided P value < 0.05 was regarded as statistically significant. All analyses were performed using R v3.6.2 (ref. [41]), and plots were created with ggplot v3.2.1 (ref. [42]) and ggalluvial v0.11.1.\n\n**Reporting Summary.** Further information on research design is available in the Nature Research Reporting Summary linked to this article.",
        "results": "**Results**\n\nOur results show that decision support with AI-based multiclass probabilities improves the accuracy of human raters from 63.6% to 77.0% (increase of 13.3%, 95% CI 11.5% to 15.2%; P = 4.9 \u00d7 10[\u2212][35], two-sided paired t-test, t = 14.5, d.f. = 301; n = 302 raters), but no improvement was observed for decision support with AI-based prediction of malignancy or with our representation of AI-based CBIR. This suggests that the form of decision support should be in accordance with the given task. The probability of malignancy may be useful for simple binary management decisions such as whether to perform a biopsy or not, but not for a multiclass diagnostic problem. The studied form of AI-based CBIR is neither simple nor concrete; it needs more extensive cognitive engagement in terms of time and decision-making, because the rater needs to extrapolate the diagnosis from similarities between the test image and images with known diagnoses. The raters needed significantly more time to interact with AI-based CBIR decision support than with other types of support. Over time, human raters also tended to ignore the AI-based CBIR decision support. However, given that a large spectrum of CBIR approaches are described in the literature, another form of CBIR may still provide benefit. \n\nAfter we established that multiclass probabilities were the best form of CNN output for the given task, we focused on this form to explore the impact of AI-based support on human performance in more detail. We show an inverse relationship between the net gain from AI-based support and rater experience (Pearson\u2019s r = \u22120.18, 95% CI \u22120.28 to \u22120.07, P = 1.5 \u00d7 10[\u2212][2]; n = 302 raters). Raters in the least experienced group changed their initial diagnosis more often than experts (mean 26.0%, 95% CI 21.3% to 30.7% versus mean 14.7%, 95% CI 9.9% to 19.6%). Expert raters benefited only marginally (net gain 13.4%, 95% CI 6.3% to 20.6%) and only if they were not confident with their initial diagnosis, but not if they were confident. If experts were confident, they were usually correct and did not need support. \n\nThis finding suggests that, if experts have high confidence in their initial diagnosis, they should ignore AI-based support or not use it at all. The least experienced raters tended to accept AI-based support that contradicted their initial diagnosis even if they were confident. In general, raters changed their initial diagnosis less often if they were confident than if they were not confident in their decision (14.7%, 95% CI 12.6% to 16.8% versus 37.5%, 95% CI 34.0% to 41.0%; P = 1.9 \u00d7 10[\u2212][25], two-sided paired t-test; n = 302 raters). \n\nHaving established a positive impact of good quality AI-based support on diagnostic accuracy, we tested the impact of \u2018faulty\u2019 AI on diagnostic accuracy. Faulty AI could result from the application of AI algorithms on examples beyond the domain of images on which the AI was trained or the more remote possibility of adversarial attacks. To represent faulty AI, we intentionally generated misleading AI-based multiclass probabilities. If the top class probability of the CNN favored the correct diagnosis, we switched the probabilities in such a way that the CNN output favored a random incorrect diagnosis. We demonstrate that any previously observed gains in accuracy with AI-based support turn into a loss when that AI support is faulty. Our results suggest that, if raters build up the trust that is necessary to benefit from AI-based support, they are also vulnerable to perform below their expected ability if there is a fault with the AI. \n\nAnother finding of importance is that the benefit of human\u2013computer collaboration is asymmetrically distributed across disease categories. Our data showed that the net gain was higher for the class of pigmented actinic keratoses and intraepithelial carcinoma (increase of 31.5%, 95% CI 22.9% to 40.1%; n = 43 images) than for other categories. This suggests that the benefit of AI-based support needs to be adapted to the given task and the expected prevalence of target conditions. \n\nWe further demonstrate that AI-based multiclass ranking and probabilities have an impact on the raters\u2019 tendency to change their initial diagnosis. Most changes occurred in favor of the AI-predicted top category. Raters typically maintained their decisions that were in disagreement with the AI prediction only if that decision was ranked by AI prediction as at least the second or third option. Furthermore, raters tended to change their assessments more frequently when the difference in the AI-predicted probability between the initially selected category and the AI top category was high. \n\nAdditionally, we demonstrate that aggregated AI-based multiclass probabilities and crowd wisdom significantly increased the number of correct diagnoses in comparison to individual raters or AI in isolation. \n\nNext, we analyzed the impact of AI-based support in clinically relevant scenarios. To examine the potential of AI-based support in telemedicine, we reused prospectively collected images of a randomized controlled trial on self-examinations in high-risk patients. While the CNN was trained only on curated images of pigmented lesions, this sample also included non-pigmented variants of keratinocyte cancer, mucosal lesions and low-quality images. Although the proportion of correct specific diagnoses was significantly lower for these images (53.9% versus 76.2%; P = 8.9 \u00d7 10[\u2212][14], chi-squared test; n = 1,430 images), the CNN was able to recognize 95.2% of patients with skin cancer at a specificity of 59.2%. \n\nSimilarly to recent findings in AI-based breast cancer screening, our results indicate that AI-based skin cancer screening could triage high-risk cases and extend the intervals between face-to-face visits in low-risk cases. The optimal operating points to balance the potential benefits of AI-based triage with the risk of filtering out patients with skin cancer remain to be determined. \n\nIn another scenario, we asked dermatologists to rethink their face-to-face decisions in suspicious cases after providing them with AI-based multiclass probabilities, but without making them aware that they had previously managed the patient. With AI-based support, dermatologists switched from \u2018excision\u2019 to \u2018monitor\u2019 in 15.5% (7 of 45) of decisions for benign lesions, without increasing the number of malignant lesions that switched contrariwise. This result illustrates how human\u2013computer collaboration could decrease the number of unwarranted interventions and costs. AI-based support in this setting increased the frequency of correct specific diagnoses from 55.6% to 75.0%. \n\nFinally, we demonstrate that explanations for AI-based predictions can be translated into a human-understandable visual concept. By analyzing gradient-weighted class activation mapping (Grad-CAM), we show that attention of the CNN outside the object is higher for the prediction of actinic keratoses than for other categories. We hypothesize that, due to visual entrenchment, humans focus on the lesion and not on the background and frequently miss this clue. Here we show that teaching medical students to pay attention to chronic sun damage in the background improved the frequency of correct diagnoses of pigmented actinic keratoses from 32.5% to 47.3%. The overall frequency of correct diagnoses in all categories combined increased from 55.2% to 59.1%. \n\nThis study examines human\u2013computer collaboration from multiple angles and under varying conditions. We used the domain of skin cancer recognition for simplicity, but our study could serve as a framework for similar research in image-based diagnostic medicine. In contrast to the current narrative, our findings suggest that the primary focus should shift from human\u2013computer competition to human\u2013computer collaboration. From a regulatory perspective, the performance of AI-based systems should be tested under real-world conditions in the hands of the intended users and not as stand-alone devices. Only then can we expect to rationally adopt and improve AI-based decision support and to accelerate its evolution.",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Human-computer collaboration for skin cancer recognition.",
        "abstract": "The rapid increase in telemedicine coupled with recent advances in diagnostic artificial intelligence (AI) create the imperative to consider the opportunities and risks of inserting AI-based support into new paradigms of care. Here we build on recent achievements in the accuracy of image-based AI for skin cancer diagnosis to address the effects of varied representations of AI-based support across different levels of clinical expertise and multiple clinical workflows. We find that good quality AI-based support of clinical decision-making improves diagnostic accuracy over that of either AI or physicians alone, and that the least experienced clinicians gain the most from AI-based support. We further find that AI-based multiclass probabilities outperformed content-based image retrieval (CBIR) representations of AI in the mobile technology environment, and AI-based support had utility in simulations of second opinions and of telemedicine triage. In addition to demonstrating the potential benefits associated with good quality AI in the hands of non-expert clinicians, we find that faulty AI can mislead the entire spectrum of clinicians, including experts. Lastly, we show that insights derived from AI class-activation maps can inform improvements in human diagnosis. Together, our approach and findings offer a framework for future studies across the spectrum of image-based diagnostics to improve human-computer collaboration in clinical practice.\n",
        "methods": "**Methods**\n\n**Network training.** We fine-tuned a CNN for classification of seven different categories of the HAM10000 dataset[17]. We performed training on NVIDIA graphics processing units (GPUs) using the Pytorch[34] framework and chose a ResNet34 (ref. [35]) architecture, with weights initiated by pretraining on ImageNet[36] data. Cross-entropy served as the loss function, with weighting dependent on the frequency of classes within the dataset. The learning rate was initialized at 0.0001, with a tenfold reduction in case of no validation loss improvement for more than three epochs, but a minimum of 1 \u00d7 10[\u2212][9]. We used adaptive moment estimation (Adam[37]) as the optimizer and performed a maximum of 100 training epochs with early stopping. Images were presented in batches of 32, randomly cropped and resized to 224 \u00d7 224 pixels without normalization of a mean pixel, randomly rotated by 90 degrees and flipped with minor jitter in color, contrast, saturation and hue. The publicly available HAM10000 dataset, which corresponds to the training set of the ISIC 2018 challenge[18], was the source of images used for training and fivefold cross-validation. We selected the single best performing network on the hold-out validation set for further interaction with raters. For inference, images were cropped to 80% and resized to 224 \u00d7 224 pixels, with minor test-time augmentation consisting of horizontal flipping and rotation by 0 or 90 degrees. For the telemedicine dataset, we also applied color normalization via Shades of Gray[38] with a Minkowski norm of 6. The multiclass probabilities presented to the raters were obtained by applying a softmax function to contain all class probabilities between 0\u2013100%. To find similar images, we used the same CNN to extract the feature vector of the target image and compared it to feature vectors of images in the HAM10000 dataset via cosine similarity[20]. We stored the four closest images of each class and presented them in the AI-based CBIR decision support.\n\n**Interaction platform and raters.** _Online interaction platform._ The web-based platform DermaChallenge, which was developed at the Medical University of Vienna, served as the interface through which the performance of human raters and AI for the diagnostic task was evaluated and quantified. The platform is split into a back end and a front end, and both are deployed on a stack of well-known web technologies (Linux, Apache, MySQL and PHP). Please refer to the Nature Research Reporting Summary for details of the specific software versions used. The back end offers a representational state transfer interface to load and persist data, as well as JavaScript Object Notation web tokens to authenticate participants. The transport layer security and secure sockets layer protocol are used to encrypt all communications. The front end is optimized for mobile devices (mobile phones and tablets) but can also be used on any other platform via a JavaScript-enabled web browser. Before public deployment, five users tested the platform.\n\n_Recruitment and characteristics of raters._ We used mailing lists and social media posts of the International Society of Dermoscopy to recruit online raters. To participate in the study, raters had to register with a username, valid email address and password. In addition, we asked raters for details on their age (age groups spanning 10 years), gender, country, profession and years of experience in dermatoscopy ((1) less than 1 year, (2) opportunistic use for more than 1 year, (3) regular use for 1 to 5 years, (4) regular use for more than 5 years or (5) more than 10 years of experience). Each rater had to perform multiple screening tests to ensure that the self-reported experience matched actual skills. Screening tests consisted of simple domain-specific tasks, for example, to assign one of the seven possible diagnoses to ten cases, to separate melanomas from non-melanomas and to separate seborrheic keratoses from other lesions. We recruited 302 raters for the first interaction study that screened different forms of AI-based support, and 155 raters were recruited for the extended interaction study (inclusion of images with faulty AI-based support) and the telemedicine study (Supplementary Table 4). The distribution of raters according to task is presented in Supplementary Table 4. Second-opinion raters consisted of eight board-certified dermatologists and three dermatology residents, who were recruited because they diagnosed and managed more than two suspicious skin lesions on a face-to-face basis between April and September 2019. For the knowledge transfer study, we invited fourth-year medical students to participate; of the 650 medical students invited, 200 agreed to participate and 189 answered more than 50% of the test questions.\n\n_Characteristics of images and patients._ The benchmark test set of the ISIC 2018 challenge served as the sample for the interaction studies[9]. Of the 1,511 dermoscopic images in this set, 928 images were collected in the Department of Dermatology at the Medical University of Vienna, 267 images were collected in the skin cancer practice of Cliff Rosendahl in Queensland and the remaining 316 images were collected in other centers in Turkey (n = 117), New Zealand (n = 87), Sweden (n = 92) and Argentina (n = 20), to ensure diversity of skin types. The mean age of patients was 50.8 years (s.d. 17.4 years), and 46.2% of patients were female. The Austrian image set consists of lesions from patients referred to a tertiary European center specializing in the early detection of melanoma in high-risk groups. This group of patients is mainly of European ancestry and have a large number of nevi and skin types I\u2013III. The Australian image set includes lesions from patients of a primary-care facility in an area with a high incidence of skin cancer. Patients are typified by Celtic complexion, skin type I or II and chronic sun damage. Routine pathology evaluation (n = 786), biology (that is, >1.5 years of sequential dermoscopic imaging without changes; n = 458), expert consensus in common, straightforward, non-melanocytic cases that were not excised (n = 260) and in vivo confocal images (n = 7) served as the ground truth. Controversial cases with ambiguous histopathologic reports were excluded. Due to random sampling, only 1,412 of 1,511 images were finally used and evaluated by the raters. The 1,412 used cases consisted of 43 AKIECs, 93 BCCs, 217 BKLs, 44 DFs, 171 MELs, 809 NVs and 35 VASCs. For the telemedicine study, we included 93 of 98 participants (mean age 41.1 years (s.d. 12.2 years); 71% female) from the intervention arm of a recently conducted prospective randomized study[27] on mobile teledermoscopy for skin self-examinations. All 93 patients permitted reuse of their images. The participants had at least two skin cancer risk factors (light skin complexion and fair hair; skin that never or rarely tans and always or mostly burns; a family history of melanoma or a personal history of skin cancer, or many nevi; and residing in Queensland) as self-reported in the eligibility survey. A teledermoscopic evaluation was performed for all lesions. Face-to-face examination by an experienced board-certified dermatologist (H.P.S.) or the histopathologic report, in cases where the lesion was removed, served as the ground truth. The set of lesions consisted of 1,521 images of 596 lesions, including 29 AKIECs, 6 BCCs, 102 BKLs, 410 NVs, 2 squamous cell carcinomas (SCCs) and 9 VASCs. For calculation of diagnostic values, ground-truth data were mapped to classes of the HAM10000 dataset, if possible. We excluded nonspecific categories (n = 38 lesions) such as \u2018other\u2019, \u2018no lesion\u2019 or \u2018previously removed\u2019, because they could be mapped to neither the \u2018benign\u2019 nor \u2018malignant\u2019 category. The sample also included images that were not represented in the training data (non-pigmented variants of keratinocyte cancers, mucosal lesions and low-quality images), which were excluded from the telemedicine support study but not from the triage study, to better simulate a realistic scenario. For the second-opinion study, we searched the database of the Department of Dermatology at the Medical University of Vienna for dermoscopy images taken between April and September 2019. We included images if the lesion was excised and had a definite histopathologic diagnosis and if lesions were examined by a physician who was responsible for the face-to-face diagnosis of at least two other cases in this time period. The final sample set (n = 79) included 3 AKIECs, 23 BCCs, 13 BKLs, 2 DFs, 15 MELs, 21 NVs, 1 \u2018other\u2019 (scar) and 1 SCC. The mean age of patients was 64.6 years (s.d. 19.8 years), and 34.5% of patients were female. Patients were mainly of European ancestry and had skin type II (41.7%), III (57.1%) or IV (1.2%). As in the telemedicine scenario, we did not exclude images of categories that were not present in the training data or images of low quality. For the knowledge transfer study, the sample cases (n = 25) were randomly selected from the ISIC 2018 test set and stratified by diagnosis (6 AKIECs, 3 BCCs, 3 BKLs, 3 DFs, 3 MELs, 4 NVs and 3 VASCs).\n\n_Design of diagnostic studies._ To test the interaction of raters with different forms of AI-based decision support, we generated batches of 28 images. Each batch contained four randomly selected examples of every class. The raters\u2019 task was to diagnose the 28 unknown test images, first without and then with one type of decision support. We created a stratified randomization procedure to ensure a balanced distribution of the four types of decision support over all disease categories. The interaction study was online from 29 May 2019 to 15 January 2020. We excluded tests if the number of correct answers was lower than expected by chance to avoid noisy random data. We included only the first five tests for each rater to avoid biasing the results toward raters with high repetitions. The extended interaction study was open for participation between 15 January 2020 and 18 February 2020, presenting only multiclass probabilities as decision support. It included one image for every diagnosis from the ISIC 2018 test set with unaltered AI-based multiclass probabilities, two images with shuffled (resulting as incorrect) AI-based multiclass probabilities and eight images from the telemedicine study (see \u2018Characteristics of images and patients\u2019). The image sources were not disclosed to the raters. The second-opinion study was performed on a local web interface. Physicians who examined the patient face to face in real life were asked to reconsider their diagnosis and decisions with AI-based support. The case presentations included metadata (age, gender and localization), overview and close-up images (if available) and dermoscopic images. Physicians were not made aware that they had treated the patient before or of their previous decision on the case. Physicians were asked to provide their best diagnosis out of the seven predefined disease categories, as well as an extra category termed \u2018other\u2019, followed by their management decision (\u2018no intervention\u2019, \u2018monitor\u2019 or \u2018excise\u2019). No time constraints were set for this task. For the knowledge transfer study, we first examined the gradient-weighted class-activation maps[29], which were created for all images of the training set. We observed that the background attention of the CNN was significantly higher for predictions of the \u2018pigmented actinic keratosis\u2019 class than for other classes (P = 4.6 \u00d7 10[\u2212][12], two-sided unpaired t-test; Extended Data Fig. 2). We interpreted this finding as a diagnostic clue that points to the severely sun-damaged skin in the background of actinic keratoses, which is usually absent or not as severe in other disease categories. To test the hypothesis that teaching this clue to humans will improve their diagnostic skills, fourth-year medical students without previous knowledge of skin cancer detection received a 30-min introductory lecture about dermoscopy, and immediately thereafter students had to diagnose 25 test images (single best diagnosis). Answers were collected with a wireless audience response and voting system. Next, the lecturer presented an additional clue of \u2018sun-damaged skin in the surrounding skin of actinic keratoses\u2019 and the students repeated the test.\n\n**Statistics.** To simulate collective ratings of realistically small human groups (Fig. 3a), we confined the dataset to images with at least three distinct ratings (resulting range of ratings per image: 3\u201369). For each image, we created 30 bootstraps of three to five randomly selected ratings, whichever was the maximum available without replacement, and determined the most common rating as the prediction of the collective (first past the post). Ties were broken randomly. Next, we calculated the proportion of correct bootstrapped predictions to obtain the mean accuracy for each image as published previously[39]. To combine human collectives with CNN-based predictions, we took the arithmetic mean of the sum of the human multiclass probabilities, which were derived from the frequencies of bootstrapped human ratings, and the corresponding CNN-based multiclass probabilities. For analyses of diagnoses, we averaged the results for each image before comparisons; for analyses of raters with and without decision support, we calculated the arithmetic mean for each user before comparisons. The mean answering time for each user in every interaction modality served as a surrogate marker for confidence; answers that were faster or slower than the individual mean were regarded as \u2018confident\u2019 or \u2018non-confident\u2019, respectively. For the filtering procedure in the telemedicine study, we used a predefined cutoff of \u22650.17 to indicate malignancy, because this cutoff was selected by human raters in the interaction study (Extended Data Fig. 2). If patients photographed a lesion more than once, a single image above the cutoff was sufficient to label the lesion as \u2018probably malignant\u2019 and likewise on the patient level. We used a one-sample t-test to distinguish whether continuous data with normal distributions deviated from zero. Comparisons of continuous data between groups were performed with paired or unpaired t-tests or Wilcoxon signed-rank test, as appropriate. A chi-squared test was used to compare proportions. All reported P values were corrected for multiple testing (Holm\u2013Bonferroni[40]), and a two-sided P value < 0.05 was regarded as statistically significant. All analyses were performed using R v3.6.2 (ref. [41]), and plots were created with ggplot v3.2.1 (ref. [42]) and ggalluvial v0.11.1.\n\n**Reporting Summary.** Further information on research design is available in the Nature Research Reporting Summary linked to this article.",
        "results": "**Results**\n\nOur results show that decision support with AI-based multiclass probabilities improves the accuracy of human raters from 63.6% to 77.0% (increase of 13.3%, 95% CI 11.5% to 15.2%; P = 4.9 \u00d7 10[\u2212][35], two-sided paired t-test, t = 14.5, d.f. = 301; n = 302 raters), but no improvement was observed for decision support with AI-based prediction of malignancy or with our representation of AI-based CBIR. This suggests that the form of decision support should be in accordance with the given task. The probability of malignancy may be useful for simple binary management decisions such as whether to perform a biopsy or not, but not for a multiclass diagnostic problem. The studied form of AI-based CBIR is neither simple nor concrete; it needs more extensive cognitive engagement in terms of time and decision-making, because the rater needs to extrapolate the diagnosis from similarities between the test image and images with known diagnoses. The raters needed significantly more time to interact with AI-based CBIR decision support than with other types of support. Over time, human raters also tended to ignore the AI-based CBIR decision support. However, given that a large spectrum of CBIR approaches are described in the literature, another form of CBIR may still provide benefit. \n\nAfter we established that multiclass probabilities were the best form of CNN output for the given task, we focused on this form to explore the impact of AI-based support on human performance in more detail. We show an inverse relationship between the net gain from AI-based support and rater experience (Pearson\u2019s r = \u22120.18, 95% CI \u22120.28 to \u22120.07, P = 1.5 \u00d7 10[\u2212][2]; n = 302 raters). Raters in the least experienced group changed their initial diagnosis more often than experts (mean 26.0%, 95% CI 21.3% to 30.7% versus mean 14.7%, 95% CI 9.9% to 19.6%). Expert raters benefited only marginally (net gain 13.4%, 95% CI 6.3% to 20.6%) and only if they were not confident with their initial diagnosis, but not if they were confident. If experts were confident, they were usually correct and did not need support. \n\nThis finding suggests that, if experts have high confidence in their initial diagnosis, they should ignore AI-based support or not use it at all. The least experienced raters tended to accept AI-based support that contradicted their initial diagnosis even if they were confident. In general, raters changed their initial diagnosis less often if they were confident than if they were not confident in their decision (14.7%, 95% CI 12.6% to 16.8% versus 37.5%, 95% CI 34.0% to 41.0%; P = 1.9 \u00d7 10[\u2212][25], two-sided paired t-test; n = 302 raters). \n\nHaving established a positive impact of good quality AI-based support on diagnostic accuracy, we tested the impact of \u2018faulty\u2019 AI on diagnostic accuracy. Faulty AI could result from the application of AI algorithms on examples beyond the domain of images on which the AI was trained or the more remote possibility of adversarial attacks. To represent faulty AI, we intentionally generated misleading AI-based multiclass probabilities. If the top class probability of the CNN favored the correct diagnosis, we switched the probabilities in such a way that the CNN output favored a random incorrect diagnosis. We demonstrate that any previously observed gains in accuracy with AI-based support turn into a loss when that AI support is faulty. Our results suggest that, if raters build up the trust that is necessary to benefit from AI-based support, they are also vulnerable to perform below their expected ability if there is a fault with the AI. \n\nAnother finding of importance is that the benefit of human\u2013computer collaboration is asymmetrically distributed across disease categories. Our data showed that the net gain was higher for the class of pigmented actinic keratoses and intraepithelial carcinoma (increase of 31.5%, 95% CI 22.9% to 40.1%; n = 43 images) than for other categories. This suggests that the benefit of AI-based support needs to be adapted to the given task and the expected prevalence of target conditions. \n\nWe further demonstrate that AI-based multiclass ranking and probabilities have an impact on the raters\u2019 tendency to change their initial diagnosis. Most changes occurred in favor of the AI-predicted top category. Raters typically maintained their decisions that were in disagreement with the AI prediction only if that decision was ranked by AI prediction as at least the second or third option. Furthermore, raters tended to change their assessments more frequently when the difference in the AI-predicted probability between the initially selected category and the AI top category was high. \n\nAdditionally, we demonstrate that aggregated AI-based multiclass probabilities and crowd wisdom significantly increased the number of correct diagnoses in comparison to individual raters or AI in isolation. \n\nNext, we analyzed the impact of AI-based support in clinically relevant scenarios. To examine the potential of AI-based support in telemedicine, we reused prospectively collected images of a randomized controlled trial on self-examinations in high-risk patients. While the CNN was trained only on curated images of pigmented lesions, this sample also included non-pigmented variants of keratinocyte cancer, mucosal lesions and low-quality images. Although the proportion of correct specific diagnoses was significantly lower for these images (53.9% versus 76.2%; P = 8.9 \u00d7 10[\u2212][14], chi-squared test; n = 1,430 images), the CNN was able to recognize 95.2% of patients with skin cancer at a specificity of 59.2%. \n\nSimilarly to recent findings in AI-based breast cancer screening, our results indicate that AI-based skin cancer screening could triage high-risk cases and extend the intervals between face-to-face visits in low-risk cases. The optimal operating points to balance the potential benefits of AI-based triage with the risk of filtering out patients with skin cancer remain to be determined. \n\nIn another scenario, we asked dermatologists to rethink their face-to-face decisions in suspicious cases after providing them with AI-based multiclass probabilities, but without making them aware that they had previously managed the patient. With AI-based support, dermatologists switched from \u2018excision\u2019 to \u2018monitor\u2019 in 15.5% (7 of 45) of decisions for benign lesions, without increasing the number of malignant lesions that switched contrariwise. This result illustrates how human\u2013computer collaboration could decrease the number of unwarranted interventions and costs. AI-based support in this setting increased the frequency of correct specific diagnoses from 55.6% to 75.0%. \n\nFinally, we demonstrate that explanations for AI-based predictions can be translated into a human-understandable visual concept. By analyzing gradient-weighted class activation mapping (Grad-CAM), we show that attention of the CNN outside the object is higher for the prediction of actinic keratoses than for other categories. We hypothesize that, due to visual entrenchment, humans focus on the lesion and not on the background and frequently miss this clue. Here we show that teaching medical students to pay attention to chronic sun damage in the background improved the frequency of correct diagnoses of pigmented actinic keratoses from 32.5% to 47.3%. The overall frequency of correct diagnoses in all categories combined increased from 55.2% to 59.1%. \n\nThis study examines human\u2013computer collaboration from multiple angles and under varying conditions. We used the domain of skin cancer recognition for simplicity, but our study could serve as a framework for similar research in image-based diagnostic medicine. In contrast to the current narrative, our findings suggest that the primary focus should shift from human\u2013computer competition to human\u2013computer collaboration. From a regulatory perspective, the performance of AI-based systems should be tested under real-world conditions in the hands of the intended users and not as stand-alone devices. Only then can we expect to rationally adopt and improve AI-based decision support and to accelerate its evolution.",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Artificial Intelligence and Its Effect on Dermatologists' Accuracy in Dermoscopic Melanoma Image Classification: Web-Based Survey Study.",
        "abstract": "## BACKGROUND\nEarly detection of melanoma can be lifesaving but this remains a challenge. Recent diagnostic studies have revealed the superiority of artificial intelligence (AI) in classifying dermoscopic images of melanoma and nevi, concluding that these algorithms should assist a dermatologist's diagnoses.\n## OBJECTIVE\nThe aim of this study was to investigate whether AI support improves the accuracy and overall diagnostic performance of dermatologists in the dichotomous image-based discrimination between melanoma and nevus.\n## METHODS\nTwelve board-certified dermatologists were presented disjoint sets of 100 unique dermoscopic images of melanomas and nevi (total of 1200 unique images), and they had to classify the images based on personal experience alone (part I) and with the support of a trained convolutional neural network (CNN, part II). Additionally, dermatologists were asked to rate their confidence in their final decision for each image.\n## RESULTS\nWhile the mean specificity of the dermatologists based on personal experience alone remained almost unchanged (70.6% vs 72.4%; P=.54) with AI support, the mean sensitivity and mean accuracy increased significantly (59.4% vs 74.6%; P=.003 and 65.0% vs 73.6%; P=.002, respectively) with AI support. Out of the 10% (10/94; 95% CI 8.4%-11.8%) of cases where dermatologists were correct and AI was incorrect, dermatologists on average changed to the incorrect answer for 39% (4/10; 95% CI 23.2%-55.6%) of cases. When dermatologists were incorrect and AI was correct (25/94, 27%; 95% CI 24.0%-30.1%), dermatologists changed their answers to the correct answer for 46% (11/25; 95% CI 33.1%-58.4%) of cases. Additionally, the dermatologists' average confidence in their decisions increased when the CNN confirmed their decision and decreased when the CNN disagreed, even when the dermatologists were correct. Reported values are based on the mean of all participants. Whenever absolute values are shown, the denominator and numerator are approximations as every dermatologist ended up rating a varying number of images due to a quality control step.\n## CONCLUSIONS\nThe findings of our study show that AI support can improve the overall accuracy of the dermatologists in the dichotomous image-based discrimination between melanoma and nevus. This supports the argument for AI-based tools to aid clinicians in skin lesion classification and provides a rationale for studies of such classifiers in real-life settings, wherein clinicians can integrate additional information such as patient age and medical history into their decisions.\n",
        "methods": "### Methods\n\n##### Study Design\n\nThis study was conducted from January 10, 2019 (design of study) to September 27, 2019 and was inspired by the study design of Sinz et al [2]. Questionnaires were sent out on June 12, 2019 and were completed by August 12, 2019. Ethical approval was waived by the ethics committee of the University of Heidelberg, Mannheim Faculty of Medicine as all the images were open source and anonymous and all participating dermatologists automatically became part of the study group.\n\nThe set of images used for the evaluation consisted of 1200 unique dermoscopic images split into 12 nonoverlapping individual test sets containing 100 images each (50 melanomas and 50 nevi). Each individual set was randomly assigned to exactly one dermatologist who diagnosed his/her set twice\u2014first without AI support (part I) and afterwards with AI support (part II). Images for each part were presented to the participants in 2 separate surveys containing 50 images each (again 50:50 split) to investigate a possible learning effect. Thus, every participant completed 4 surveys (2 surveys for part I and 2 surveys for part II) where images seen in part I were identical to those seen in part II. Both surveys from part I were carried out before any survey from part II. The participants received detailed instructions by email in which the survey structure and the procedure were discussed and the lesion distribution was disclosed. For part II, participants were made aware of the classifier\u2019s performance, which was established beforehand on a separate validation set.\n\n##### Classifier Training\n\nImages were obtained from the ISIC archive [19], with a large fraction of images coming from the HAM10000 dataset [20]. The archive is publicly accessible and contains anonymous dermoscopic images from multiple sources and various camera systems. From the available pigmented lesions, only images showing a biopsy-verified melanoma (n=1633) or nevus (n=3311) were selected (total n=4944).\n\nAs each of the 12 individual test sets consisted of 2 subsets (with 50 images each), there were 24 test subsets. A designated training set was constructed for each test subset by removing its images from the total image data, resulting in 24 training sets, each containing 4894 images (ie, 4944 \u2013 50). The class distribution across each training set was unevenly distributed with 1608 melanoma images in contrast to 3286 nevus images. To counter the class imbalance, the set of melanoma images was duplicated. Online data augmentation was used to increase the diversity of the training data and to modify the duplicated images. Before balancing, an evenly balanced validation set was removed from each training set for calibration and validation purposes later. A detailed description of the classifier training is given in Multimedia Appendix 1.\n\n##### Classifier Integration\n\nFor AI support to be used in an effective manner, a confidence measure was displayed in addition to the binary decision (melanoma or nevus), which indicated the classifier\u2019s confidence in its decision. The measure was obtained by mapping the classifier\u2019s raw output probability for the predicted class into a percentage range from 0% to 100%.\n\nParticipants were made aware of the classifier\u2019s overall performance by establishing the overall performance of all 24 classifiers on their validation set and by taking the average, which resulted in mean sensitivity, mean specificity, and mean accuracy of 78.0%, 81.0%, and 80.0% respectively.\n\n##### Electronic Survey and Participants\n\nFifteen participants were invited by personal invitation via their university email accounts by the principal investigator (TJB) based on previous cooperation. Three had to be excluded because one reported to not have time to do the second survey properly, and 2 did not attempt or finish the survey in the proposed timeframe, leaving 12 dermatologists (all board-certified and coauthors of this publication) from 9 German university hospitals for analysis. The Google Forms closed web-based survey was not anonymous and each participant had to disclose his or her full name and additional metadata. For part I, the dermatologists were shown images of biopsy-verified skin lesions and asked 3 mandatory questions. First, a personal rating of the image quality had to be given (excellent, good, sufficient, poor, other image problems, no image visible). Second, the participant was asked for a diagnosis of the displayed lesion with the options being benign or malignant. Third, the participant was requested to quantify his or her confidence in the decision on a scale of 0 (=very uncertain) to 10 (=very certain). Part II showed identical images and questions and was additionally labeled with the dermatologist\u2019s previous response and the diagnosis of the CNN together with its confidence level. For details on the participants, see Multimedia Appendix 1.\n\n##### Performance and Statistical Analysis\n\nThe primary endpoint of this study was whether dermatologists\u2019 accuracy (overall proportion of the correct predictions among the total number of examined cases) would increase with AI support with the secondary endpoint investigating how sensitivity and specificity would change (tested for significance using Wilcoxon test). Additional analyses investigated whether dermatologists experienced a learning effect when working with AI support (tested for significance using Pearson\u2019s chi-squared test) and how they would respond to this form of computer aid. Statistical significance was considered at P<.05, corrected to 0.016 (Bonferroni correction for the primary and secondary endpoint) to account for multiple testing. The details on how survey answers were evaluated are listed in Multimedia Appendix 1. As a pair of classifiers was responsible for scoring the test set of one participant (one classifier for survey 1, the other for survey 2), the results of both classifiers were combined for analysis reasons so that every dermatologist had a respective counterpart. Values shown in the result sections are based on the mean of all the participants. If absolute values are shown, the denominator and numerator are approximated. In theory, every dermatologist rated the exact same number of images, but due to the quality control step, not every rating was counted. Therefore, each dermatologist ended up rating a varying number of images.",
        "results": "### Results\n\n##### Classification Results\n\nIn the first part of the study, the dermatologists had an overall mean sensitivity of 59.4% (95% CI 53.3%-65.5%), specificity of 70.6% (95% CI 62.3%-78.9%), and accuracy of 65.0% (95% CI 62.3%-67.6%, see Table 1).\n\nIn the second part of the study, where participants could integrate the results of the CNN-based classifier in their decision-making, their overall sensitivity increased significantly to 74.6% (95% CI 69.9%-79.3%; _P=.003). Their mean specificity also showed a positive trend (72.4%; 95% CI 66.2%-78.6%; P=.54), so that the overall accuracy also increased significantly to 73.6% (95% CI 70.9%-76.3%; P=.002).\n\nThe CNN on its own had an even higher sensitivity (84.7%; 95% CI 81.9%-87.6%), specificity (79.1%; 95% CI 74.8%-83.4%), and accuracy (81.9%; 95% CI 79.7%-84.2%) on the test set, which was comparable to its performance on the validation set. The overall performance of the dermatologists and the CNN are summarized in Table 1 while individual performances are captured in Figure 1 and Figure 2, which show an overview of the various metrics for every participant with and without AI support.\n\n**Table 1.** Overall mean performance of dermatologists without artificial intelligence (AI) support (-AI) compared to that of dermatologists with AI support (+AI) and AI on its own. Performance is split into 3 categories, that is, sensitivity, specificity, and accuracy.\n\nPerformance AI Dermatologist -AI Dermatologist +AI\n\nSensitivity (95% CI) 84.7% (81.9%-87.6%) 59.4% (53.3%-65.5%) 74.6% (69.9%-79.3%)\n\nSpecificity (95% CI) 79.1% (74.8%-83.4%) 70.6% (62.3%-78.9%) 72.4% (66.2%-78.6%)\n\nAccuracy (95% CI) 81.9% (79.7%-84.2%) 65.0% (62.3%-67.6%) 73.6% (70.9%-76.3%)\n\n**Figure 1.** Combined and individual dermatologists\u2019 performance without and with artificial intelligence (AI) support. Every dot represents a single participant. A line between 2 dots connects the participants' metric without AI support to the corresponding metric with AI support. Highlighted dots represent the participants combined. Boxes indicate 25th and 75th percentile while the horizontal line within shows the median (50th percentile). Whiskers indicate the data range (1.5*IQR) where points beyond are considered as outliers.\n\n**Figure 2.** Combined and individual dermatologists\u2019 diagnostic accuracy without and with artificial intelligence (AI) support. Diagnostic accuracy is measured using sensitivity and specificity. Arrows represent the change in the diagnostic accuracy from without AI support to with AI support. Highlighted arrows represent the participants combined. In addition, the black curve denotes the mean receiver operating characteristic curve of the classifier. ROC: receiver operating characteristic; CNN: convolutional neural network; AUC: area under the curve.\n\n##### Effects of the Use of the Classifier on the Decisions of the Dermatologists\n\nThe mean overall confidence of the dermatologists increased only marginally with the use of the classifier (65.0%; 95% CI 58.5%-71.6% without AI to 66.8%; 95% CI 60.2%-73.4% with AI). The confidence of the classifier in its diagnostic decisions was at 65.0% (95% CI 62.8%-67.4%). Upon detailed analysis, one can detect opposite effects on the dermatologists\u2019 confidence and switching behavior depending on agreement or disagreement with the classifier. On average, dermatologists and classifiers agreed in 63% (59/94; 95% CI 59.4%-66.1%) of cases. Among these, 8% (7/94) of cases were diagnosed incorrectly and 55% (52/94) of cases were classified correctly (see Table 2).\n\n**Table 2.** Distribution of the correct and incorrect predictions by classifier and dermatologists without artificial intelligence (AI) support and switching in response to AI support. Percentages displayed below show the amount of times a switch did or did not occur for dermatologists when answering part II of the survey.\n\nGroupings Proportion, n (%)[a] 95% CI\n\n**Both incorrect, n=94** 7 (8) 5.5%-10.4%\n\nDermatologist switched, n=7 0 (1) 0%-2.2%\n\nDermatologist stayed, n=7 7 (99) 97.8%-100%\n\n**AI correct, Dermatologist incorrect, n=94** 25 (27) 24.0%-30.1%\n\nDermatologist switched, n=25 11 (46) 33.1%-58.4%\n\nDermatologist stayed, n=25 14 (54) 41.6%-66.9%\n\n**AI incorrect, Dermatologist correct, n=94** 10 (10) 8.4%-11.8%\n\nDermatologist switched, n=10 4 (39) 23.2%-55.6%\n\nDermatologist stayed, n=10 6 (61) 44.4%-76.9%\n\n**Both correct, n=94** 52 (55) 52.4%-57.1%\n\nDermatologist switched, n=52 0 (0) 0%-0.5%\n\nDermatologist stayed, n=52 52 (100) 99.5%-100%\n\naAs the mean of all the participants was taken and every participant ended up rating a varying amount of images due to the quality control step, the reported absolute values are approximations.\n\nOf the 37% (35/94; 95% CI 33.9%-40.6%) of cases in which dermatologists and classifier came to different conclusions, 10% (10/94) of cases were diagnosed correctly by the dermatologists and 27% (25/94) of cases were diagnosed correctly by the classifier.\n\nIn cases of agreement, the mean confidence of the dermatologists increased substantially from the first part of the study to the second part of the study (from 67.0%; 95% CI 60.6%-73.5% to 79.1%; 95% CI 75.0%-83.0%, respectively). In cases of disagreement, it decreased from 61.7% (95% CI 54.6%-68.9%) in the first to 44.3% (95% CI 31.8%-56.7%) in the second part of the study. This was also reflected in the dermatologists\u2019 switching behavior. In cases of agreement, the dermatologists basically never altered their classifications (0/59, 0%; 95% CI 0%-0.8%). In contrast, in cases of disagreement, the dermatologists altered their decision in 43% (15/35; 95% CI 30.8%-56.3%) of those cases. The dermatologists altered their diagnosis less often when they had initially diagnosed the lesion correctly (subsequently changed answer in 4/10, 39% of those cases) than when they had initially diagnosed it incorrectly (subsequently changed answer in 11/25, 46% of those cases, see Table 2). Altogether, this resulted in the observed overall increase in the correct diagnoses by the dermatologists. Out of all the occurring switches, dermatologists showed a higher willingness to switch from benign to malignant (11/15, 72%; 95% CI 53.4%-90.9%) than from malignant to benign (4/15, 28%; 95% CI 9.1%-46.6%).\n\nThe mean confidence levels of the classifier were much higher when the correct conclusion was reached (71.7%; 95% CI 69.3%-74.1%) than when the diagnosis was incorrect (34.8%; 95% CI 30.9%-38.7%). Regarding the dermatologists, a similar trend was less striking for part I (67.3%; 95% CI 61.1%-73.5% correct vs 60.8%; 95% CI 53.3%-68.2% incorrect) but became more pronounced for part II (71.8%; 95% CI 66.4%-77.1% correct vs 52.3%; 95% CI 41.0%-63.5% incorrect). A more detailed breakdown of the dermatologists\u2019 and classifier\u2019s confidence levels is shown in Table 3. Finally, the dermatologists altered their diagnoses in divergent cases more often when the CNN\u2019s confidence levels were high than when they were low (mean CNN confidence levels 63.3%; 95% CI 56.4%-70.1% with subsequent switch vs 53.7%; 95% CI 46.7%-60.7% with no subsequent switch).\n\n**Table 3.** Confidence distribution of the classifier and dermatologists for part I and part II.\n\nGroupings Confidence (95% CI)\n\n**Both incorrect**\n\nAI[a] 35.8% (27.6%-44.1%)\n\nDermatologist part I 62.3% (54.6%-69.9%)\n\nDermatologist part II 69.8% (62.4%-77.3%)\n\n**AI correct, Dermatologist incorrect**\n\nAI 65.4% (60.7%-70.2%)\n\nDermatologist part I 60.8% (53.3%-68.3%)\n\nDermatologist part II 46.0% (33.6%-58.5%)\n\n**AI incorrect, Dermatologist correct**\n\nAI 32.9% (28.1%-37.6%)\n\nDermatologist part I 63.2% (56.3%-70.2%)\n\nDermatologist part II 47.9% (36.4%-59.4%)\n\n**Both correct**\n\nAI 74.5% (72.0%-76.9%)\n\nDermatologist part I 67.9% (61.5%-74.3%)\n\nDermatologist part II 80.3% (76.2%-84.3%)\n\naAI: artificial intelligence.\n\n##### Learning Effect\n\nThe dermatologists received their results immediately after completing the first survey of the second part of the study. Thus, they had the opportunity to implement the lessons learned in the first AI-supported survey during the completion of the second survey in part II. A subanalysis showed that there was no detectable change in the dermatologists\u2019 performance from the first to the second survey in part I of the study (see figure in Multimedia Appendix 2) with no significant difference between sensitivity (P=.50) and specificity (P=.76). In contrast, the dermatologists tended to perform better in the second survey of part II compared to the first survey, suggesting that they had learned how to better incorporate the CNN results into their diagnostic procedures, albeit with differences in sensitivity (P=.21) and specificity (P=.43) being still insignificant. In divergent cases, participants switched their diagnosis in 49% (8/17; 95% CI 36.2%-62.3%) of cases in survey 2, whereas they had only done so in 38% (7/18; 95% CI 21.1%-54.5%) of these cases for survey 1. The tendency to switch at the correct moments in cases of disagreement was reinforced in the second survey wherein dermatologists switched far more often when they were incorrect (5/13, 39%; 95% CI 23.0%-55.7% in the first survey vs 6/12, 52%; 95% CI 39.8%-64.3% in the second survey).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Artificial Intelligence and Its Effect on Dermatologists' Accuracy in Dermoscopic Melanoma Image Classification: Web-Based Survey Study.",
        "abstract": "## BACKGROUND\nEarly detection of melanoma can be lifesaving but this remains a challenge. Recent diagnostic studies have revealed the superiority of artificial intelligence (AI) in classifying dermoscopic images of melanoma and nevi, concluding that these algorithms should assist a dermatologist's diagnoses.\n## OBJECTIVE\nThe aim of this study was to investigate whether AI support improves the accuracy and overall diagnostic performance of dermatologists in the dichotomous image-based discrimination between melanoma and nevus.\n## METHODS\nTwelve board-certified dermatologists were presented disjoint sets of 100 unique dermoscopic images of melanomas and nevi (total of 1200 unique images), and they had to classify the images based on personal experience alone (part I) and with the support of a trained convolutional neural network (CNN, part II). Additionally, dermatologists were asked to rate their confidence in their final decision for each image.\n## RESULTS\nWhile the mean specificity of the dermatologists based on personal experience alone remained almost unchanged (70.6% vs 72.4%; P=.54) with AI support, the mean sensitivity and mean accuracy increased significantly (59.4% vs 74.6%; P=.003 and 65.0% vs 73.6%; P=.002, respectively) with AI support. Out of the 10% (10/94; 95% CI 8.4%-11.8%) of cases where dermatologists were correct and AI was incorrect, dermatologists on average changed to the incorrect answer for 39% (4/10; 95% CI 23.2%-55.6%) of cases. When dermatologists were incorrect and AI was correct (25/94, 27%; 95% CI 24.0%-30.1%), dermatologists changed their answers to the correct answer for 46% (11/25; 95% CI 33.1%-58.4%) of cases. Additionally, the dermatologists' average confidence in their decisions increased when the CNN confirmed their decision and decreased when the CNN disagreed, even when the dermatologists were correct. Reported values are based on the mean of all participants. Whenever absolute values are shown, the denominator and numerator are approximations as every dermatologist ended up rating a varying number of images due to a quality control step.\n## CONCLUSIONS\nThe findings of our study show that AI support can improve the overall accuracy of the dermatologists in the dichotomous image-based discrimination between melanoma and nevus. This supports the argument for AI-based tools to aid clinicians in skin lesion classification and provides a rationale for studies of such classifiers in real-life settings, wherein clinicians can integrate additional information such as patient age and medical history into their decisions.\n",
        "methods": "### Methods\n\n##### Study Design\n\nThis study was conducted from January 10, 2019 (design of study) to September 27, 2019 and was inspired by the study design of Sinz et al [2]. Questionnaires were sent out on June 12, 2019 and were completed by August 12, 2019. Ethical approval was waived by the ethics committee of the University of Heidelberg, Mannheim Faculty of Medicine as all the images were open source and anonymous and all participating dermatologists automatically became part of the study group.\n\nThe set of images used for the evaluation consisted of 1200 unique dermoscopic images split into 12 nonoverlapping individual test sets containing 100 images each (50 melanomas and 50 nevi). Each individual set was randomly assigned to exactly one dermatologist who diagnosed his/her set twice\u2014first without AI support (part I) and afterwards with AI support (part II). Images for each part were presented to the participants in 2 separate surveys containing 50 images each (again 50:50 split) to investigate a possible learning effect. Thus, every participant completed 4 surveys (2 surveys for part I and 2 surveys for part II) where images seen in part I were identical to those seen in part II. Both surveys from part I were carried out before any survey from part II. The participants received detailed instructions by email in which the survey structure and the procedure were discussed and the lesion distribution was disclosed. For part II, participants were made aware of the classifier\u2019s performance, which was established beforehand on a separate validation set.\n\n##### Classifier Training\n\nImages were obtained from the ISIC archive [19], with a large fraction of images coming from the HAM10000 dataset [20]. The archive is publicly accessible and contains anonymous dermoscopic images from multiple sources and various camera systems. From the available pigmented lesions, only images showing a biopsy-verified melanoma (n=1633) or nevus (n=3311) were selected (total n=4944).\n\nAs each of the 12 individual test sets consisted of 2 subsets (with 50 images each), there were 24 test subsets. A designated training set was constructed for each test subset by removing its images from the total image data, resulting in 24 training sets, each containing 4894 images (ie, 4944 \u2013 50). The class distribution across each training set was unevenly distributed with 1608 melanoma images in contrast to 3286 nevus images. To counter the class imbalance, the set of melanoma images was duplicated. Online data augmentation was used to increase the diversity of the training data and to modify the duplicated images. Before balancing, an evenly balanced validation set was removed from each training set for calibration and validation purposes later. A detailed description of the classifier training is given in Multimedia Appendix 1.\n\n##### Classifier Integration\n\nFor AI support to be used in an effective manner, a confidence measure was displayed in addition to the binary decision (melanoma or nevus), which indicated the classifier\u2019s confidence in its decision. The measure was obtained by mapping the classifier\u2019s raw output probability for the predicted class into a percentage range from 0% to 100%.\n\nParticipants were made aware of the classifier\u2019s overall performance by establishing the overall performance of all 24 classifiers on their validation set and by taking the average, which resulted in mean sensitivity, mean specificity, and mean accuracy of 78.0%, 81.0%, and 80.0% respectively.\n\n##### Electronic Survey and Participants\n\nFifteen participants were invited by personal invitation via their university email accounts by the principal investigator (TJB) based on previous cooperation. Three had to be excluded because one reported to not have time to do the second survey properly, and 2 did not attempt or finish the survey in the proposed timeframe, leaving 12 dermatologists (all board-certified and coauthors of this publication) from 9 German university hospitals for analysis. The Google Forms closed web-based survey was not anonymous and each participant had to disclose his or her full name and additional metadata. For part I, the dermatologists were shown images of biopsy-verified skin lesions and asked 3 mandatory questions. First, a personal rating of the image quality had to be given (excellent, good, sufficient, poor, other image problems, no image visible). Second, the participant was asked for a diagnosis of the displayed lesion with the options being benign or malignant. Third, the participant was requested to quantify his or her confidence in the decision on a scale of 0 (=very uncertain) to 10 (=very certain). Part II showed identical images and questions and was additionally labeled with the dermatologist\u2019s previous response and the diagnosis of the CNN together with its confidence level. For details on the participants, see Multimedia Appendix 1.\n\n##### Performance and Statistical Analysis\n\nThe primary endpoint of this study was whether dermatologists\u2019 accuracy (overall proportion of the correct predictions among the total number of examined cases) would increase with AI support with the secondary endpoint investigating how sensitivity and specificity would change (tested for significance using Wilcoxon test). Additional analyses investigated whether dermatologists experienced a learning effect when working with AI support (tested for significance using Pearson\u2019s chi-squared test) and how they would respond to this form of computer aid. Statistical significance was considered at P<.05, corrected to 0.016 (Bonferroni correction for the primary and secondary endpoint) to account for multiple testing. The details on how survey answers were evaluated are listed in Multimedia Appendix 1. As a pair of classifiers was responsible for scoring the test set of one participant (one classifier for survey 1, the other for survey 2), the results of both classifiers were combined for analysis reasons so that every dermatologist had a respective counterpart. Values shown in the result sections are based on the mean of all the participants. If absolute values are shown, the denominator and numerator are approximated. In theory, every dermatologist rated the exact same number of images, but due to the quality control step, not every rating was counted. Therefore, each dermatologist ended up rating a varying number of images.",
        "results": "### Results\n\n##### Classification Results\n\nIn the first part of the study, the dermatologists had an overall mean sensitivity of 59.4% (95% CI 53.3%-65.5%), specificity of 70.6% (95% CI 62.3%-78.9%), and accuracy of 65.0% (95% CI 62.3%-67.6%, see Table 1).\n\nIn the second part of the study, where participants could integrate the results of the CNN-based classifier in their decision-making, their overall sensitivity increased significantly to 74.6% (95% CI 69.9%-79.3%; _P=.003). Their mean specificity also showed a positive trend (72.4%; 95% CI 66.2%-78.6%; P=.54), so that the overall accuracy also increased significantly to 73.6% (95% CI 70.9%-76.3%; P=.002).\n\nThe CNN on its own had an even higher sensitivity (84.7%; 95% CI 81.9%-87.6%), specificity (79.1%; 95% CI 74.8%-83.4%), and accuracy (81.9%; 95% CI 79.7%-84.2%) on the test set, which was comparable to its performance on the validation set. The overall performance of the dermatologists and the CNN are summarized in Table 1 while individual performances are captured in Figure 1 and Figure 2, which show an overview of the various metrics for every participant with and without AI support.\n\n**Table 1.** Overall mean performance of dermatologists without artificial intelligence (AI) support (-AI) compared to that of dermatologists with AI support (+AI) and AI on its own. Performance is split into 3 categories, that is, sensitivity, specificity, and accuracy.\n\nPerformance AI Dermatologist -AI Dermatologist +AI\n\nSensitivity (95% CI) 84.7% (81.9%-87.6%) 59.4% (53.3%-65.5%) 74.6% (69.9%-79.3%)\n\nSpecificity (95% CI) 79.1% (74.8%-83.4%) 70.6% (62.3%-78.9%) 72.4% (66.2%-78.6%)\n\nAccuracy (95% CI) 81.9% (79.7%-84.2%) 65.0% (62.3%-67.6%) 73.6% (70.9%-76.3%)\n\n**Figure 1.** Combined and individual dermatologists\u2019 performance without and with artificial intelligence (AI) support. Every dot represents a single participant. A line between 2 dots connects the participants' metric without AI support to the corresponding metric with AI support. Highlighted dots represent the participants combined. Boxes indicate 25th and 75th percentile while the horizontal line within shows the median (50th percentile). Whiskers indicate the data range (1.5*IQR) where points beyond are considered as outliers.\n\n**Figure 2.** Combined and individual dermatologists\u2019 diagnostic accuracy without and with artificial intelligence (AI) support. Diagnostic accuracy is measured using sensitivity and specificity. Arrows represent the change in the diagnostic accuracy from without AI support to with AI support. Highlighted arrows represent the participants combined. In addition, the black curve denotes the mean receiver operating characteristic curve of the classifier. ROC: receiver operating characteristic; CNN: convolutional neural network; AUC: area under the curve.\n\n##### Effects of the Use of the Classifier on the Decisions of the Dermatologists\n\nThe mean overall confidence of the dermatologists increased only marginally with the use of the classifier (65.0%; 95% CI 58.5%-71.6% without AI to 66.8%; 95% CI 60.2%-73.4% with AI). The confidence of the classifier in its diagnostic decisions was at 65.0% (95% CI 62.8%-67.4%). Upon detailed analysis, one can detect opposite effects on the dermatologists\u2019 confidence and switching behavior depending on agreement or disagreement with the classifier. On average, dermatologists and classifiers agreed in 63% (59/94; 95% CI 59.4%-66.1%) of cases. Among these, 8% (7/94) of cases were diagnosed incorrectly and 55% (52/94) of cases were classified correctly (see Table 2).\n\n**Table 2.** Distribution of the correct and incorrect predictions by classifier and dermatologists without artificial intelligence (AI) support and switching in response to AI support. Percentages displayed below show the amount of times a switch did or did not occur for dermatologists when answering part II of the survey.\n\nGroupings Proportion, n (%)[a] 95% CI\n\n**Both incorrect, n=94** 7 (8) 5.5%-10.4%\n\nDermatologist switched, n=7 0 (1) 0%-2.2%\n\nDermatologist stayed, n=7 7 (99) 97.8%-100%\n\n**AI correct, Dermatologist incorrect, n=94** 25 (27) 24.0%-30.1%\n\nDermatologist switched, n=25 11 (46) 33.1%-58.4%\n\nDermatologist stayed, n=25 14 (54) 41.6%-66.9%\n\n**AI incorrect, Dermatologist correct, n=94** 10 (10) 8.4%-11.8%\n\nDermatologist switched, n=10 4 (39) 23.2%-55.6%\n\nDermatologist stayed, n=10 6 (61) 44.4%-76.9%\n\n**Both correct, n=94** 52 (55) 52.4%-57.1%\n\nDermatologist switched, n=52 0 (0) 0%-0.5%\n\nDermatologist stayed, n=52 52 (100) 99.5%-100%\n\naAs the mean of all the participants was taken and every participant ended up rating a varying amount of images due to the quality control step, the reported absolute values are approximations.\n\nOf the 37% (35/94; 95% CI 33.9%-40.6%) of cases in which dermatologists and classifier came to different conclusions, 10% (10/94) of cases were diagnosed correctly by the dermatologists and 27% (25/94) of cases were diagnosed correctly by the classifier.\n\nIn cases of agreement, the mean confidence of the dermatologists increased substantially from the first part of the study to the second part of the study (from 67.0%; 95% CI 60.6%-73.5% to 79.1%; 95% CI 75.0%-83.0%, respectively). In cases of disagreement, it decreased from 61.7% (95% CI 54.6%-68.9%) in the first to 44.3% (95% CI 31.8%-56.7%) in the second part of the study. This was also reflected in the dermatologists\u2019 switching behavior. In cases of agreement, the dermatologists basically never altered their classifications (0/59, 0%; 95% CI 0%-0.8%). In contrast, in cases of disagreement, the dermatologists altered their decision in 43% (15/35; 95% CI 30.8%-56.3%) of those cases. The dermatologists altered their diagnosis less often when they had initially diagnosed the lesion correctly (subsequently changed answer in 4/10, 39% of those cases) than when they had initially diagnosed it incorrectly (subsequently changed answer in 11/25, 46% of those cases, see Table 2). Altogether, this resulted in the observed overall increase in the correct diagnoses by the dermatologists. Out of all the occurring switches, dermatologists showed a higher willingness to switch from benign to malignant (11/15, 72%; 95% CI 53.4%-90.9%) than from malignant to benign (4/15, 28%; 95% CI 9.1%-46.6%).\n\nThe mean confidence levels of the classifier were much higher when the correct conclusion was reached (71.7%; 95% CI 69.3%-74.1%) than when the diagnosis was incorrect (34.8%; 95% CI 30.9%-38.7%). Regarding the dermatologists, a similar trend was less striking for part I (67.3%; 95% CI 61.1%-73.5% correct vs 60.8%; 95% CI 53.3%-68.2% incorrect) but became more pronounced for part II (71.8%; 95% CI 66.4%-77.1% correct vs 52.3%; 95% CI 41.0%-63.5% incorrect). A more detailed breakdown of the dermatologists\u2019 and classifier\u2019s confidence levels is shown in Table 3. Finally, the dermatologists altered their diagnoses in divergent cases more often when the CNN\u2019s confidence levels were high than when they were low (mean CNN confidence levels 63.3%; 95% CI 56.4%-70.1% with subsequent switch vs 53.7%; 95% CI 46.7%-60.7% with no subsequent switch).\n\n**Table 3.** Confidence distribution of the classifier and dermatologists for part I and part II.\n\nGroupings Confidence (95% CI)\n\n**Both incorrect**\n\nAI[a] 35.8% (27.6%-44.1%)\n\nDermatologist part I 62.3% (54.6%-69.9%)\n\nDermatologist part II 69.8% (62.4%-77.3%)\n\n**AI correct, Dermatologist incorrect**\n\nAI 65.4% (60.7%-70.2%)\n\nDermatologist part I 60.8% (53.3%-68.3%)\n\nDermatologist part II 46.0% (33.6%-58.5%)\n\n**AI incorrect, Dermatologist correct**\n\nAI 32.9% (28.1%-37.6%)\n\nDermatologist part I 63.2% (56.3%-70.2%)\n\nDermatologist part II 47.9% (36.4%-59.4%)\n\n**Both correct**\n\nAI 74.5% (72.0%-76.9%)\n\nDermatologist part I 67.9% (61.5%-74.3%)\n\nDermatologist part II 80.3% (76.2%-84.3%)\n\naAI: artificial intelligence.\n\n##### Learning Effect\n\nThe dermatologists received their results immediately after completing the first survey of the second part of the study. Thus, they had the opportunity to implement the lessons learned in the first AI-supported survey during the completion of the second survey in part II. A subanalysis showed that there was no detectable change in the dermatologists\u2019 performance from the first to the second survey in part I of the study (see figure in Multimedia Appendix 2) with no significant difference between sensitivity (P=.50) and specificity (P=.76). In contrast, the dermatologists tended to perform better in the second survey of part II compared to the first survey, suggesting that they had learned how to better incorporate the CNN results into their diagnostic procedures, albeit with differences in sensitivity (P=.21) and specificity (P=.43) being still insignificant. In divergent cases, participants switched their diagnosis in 49% (8/17; 95% CI 36.2%-62.3%) of cases in survey 2, whereas they had only done so in 38% (7/18; 95% CI 21.1%-54.5%) of these cases for survey 1. The tendency to switch at the correct moments in cases of disagreement was reinforced in the second survey wherein dermatologists switched far more often when they were incorrect (5/13, 39%; 95% CI 23.0%-55.7% in the first survey vs 6/12, 52%; 95% CI 39.8%-64.3% in the second survey).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Artificial Intelligence and Its Effect on Dermatologists' Accuracy in Dermoscopic Melanoma Image Classification: Web-Based Survey Study.",
        "abstract": "## BACKGROUND\nEarly detection of melanoma can be lifesaving but this remains a challenge. Recent diagnostic studies have revealed the superiority of artificial intelligence (AI) in classifying dermoscopic images of melanoma and nevi, concluding that these algorithms should assist a dermatologist's diagnoses.\n## OBJECTIVE\nThe aim of this study was to investigate whether AI support improves the accuracy and overall diagnostic performance of dermatologists in the dichotomous image-based discrimination between melanoma and nevus.\n## METHODS\nTwelve board-certified dermatologists were presented disjoint sets of 100 unique dermoscopic images of melanomas and nevi (total of 1200 unique images), and they had to classify the images based on personal experience alone (part I) and with the support of a trained convolutional neural network (CNN, part II). Additionally, dermatologists were asked to rate their confidence in their final decision for each image.\n## RESULTS\nWhile the mean specificity of the dermatologists based on personal experience alone remained almost unchanged (70.6% vs 72.4%; P=.54) with AI support, the mean sensitivity and mean accuracy increased significantly (59.4% vs 74.6%; P=.003 and 65.0% vs 73.6%; P=.002, respectively) with AI support. Out of the 10% (10/94; 95% CI 8.4%-11.8%) of cases where dermatologists were correct and AI was incorrect, dermatologists on average changed to the incorrect answer for 39% (4/10; 95% CI 23.2%-55.6%) of cases. When dermatologists were incorrect and AI was correct (25/94, 27%; 95% CI 24.0%-30.1%), dermatologists changed their answers to the correct answer for 46% (11/25; 95% CI 33.1%-58.4%) of cases. Additionally, the dermatologists' average confidence in their decisions increased when the CNN confirmed their decision and decreased when the CNN disagreed, even when the dermatologists were correct. Reported values are based on the mean of all participants. Whenever absolute values are shown, the denominator and numerator are approximations as every dermatologist ended up rating a varying number of images due to a quality control step.\n## CONCLUSIONS\nThe findings of our study show that AI support can improve the overall accuracy of the dermatologists in the dichotomous image-based discrimination between melanoma and nevus. This supports the argument for AI-based tools to aid clinicians in skin lesion classification and provides a rationale for studies of such classifiers in real-life settings, wherein clinicians can integrate additional information such as patient age and medical history into their decisions.\n",
        "methods": "### Methods\n\n##### Study Design\n\nThis study was conducted from January 10, 2019 (design of study) to September 27, 2019 and was inspired by the study design of Sinz et al [2]. Questionnaires were sent out on June 12, 2019 and were completed by August 12, 2019. Ethical approval was waived by the ethics committee of the University of Heidelberg, Mannheim Faculty of Medicine as all the images were open source and anonymous and all participating dermatologists automatically became part of the study group.\n\nThe set of images used for the evaluation consisted of 1200 unique dermoscopic images split into 12 nonoverlapping individual test sets containing 100 images each (50 melanomas and 50 nevi). Each individual set was randomly assigned to exactly one dermatologist who diagnosed his/her set twice\u2014first without AI support (part I) and afterwards with AI support (part II). Images for each part were presented to the participants in 2 separate surveys containing 50 images each (again 50:50 split) to investigate a possible learning effect. Thus, every participant completed 4 surveys (2 surveys for part I and 2 surveys for part II) where images seen in part I were identical to those seen in part II. Both surveys from part I were carried out before any survey from part II. The participants received detailed instructions by email in which the survey structure and the procedure were discussed and the lesion distribution was disclosed. For part II, participants were made aware of the classifier\u2019s performance, which was established beforehand on a separate validation set.\n\n##### Classifier Training\n\nImages were obtained from the ISIC archive [19], with a large fraction of images coming from the HAM10000 dataset [20]. The archive is publicly accessible and contains anonymous dermoscopic images from multiple sources and various camera systems. From the available pigmented lesions, only images showing a biopsy-verified melanoma (n=1633) or nevus (n=3311) were selected (total n=4944).\n\nAs each of the 12 individual test sets consisted of 2 subsets (with 50 images each), there were 24 test subsets. A designated training set was constructed for each test subset by removing its images from the total image data, resulting in 24 training sets, each containing 4894 images (ie, 4944 \u2013 50). The class distribution across each training set was unevenly distributed with 1608 melanoma images in contrast to 3286 nevus images. To counter the class imbalance, the set of melanoma images was duplicated. Online data augmentation was used to increase the diversity of the training data and to modify the duplicated images. Before balancing, an evenly balanced validation set was removed from each training set for calibration and validation purposes later. A detailed description of the classifier training is given in Multimedia Appendix 1.\n\n##### Classifier Integration\n\nFor AI support to be used in an effective manner, a confidence measure was displayed in addition to the binary decision (melanoma or nevus), which indicated the classifier\u2019s confidence in its decision. The measure was obtained by mapping the classifier\u2019s raw output probability for the predicted class into a percentage range from 0% to 100%.\n\nParticipants were made aware of the classifier\u2019s overall performance by establishing the overall performance of all 24 classifiers on their validation set and by taking the average, which resulted in mean sensitivity, mean specificity, and mean accuracy of 78.0%, 81.0%, and 80.0% respectively.\n\n##### Electronic Survey and Participants\n\nFifteen participants were invited by personal invitation via their university email accounts by the principal investigator (TJB) based on previous cooperation. Three had to be excluded because one reported to not have time to do the second survey properly, and 2 did not attempt or finish the survey in the proposed timeframe, leaving 12 dermatologists (all board-certified and coauthors of this publication) from 9 German university hospitals for analysis. The Google Forms closed web-based survey was not anonymous and each participant had to disclose his or her full name and additional metadata. For part I, the dermatologists were shown images of biopsy-verified skin lesions and asked 3 mandatory questions. First, a personal rating of the image quality had to be given (excellent, good, sufficient, poor, other image problems, no image visible). Second, the participant was asked for a diagnosis of the displayed lesion with the options being benign or malignant. Third, the participant was requested to quantify his or her confidence in the decision on a scale of 0 (=very uncertain) to 10 (=very certain). Part II showed identical images and questions and was additionally labeled with the dermatologist\u2019s previous response and the diagnosis of the CNN together with its confidence level. For details on the participants, see Multimedia Appendix 1.\n\n##### Performance and Statistical Analysis\n\nThe primary endpoint of this study was whether dermatologists\u2019 accuracy (overall proportion of the correct predictions among the total number of examined cases) would increase with AI support with the secondary endpoint investigating how sensitivity and specificity would change (tested for significance using Wilcoxon test). Additional analyses investigated whether dermatologists experienced a learning effect when working with AI support (tested for significance using Pearson\u2019s chi-squared test) and how they would respond to this form of computer aid. Statistical significance was considered at P<.05, corrected to 0.016 (Bonferroni correction for the primary and secondary endpoint) to account for multiple testing. The details on how survey answers were evaluated are listed in Multimedia Appendix 1. As a pair of classifiers was responsible for scoring the test set of one participant (one classifier for survey 1, the other for survey 2), the results of both classifiers were combined for analysis reasons so that every dermatologist had a respective counterpart. Values shown in the result sections are based on the mean of all the participants. If absolute values are shown, the denominator and numerator are approximated. In theory, every dermatologist rated the exact same number of images, but due to the quality control step, not every rating was counted. Therefore, each dermatologist ended up rating a varying number of images.",
        "results": "### Results\n\n##### Classification Results\n\nIn the first part of the study, the dermatologists had an overall mean sensitivity of 59.4% (95% CI 53.3%-65.5%), specificity of 70.6% (95% CI 62.3%-78.9%), and accuracy of 65.0% (95% CI 62.3%-67.6%, see Table 1).\n\nIn the second part of the study, where participants could integrate the results of the CNN-based classifier in their decision-making, their overall sensitivity increased significantly to 74.6% (95% CI 69.9%-79.3%; _P=.003). Their mean specificity also showed a positive trend (72.4%; 95% CI 66.2%-78.6%; P=.54), so that the overall accuracy also increased significantly to 73.6% (95% CI 70.9%-76.3%; P=.002).\n\nThe CNN on its own had an even higher sensitivity (84.7%; 95% CI 81.9%-87.6%), specificity (79.1%; 95% CI 74.8%-83.4%), and accuracy (81.9%; 95% CI 79.7%-84.2%) on the test set, which was comparable to its performance on the validation set. The overall performance of the dermatologists and the CNN are summarized in Table 1 while individual performances are captured in Figure 1 and Figure 2, which show an overview of the various metrics for every participant with and without AI support.\n\n**Table 1.** Overall mean performance of dermatologists without artificial intelligence (AI) support (-AI) compared to that of dermatologists with AI support (+AI) and AI on its own. Performance is split into 3 categories, that is, sensitivity, specificity, and accuracy.\n\nPerformance AI Dermatologist -AI Dermatologist +AI\n\nSensitivity (95% CI) 84.7% (81.9%-87.6%) 59.4% (53.3%-65.5%) 74.6% (69.9%-79.3%)\n\nSpecificity (95% CI) 79.1% (74.8%-83.4%) 70.6% (62.3%-78.9%) 72.4% (66.2%-78.6%)\n\nAccuracy (95% CI) 81.9% (79.7%-84.2%) 65.0% (62.3%-67.6%) 73.6% (70.9%-76.3%)\n\n**Figure 1.** Combined and individual dermatologists\u2019 performance without and with artificial intelligence (AI) support. Every dot represents a single participant. A line between 2 dots connects the participants' metric without AI support to the corresponding metric with AI support. Highlighted dots represent the participants combined. Boxes indicate 25th and 75th percentile while the horizontal line within shows the median (50th percentile). Whiskers indicate the data range (1.5*IQR) where points beyond are considered as outliers.\n\n**Figure 2.** Combined and individual dermatologists\u2019 diagnostic accuracy without and with artificial intelligence (AI) support. Diagnostic accuracy is measured using sensitivity and specificity. Arrows represent the change in the diagnostic accuracy from without AI support to with AI support. Highlighted arrows represent the participants combined. In addition, the black curve denotes the mean receiver operating characteristic curve of the classifier. ROC: receiver operating characteristic; CNN: convolutional neural network; AUC: area under the curve.\n\n##### Effects of the Use of the Classifier on the Decisions of the Dermatologists\n\nThe mean overall confidence of the dermatologists increased only marginally with the use of the classifier (65.0%; 95% CI 58.5%-71.6% without AI to 66.8%; 95% CI 60.2%-73.4% with AI). The confidence of the classifier in its diagnostic decisions was at 65.0% (95% CI 62.8%-67.4%). Upon detailed analysis, one can detect opposite effects on the dermatologists\u2019 confidence and switching behavior depending on agreement or disagreement with the classifier. On average, dermatologists and classifiers agreed in 63% (59/94; 95% CI 59.4%-66.1%) of cases. Among these, 8% (7/94) of cases were diagnosed incorrectly and 55% (52/94) of cases were classified correctly (see Table 2).\n\n**Table 2.** Distribution of the correct and incorrect predictions by classifier and dermatologists without artificial intelligence (AI) support and switching in response to AI support. Percentages displayed below show the amount of times a switch did or did not occur for dermatologists when answering part II of the survey.\n\nGroupings Proportion, n (%)[a] 95% CI\n\n**Both incorrect, n=94** 7 (8) 5.5%-10.4%\n\nDermatologist switched, n=7 0 (1) 0%-2.2%\n\nDermatologist stayed, n=7 7 (99) 97.8%-100%\n\n**AI correct, Dermatologist incorrect, n=94** 25 (27) 24.0%-30.1%\n\nDermatologist switched, n=25 11 (46) 33.1%-58.4%\n\nDermatologist stayed, n=25 14 (54) 41.6%-66.9%\n\n**AI incorrect, Dermatologist correct, n=94** 10 (10) 8.4%-11.8%\n\nDermatologist switched, n=10 4 (39) 23.2%-55.6%\n\nDermatologist stayed, n=10 6 (61) 44.4%-76.9%\n\n**Both correct, n=94** 52 (55) 52.4%-57.1%\n\nDermatologist switched, n=52 0 (0) 0%-0.5%\n\nDermatologist stayed, n=52 52 (100) 99.5%-100%\n\naAs the mean of all the participants was taken and every participant ended up rating a varying amount of images due to the quality control step, the reported absolute values are approximations.\n\nOf the 37% (35/94; 95% CI 33.9%-40.6%) of cases in which dermatologists and classifier came to different conclusions, 10% (10/94) of cases were diagnosed correctly by the dermatologists and 27% (25/94) of cases were diagnosed correctly by the classifier.\n\nIn cases of agreement, the mean confidence of the dermatologists increased substantially from the first part of the study to the second part of the study (from 67.0%; 95% CI 60.6%-73.5% to 79.1%; 95% CI 75.0%-83.0%, respectively). In cases of disagreement, it decreased from 61.7% (95% CI 54.6%-68.9%) in the first to 44.3% (95% CI 31.8%-56.7%) in the second part of the study. This was also reflected in the dermatologists\u2019 switching behavior. In cases of agreement, the dermatologists basically never altered their classifications (0/59, 0%; 95% CI 0%-0.8%). In contrast, in cases of disagreement, the dermatologists altered their decision in 43% (15/35; 95% CI 30.8%-56.3%) of those cases. The dermatologists altered their diagnosis less often when they had initially diagnosed the lesion correctly (subsequently changed answer in 4/10, 39% of those cases) than when they had initially diagnosed it incorrectly (subsequently changed answer in 11/25, 46% of those cases, see Table 2). Altogether, this resulted in the observed overall increase in the correct diagnoses by the dermatologists. Out of all the occurring switches, dermatologists showed a higher willingness to switch from benign to malignant (11/15, 72%; 95% CI 53.4%-90.9%) than from malignant to benign (4/15, 28%; 95% CI 9.1%-46.6%).\n\nThe mean confidence levels of the classifier were much higher when the correct conclusion was reached (71.7%; 95% CI 69.3%-74.1%) than when the diagnosis was incorrect (34.8%; 95% CI 30.9%-38.7%). Regarding the dermatologists, a similar trend was less striking for part I (67.3%; 95% CI 61.1%-73.5% correct vs 60.8%; 95% CI 53.3%-68.2% incorrect) but became more pronounced for part II (71.8%; 95% CI 66.4%-77.1% correct vs 52.3%; 95% CI 41.0%-63.5% incorrect). A more detailed breakdown of the dermatologists\u2019 and classifier\u2019s confidence levels is shown in Table 3. Finally, the dermatologists altered their diagnoses in divergent cases more often when the CNN\u2019s confidence levels were high than when they were low (mean CNN confidence levels 63.3%; 95% CI 56.4%-70.1% with subsequent switch vs 53.7%; 95% CI 46.7%-60.7% with no subsequent switch).\n\n**Table 3.** Confidence distribution of the classifier and dermatologists for part I and part II.\n\nGroupings Confidence (95% CI)\n\n**Both incorrect**\n\nAI[a] 35.8% (27.6%-44.1%)\n\nDermatologist part I 62.3% (54.6%-69.9%)\n\nDermatologist part II 69.8% (62.4%-77.3%)\n\n**AI correct, Dermatologist incorrect**\n\nAI 65.4% (60.7%-70.2%)\n\nDermatologist part I 60.8% (53.3%-68.3%)\n\nDermatologist part II 46.0% (33.6%-58.5%)\n\n**AI incorrect, Dermatologist correct**\n\nAI 32.9% (28.1%-37.6%)\n\nDermatologist part I 63.2% (56.3%-70.2%)\n\nDermatologist part II 47.9% (36.4%-59.4%)\n\n**Both correct**\n\nAI 74.5% (72.0%-76.9%)\n\nDermatologist part I 67.9% (61.5%-74.3%)\n\nDermatologist part II 80.3% (76.2%-84.3%)\n\naAI: artificial intelligence.\n\n##### Learning Effect\n\nThe dermatologists received their results immediately after completing the first survey of the second part of the study. Thus, they had the opportunity to implement the lessons learned in the first AI-supported survey during the completion of the second survey in part II. A subanalysis showed that there was no detectable change in the dermatologists\u2019 performance from the first to the second survey in part I of the study (see figure in Multimedia Appendix 2) with no significant difference between sensitivity (P=.50) and specificity (P=.76). In contrast, the dermatologists tended to perform better in the second survey of part II compared to the first survey, suggesting that they had learned how to better incorporate the CNN results into their diagnostic procedures, albeit with differences in sensitivity (P=.21) and specificity (P=.43) being still insignificant. In divergent cases, participants switched their diagnosis in 49% (8/17; 95% CI 36.2%-62.3%) of cases in survey 2, whereas they had only done so in 38% (7/18; 95% CI 21.1%-54.5%) of these cases for survey 1. The tendency to switch at the correct moments in cases of disagreement was reinforced in the second survey wherein dermatologists switched far more often when they were incorrect (5/13, 39%; 95% CI 23.0%-55.7% in the first survey vs 6/12, 52%; 95% CI 39.8%-64.3% in the second survey).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Artificial Intelligence and Its Effect on Dermatologists' Accuracy in Dermoscopic Melanoma Image Classification: Web-Based Survey Study.",
        "abstract": "## BACKGROUND\nEarly detection of melanoma can be lifesaving but this remains a challenge. Recent diagnostic studies have revealed the superiority of artificial intelligence (AI) in classifying dermoscopic images of melanoma and nevi, concluding that these algorithms should assist a dermatologist's diagnoses.\n## OBJECTIVE\nThe aim of this study was to investigate whether AI support improves the accuracy and overall diagnostic performance of dermatologists in the dichotomous image-based discrimination between melanoma and nevus.\n## METHODS\nTwelve board-certified dermatologists were presented disjoint sets of 100 unique dermoscopic images of melanomas and nevi (total of 1200 unique images), and they had to classify the images based on personal experience alone (part I) and with the support of a trained convolutional neural network (CNN, part II). Additionally, dermatologists were asked to rate their confidence in their final decision for each image.\n## RESULTS\nWhile the mean specificity of the dermatologists based on personal experience alone remained almost unchanged (70.6% vs 72.4%; P=.54) with AI support, the mean sensitivity and mean accuracy increased significantly (59.4% vs 74.6%; P=.003 and 65.0% vs 73.6%; P=.002, respectively) with AI support. Out of the 10% (10/94; 95% CI 8.4%-11.8%) of cases where dermatologists were correct and AI was incorrect, dermatologists on average changed to the incorrect answer for 39% (4/10; 95% CI 23.2%-55.6%) of cases. When dermatologists were incorrect and AI was correct (25/94, 27%; 95% CI 24.0%-30.1%), dermatologists changed their answers to the correct answer for 46% (11/25; 95% CI 33.1%-58.4%) of cases. Additionally, the dermatologists' average confidence in their decisions increased when the CNN confirmed their decision and decreased when the CNN disagreed, even when the dermatologists were correct. Reported values are based on the mean of all participants. Whenever absolute values are shown, the denominator and numerator are approximations as every dermatologist ended up rating a varying number of images due to a quality control step.\n## CONCLUSIONS\nThe findings of our study show that AI support can improve the overall accuracy of the dermatologists in the dichotomous image-based discrimination between melanoma and nevus. This supports the argument for AI-based tools to aid clinicians in skin lesion classification and provides a rationale for studies of such classifiers in real-life settings, wherein clinicians can integrate additional information such as patient age and medical history into their decisions.\n",
        "methods": "### Methods\n\n##### Study Design\n\nThis study was conducted from January 10, 2019 (design of study) to September 27, 2019 and was inspired by the study design of Sinz et al [2]. Questionnaires were sent out on June 12, 2019 and were completed by August 12, 2019. Ethical approval was waived by the ethics committee of the University of Heidelberg, Mannheim Faculty of Medicine as all the images were open source and anonymous and all participating dermatologists automatically became part of the study group.\n\nThe set of images used for the evaluation consisted of 1200 unique dermoscopic images split into 12 nonoverlapping individual test sets containing 100 images each (50 melanomas and 50 nevi). Each individual set was randomly assigned to exactly one dermatologist who diagnosed his/her set twice\u2014first without AI support (part I) and afterwards with AI support (part II). Images for each part were presented to the participants in 2 separate surveys containing 50 images each (again 50:50 split) to investigate a possible learning effect. Thus, every participant completed 4 surveys (2 surveys for part I and 2 surveys for part II) where images seen in part I were identical to those seen in part II. Both surveys from part I were carried out before any survey from part II. The participants received detailed instructions by email in which the survey structure and the procedure were discussed and the lesion distribution was disclosed. For part II, participants were made aware of the classifier\u2019s performance, which was established beforehand on a separate validation set.\n\n##### Classifier Training\n\nImages were obtained from the ISIC archive [19], with a large fraction of images coming from the HAM10000 dataset [20]. The archive is publicly accessible and contains anonymous dermoscopic images from multiple sources and various camera systems. From the available pigmented lesions, only images showing a biopsy-verified melanoma (n=1633) or nevus (n=3311) were selected (total n=4944).\n\nAs each of the 12 individual test sets consisted of 2 subsets (with 50 images each), there were 24 test subsets. A designated training set was constructed for each test subset by removing its images from the total image data, resulting in 24 training sets, each containing 4894 images (ie, 4944 \u2013 50). The class distribution across each training set was unevenly distributed with 1608 melanoma images in contrast to 3286 nevus images. To counter the class imbalance, the set of melanoma images was duplicated. Online data augmentation was used to increase the diversity of the training data and to modify the duplicated images. Before balancing, an evenly balanced validation set was removed from each training set for calibration and validation purposes later. A detailed description of the classifier training is given in Multimedia Appendix 1.\n\n##### Classifier Integration\n\nFor AI support to be used in an effective manner, a confidence measure was displayed in addition to the binary decision (melanoma or nevus), which indicated the classifier\u2019s confidence in its decision. The measure was obtained by mapping the classifier\u2019s raw output probability for the predicted class into a percentage range from 0% to 100%.\n\nParticipants were made aware of the classifier\u2019s overall performance by establishing the overall performance of all 24 classifiers on their validation set and by taking the average, which resulted in mean sensitivity, mean specificity, and mean accuracy of 78.0%, 81.0%, and 80.0% respectively.\n\n##### Electronic Survey and Participants\n\nFifteen participants were invited by personal invitation via their university email accounts by the principal investigator (TJB) based on previous cooperation. Three had to be excluded because one reported to not have time to do the second survey properly, and 2 did not attempt or finish the survey in the proposed timeframe, leaving 12 dermatologists (all board-certified and coauthors of this publication) from 9 German university hospitals for analysis. The Google Forms closed web-based survey was not anonymous and each participant had to disclose his or her full name and additional metadata. For part I, the dermatologists were shown images of biopsy-verified skin lesions and asked 3 mandatory questions. First, a personal rating of the image quality had to be given (excellent, good, sufficient, poor, other image problems, no image visible). Second, the participant was asked for a diagnosis of the displayed lesion with the options being benign or malignant. Third, the participant was requested to quantify his or her confidence in the decision on a scale of 0 (=very uncertain) to 10 (=very certain). Part II showed identical images and questions and was additionally labeled with the dermatologist\u2019s previous response and the diagnosis of the CNN together with its confidence level. For details on the participants, see Multimedia Appendix 1.\n\n##### Performance and Statistical Analysis\n\nThe primary endpoint of this study was whether dermatologists\u2019 accuracy (overall proportion of the correct predictions among the total number of examined cases) would increase with AI support with the secondary endpoint investigating how sensitivity and specificity would change (tested for significance using Wilcoxon test). Additional analyses investigated whether dermatologists experienced a learning effect when working with AI support (tested for significance using Pearson\u2019s chi-squared test) and how they would respond to this form of computer aid. Statistical significance was considered at P<.05, corrected to 0.016 (Bonferroni correction for the primary and secondary endpoint) to account for multiple testing. The details on how survey answers were evaluated are listed in Multimedia Appendix 1. As a pair of classifiers was responsible for scoring the test set of one participant (one classifier for survey 1, the other for survey 2), the results of both classifiers were combined for analysis reasons so that every dermatologist had a respective counterpart. Values shown in the result sections are based on the mean of all the participants. If absolute values are shown, the denominator and numerator are approximated. In theory, every dermatologist rated the exact same number of images, but due to the quality control step, not every rating was counted. Therefore, each dermatologist ended up rating a varying number of images.",
        "results": "### Results\n\n##### Classification Results\n\nIn the first part of the study, the dermatologists had an overall mean sensitivity of 59.4% (95% CI 53.3%-65.5%), specificity of 70.6% (95% CI 62.3%-78.9%), and accuracy of 65.0% (95% CI 62.3%-67.6%, see Table 1).\n\nIn the second part of the study, where participants could integrate the results of the CNN-based classifier in their decision-making, their overall sensitivity increased significantly to 74.6% (95% CI 69.9%-79.3%; _P=.003). Their mean specificity also showed a positive trend (72.4%; 95% CI 66.2%-78.6%; P=.54), so that the overall accuracy also increased significantly to 73.6% (95% CI 70.9%-76.3%; P=.002).\n\nThe CNN on its own had an even higher sensitivity (84.7%; 95% CI 81.9%-87.6%), specificity (79.1%; 95% CI 74.8%-83.4%), and accuracy (81.9%; 95% CI 79.7%-84.2%) on the test set, which was comparable to its performance on the validation set. The overall performance of the dermatologists and the CNN are summarized in Table 1 while individual performances are captured in Figure 1 and Figure 2, which show an overview of the various metrics for every participant with and without AI support.\n\n**Table 1.** Overall mean performance of dermatologists without artificial intelligence (AI) support (-AI) compared to that of dermatologists with AI support (+AI) and AI on its own. Performance is split into 3 categories, that is, sensitivity, specificity, and accuracy.\n\nPerformance AI Dermatologist -AI Dermatologist +AI\n\nSensitivity (95% CI) 84.7% (81.9%-87.6%) 59.4% (53.3%-65.5%) 74.6% (69.9%-79.3%)\n\nSpecificity (95% CI) 79.1% (74.8%-83.4%) 70.6% (62.3%-78.9%) 72.4% (66.2%-78.6%)\n\nAccuracy (95% CI) 81.9% (79.7%-84.2%) 65.0% (62.3%-67.6%) 73.6% (70.9%-76.3%)\n\n**Figure 1.** Combined and individual dermatologists\u2019 performance without and with artificial intelligence (AI) support. Every dot represents a single participant. A line between 2 dots connects the participants' metric without AI support to the corresponding metric with AI support. Highlighted dots represent the participants combined. Boxes indicate 25th and 75th percentile while the horizontal line within shows the median (50th percentile). Whiskers indicate the data range (1.5*IQR) where points beyond are considered as outliers.\n\n**Figure 2.** Combined and individual dermatologists\u2019 diagnostic accuracy without and with artificial intelligence (AI) support. Diagnostic accuracy is measured using sensitivity and specificity. Arrows represent the change in the diagnostic accuracy from without AI support to with AI support. Highlighted arrows represent the participants combined. In addition, the black curve denotes the mean receiver operating characteristic curve of the classifier. ROC: receiver operating characteristic; CNN: convolutional neural network; AUC: area under the curve.\n\n##### Effects of the Use of the Classifier on the Decisions of the Dermatologists\n\nThe mean overall confidence of the dermatologists increased only marginally with the use of the classifier (65.0%; 95% CI 58.5%-71.6% without AI to 66.8%; 95% CI 60.2%-73.4% with AI). The confidence of the classifier in its diagnostic decisions was at 65.0% (95% CI 62.8%-67.4%). Upon detailed analysis, one can detect opposite effects on the dermatologists\u2019 confidence and switching behavior depending on agreement or disagreement with the classifier. On average, dermatologists and classifiers agreed in 63% (59/94; 95% CI 59.4%-66.1%) of cases. Among these, 8% (7/94) of cases were diagnosed incorrectly and 55% (52/94) of cases were classified correctly (see Table 2).\n\n**Table 2.** Distribution of the correct and incorrect predictions by classifier and dermatologists without artificial intelligence (AI) support and switching in response to AI support. Percentages displayed below show the amount of times a switch did or did not occur for dermatologists when answering part II of the survey.\n\nGroupings Proportion, n (%)[a] 95% CI\n\n**Both incorrect, n=94** 7 (8) 5.5%-10.4%\n\nDermatologist switched, n=7 0 (1) 0%-2.2%\n\nDermatologist stayed, n=7 7 (99) 97.8%-100%\n\n**AI correct, Dermatologist incorrect, n=94** 25 (27) 24.0%-30.1%\n\nDermatologist switched, n=25 11 (46) 33.1%-58.4%\n\nDermatologist stayed, n=25 14 (54) 41.6%-66.9%\n\n**AI incorrect, Dermatologist correct, n=94** 10 (10) 8.4%-11.8%\n\nDermatologist switched, n=10 4 (39) 23.2%-55.6%\n\nDermatologist stayed, n=10 6 (61) 44.4%-76.9%\n\n**Both correct, n=94** 52 (55) 52.4%-57.1%\n\nDermatologist switched, n=52 0 (0) 0%-0.5%\n\nDermatologist stayed, n=52 52 (100) 99.5%-100%\n\naAs the mean of all the participants was taken and every participant ended up rating a varying amount of images due to the quality control step, the reported absolute values are approximations.\n\nOf the 37% (35/94; 95% CI 33.9%-40.6%) of cases in which dermatologists and classifier came to different conclusions, 10% (10/94) of cases were diagnosed correctly by the dermatologists and 27% (25/94) of cases were diagnosed correctly by the classifier.\n\nIn cases of agreement, the mean confidence of the dermatologists increased substantially from the first part of the study to the second part of the study (from 67.0%; 95% CI 60.6%-73.5% to 79.1%; 95% CI 75.0%-83.0%, respectively). In cases of disagreement, it decreased from 61.7% (95% CI 54.6%-68.9%) in the first to 44.3% (95% CI 31.8%-56.7%) in the second part of the study. This was also reflected in the dermatologists\u2019 switching behavior. In cases of agreement, the dermatologists basically never altered their classifications (0/59, 0%; 95% CI 0%-0.8%). In contrast, in cases of disagreement, the dermatologists altered their decision in 43% (15/35; 95% CI 30.8%-56.3%) of those cases. The dermatologists altered their diagnosis less often when they had initially diagnosed the lesion correctly (subsequently changed answer in 4/10, 39% of those cases) than when they had initially diagnosed it incorrectly (subsequently changed answer in 11/25, 46% of those cases, see Table 2). Altogether, this resulted in the observed overall increase in the correct diagnoses by the dermatologists. Out of all the occurring switches, dermatologists showed a higher willingness to switch from benign to malignant (11/15, 72%; 95% CI 53.4%-90.9%) than from malignant to benign (4/15, 28%; 95% CI 9.1%-46.6%).\n\nThe mean confidence levels of the classifier were much higher when the correct conclusion was reached (71.7%; 95% CI 69.3%-74.1%) than when the diagnosis was incorrect (34.8%; 95% CI 30.9%-38.7%). Regarding the dermatologists, a similar trend was less striking for part I (67.3%; 95% CI 61.1%-73.5% correct vs 60.8%; 95% CI 53.3%-68.2% incorrect) but became more pronounced for part II (71.8%; 95% CI 66.4%-77.1% correct vs 52.3%; 95% CI 41.0%-63.5% incorrect). A more detailed breakdown of the dermatologists\u2019 and classifier\u2019s confidence levels is shown in Table 3. Finally, the dermatologists altered their diagnoses in divergent cases more often when the CNN\u2019s confidence levels were high than when they were low (mean CNN confidence levels 63.3%; 95% CI 56.4%-70.1% with subsequent switch vs 53.7%; 95% CI 46.7%-60.7% with no subsequent switch).\n\n**Table 3.** Confidence distribution of the classifier and dermatologists for part I and part II.\n\nGroupings Confidence (95% CI)\n\n**Both incorrect**\n\nAI[a] 35.8% (27.6%-44.1%)\n\nDermatologist part I 62.3% (54.6%-69.9%)\n\nDermatologist part II 69.8% (62.4%-77.3%)\n\n**AI correct, Dermatologist incorrect**\n\nAI 65.4% (60.7%-70.2%)\n\nDermatologist part I 60.8% (53.3%-68.3%)\n\nDermatologist part II 46.0% (33.6%-58.5%)\n\n**AI incorrect, Dermatologist correct**\n\nAI 32.9% (28.1%-37.6%)\n\nDermatologist part I 63.2% (56.3%-70.2%)\n\nDermatologist part II 47.9% (36.4%-59.4%)\n\n**Both correct**\n\nAI 74.5% (72.0%-76.9%)\n\nDermatologist part I 67.9% (61.5%-74.3%)\n\nDermatologist part II 80.3% (76.2%-84.3%)\n\naAI: artificial intelligence.\n\n##### Learning Effect\n\nThe dermatologists received their results immediately after completing the first survey of the second part of the study. Thus, they had the opportunity to implement the lessons learned in the first AI-supported survey during the completion of the second survey in part II. A subanalysis showed that there was no detectable change in the dermatologists\u2019 performance from the first to the second survey in part I of the study (see figure in Multimedia Appendix 2) with no significant difference between sensitivity (P=.50) and specificity (P=.76). In contrast, the dermatologists tended to perform better in the second survey of part II compared to the first survey, suggesting that they had learned how to better incorporate the CNN results into their diagnostic procedures, albeit with differences in sensitivity (P=.21) and specificity (P=.43) being still insignificant. In divergent cases, participants switched their diagnosis in 49% (8/17; 95% CI 36.2%-62.3%) of cases in survey 2, whereas they had only done so in 38% (7/18; 95% CI 21.1%-54.5%) of these cases for survey 1. The tendency to switch at the correct moments in cases of disagreement was reinforced in the second survey wherein dermatologists switched far more often when they were incorrect (5/13, 39%; 95% CI 23.0%-55.7% in the first survey vs 6/12, 52%; 95% CI 39.8%-64.3% in the second survey).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Unclear"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Artificial Intelligence and Its Effect on Dermatologists' Accuracy in Dermoscopic Melanoma Image Classification: Web-Based Survey Study.",
        "abstract": "## BACKGROUND\nEarly detection of melanoma can be lifesaving but this remains a challenge. Recent diagnostic studies have revealed the superiority of artificial intelligence (AI) in classifying dermoscopic images of melanoma and nevi, concluding that these algorithms should assist a dermatologist's diagnoses.\n## OBJECTIVE\nThe aim of this study was to investigate whether AI support improves the accuracy and overall diagnostic performance of dermatologists in the dichotomous image-based discrimination between melanoma and nevus.\n## METHODS\nTwelve board-certified dermatologists were presented disjoint sets of 100 unique dermoscopic images of melanomas and nevi (total of 1200 unique images), and they had to classify the images based on personal experience alone (part I) and with the support of a trained convolutional neural network (CNN, part II). Additionally, dermatologists were asked to rate their confidence in their final decision for each image.\n## RESULTS\nWhile the mean specificity of the dermatologists based on personal experience alone remained almost unchanged (70.6% vs 72.4%; P=.54) with AI support, the mean sensitivity and mean accuracy increased significantly (59.4% vs 74.6%; P=.003 and 65.0% vs 73.6%; P=.002, respectively) with AI support. Out of the 10% (10/94; 95% CI 8.4%-11.8%) of cases where dermatologists were correct and AI was incorrect, dermatologists on average changed to the incorrect answer for 39% (4/10; 95% CI 23.2%-55.6%) of cases. When dermatologists were incorrect and AI was correct (25/94, 27%; 95% CI 24.0%-30.1%), dermatologists changed their answers to the correct answer for 46% (11/25; 95% CI 33.1%-58.4%) of cases. Additionally, the dermatologists' average confidence in their decisions increased when the CNN confirmed their decision and decreased when the CNN disagreed, even when the dermatologists were correct. Reported values are based on the mean of all participants. Whenever absolute values are shown, the denominator and numerator are approximations as every dermatologist ended up rating a varying number of images due to a quality control step.\n## CONCLUSIONS\nThe findings of our study show that AI support can improve the overall accuracy of the dermatologists in the dichotomous image-based discrimination between melanoma and nevus. This supports the argument for AI-based tools to aid clinicians in skin lesion classification and provides a rationale for studies of such classifiers in real-life settings, wherein clinicians can integrate additional information such as patient age and medical history into their decisions.\n",
        "methods": "### Methods\n\n##### Study Design\n\nThis study was conducted from January 10, 2019 (design of study) to September 27, 2019 and was inspired by the study design of Sinz et al [2]. Questionnaires were sent out on June 12, 2019 and were completed by August 12, 2019. Ethical approval was waived by the ethics committee of the University of Heidelberg, Mannheim Faculty of Medicine as all the images were open source and anonymous and all participating dermatologists automatically became part of the study group.\n\nThe set of images used for the evaluation consisted of 1200 unique dermoscopic images split into 12 nonoverlapping individual test sets containing 100 images each (50 melanomas and 50 nevi). Each individual set was randomly assigned to exactly one dermatologist who diagnosed his/her set twice\u2014first without AI support (part I) and afterwards with AI support (part II). Images for each part were presented to the participants in 2 separate surveys containing 50 images each (again 50:50 split) to investigate a possible learning effect. Thus, every participant completed 4 surveys (2 surveys for part I and 2 surveys for part II) where images seen in part I were identical to those seen in part II. Both surveys from part I were carried out before any survey from part II. The participants received detailed instructions by email in which the survey structure and the procedure were discussed and the lesion distribution was disclosed. For part II, participants were made aware of the classifier\u2019s performance, which was established beforehand on a separate validation set.\n\n##### Classifier Training\n\nImages were obtained from the ISIC archive [19], with a large fraction of images coming from the HAM10000 dataset [20]. The archive is publicly accessible and contains anonymous dermoscopic images from multiple sources and various camera systems. From the available pigmented lesions, only images showing a biopsy-verified melanoma (n=1633) or nevus (n=3311) were selected (total n=4944).\n\nAs each of the 12 individual test sets consisted of 2 subsets (with 50 images each), there were 24 test subsets. A designated training set was constructed for each test subset by removing its images from the total image data, resulting in 24 training sets, each containing 4894 images (ie, 4944 \u2013 50). The class distribution across each training set was unevenly distributed with 1608 melanoma images in contrast to 3286 nevus images. To counter the class imbalance, the set of melanoma images was duplicated. Online data augmentation was used to increase the diversity of the training data and to modify the duplicated images. Before balancing, an evenly balanced validation set was removed from each training set for calibration and validation purposes later. A detailed description of the classifier training is given in Multimedia Appendix 1.\n\n##### Classifier Integration\n\nFor AI support to be used in an effective manner, a confidence measure was displayed in addition to the binary decision (melanoma or nevus), which indicated the classifier\u2019s confidence in its decision. The measure was obtained by mapping the classifier\u2019s raw output probability for the predicted class into a percentage range from 0% to 100%.\n\nParticipants were made aware of the classifier\u2019s overall performance by establishing the overall performance of all 24 classifiers on their validation set and by taking the average, which resulted in mean sensitivity, mean specificity, and mean accuracy of 78.0%, 81.0%, and 80.0% respectively.\n\n##### Electronic Survey and Participants\n\nFifteen participants were invited by personal invitation via their university email accounts by the principal investigator (TJB) based on previous cooperation. Three had to be excluded because one reported to not have time to do the second survey properly, and 2 did not attempt or finish the survey in the proposed timeframe, leaving 12 dermatologists (all board-certified and coauthors of this publication) from 9 German university hospitals for analysis. The Google Forms closed web-based survey was not anonymous and each participant had to disclose his or her full name and additional metadata. For part I, the dermatologists were shown images of biopsy-verified skin lesions and asked 3 mandatory questions. First, a personal rating of the image quality had to be given (excellent, good, sufficient, poor, other image problems, no image visible). Second, the participant was asked for a diagnosis of the displayed lesion with the options being benign or malignant. Third, the participant was requested to quantify his or her confidence in the decision on a scale of 0 (=very uncertain) to 10 (=very certain). Part II showed identical images and questions and was additionally labeled with the dermatologist\u2019s previous response and the diagnosis of the CNN together with its confidence level. For details on the participants, see Multimedia Appendix 1.\n\n##### Performance and Statistical Analysis\n\nThe primary endpoint of this study was whether dermatologists\u2019 accuracy (overall proportion of the correct predictions among the total number of examined cases) would increase with AI support with the secondary endpoint investigating how sensitivity and specificity would change (tested for significance using Wilcoxon test). Additional analyses investigated whether dermatologists experienced a learning effect when working with AI support (tested for significance using Pearson\u2019s chi-squared test) and how they would respond to this form of computer aid. Statistical significance was considered at P<.05, corrected to 0.016 (Bonferroni correction for the primary and secondary endpoint) to account for multiple testing. The details on how survey answers were evaluated are listed in Multimedia Appendix 1. As a pair of classifiers was responsible for scoring the test set of one participant (one classifier for survey 1, the other for survey 2), the results of both classifiers were combined for analysis reasons so that every dermatologist had a respective counterpart. Values shown in the result sections are based on the mean of all the participants. If absolute values are shown, the denominator and numerator are approximated. In theory, every dermatologist rated the exact same number of images, but due to the quality control step, not every rating was counted. Therefore, each dermatologist ended up rating a varying number of images.",
        "results": "### Results\n\n##### Classification Results\n\nIn the first part of the study, the dermatologists had an overall mean sensitivity of 59.4% (95% CI 53.3%-65.5%), specificity of 70.6% (95% CI 62.3%-78.9%), and accuracy of 65.0% (95% CI 62.3%-67.6%, see Table 1).\n\nIn the second part of the study, where participants could integrate the results of the CNN-based classifier in their decision-making, their overall sensitivity increased significantly to 74.6% (95% CI 69.9%-79.3%; _P=.003). Their mean specificity also showed a positive trend (72.4%; 95% CI 66.2%-78.6%; P=.54), so that the overall accuracy also increased significantly to 73.6% (95% CI 70.9%-76.3%; P=.002).\n\nThe CNN on its own had an even higher sensitivity (84.7%; 95% CI 81.9%-87.6%), specificity (79.1%; 95% CI 74.8%-83.4%), and accuracy (81.9%; 95% CI 79.7%-84.2%) on the test set, which was comparable to its performance on the validation set. The overall performance of the dermatologists and the CNN are summarized in Table 1 while individual performances are captured in Figure 1 and Figure 2, which show an overview of the various metrics for every participant with and without AI support.\n\n**Table 1.** Overall mean performance of dermatologists without artificial intelligence (AI) support (-AI) compared to that of dermatologists with AI support (+AI) and AI on its own. Performance is split into 3 categories, that is, sensitivity, specificity, and accuracy.\n\nPerformance AI Dermatologist -AI Dermatologist +AI\n\nSensitivity (95% CI) 84.7% (81.9%-87.6%) 59.4% (53.3%-65.5%) 74.6% (69.9%-79.3%)\n\nSpecificity (95% CI) 79.1% (74.8%-83.4%) 70.6% (62.3%-78.9%) 72.4% (66.2%-78.6%)\n\nAccuracy (95% CI) 81.9% (79.7%-84.2%) 65.0% (62.3%-67.6%) 73.6% (70.9%-76.3%)\n\n**Figure 1.** Combined and individual dermatologists\u2019 performance without and with artificial intelligence (AI) support. Every dot represents a single participant. A line between 2 dots connects the participants' metric without AI support to the corresponding metric with AI support. Highlighted dots represent the participants combined. Boxes indicate 25th and 75th percentile while the horizontal line within shows the median (50th percentile). Whiskers indicate the data range (1.5*IQR) where points beyond are considered as outliers.\n\n**Figure 2.** Combined and individual dermatologists\u2019 diagnostic accuracy without and with artificial intelligence (AI) support. Diagnostic accuracy is measured using sensitivity and specificity. Arrows represent the change in the diagnostic accuracy from without AI support to with AI support. Highlighted arrows represent the participants combined. In addition, the black curve denotes the mean receiver operating characteristic curve of the classifier. ROC: receiver operating characteristic; CNN: convolutional neural network; AUC: area under the curve.\n\n##### Effects of the Use of the Classifier on the Decisions of the Dermatologists\n\nThe mean overall confidence of the dermatologists increased only marginally with the use of the classifier (65.0%; 95% CI 58.5%-71.6% without AI to 66.8%; 95% CI 60.2%-73.4% with AI). The confidence of the classifier in its diagnostic decisions was at 65.0% (95% CI 62.8%-67.4%). Upon detailed analysis, one can detect opposite effects on the dermatologists\u2019 confidence and switching behavior depending on agreement or disagreement with the classifier. On average, dermatologists and classifiers agreed in 63% (59/94; 95% CI 59.4%-66.1%) of cases. Among these, 8% (7/94) of cases were diagnosed incorrectly and 55% (52/94) of cases were classified correctly (see Table 2).\n\n**Table 2.** Distribution of the correct and incorrect predictions by classifier and dermatologists without artificial intelligence (AI) support and switching in response to AI support. Percentages displayed below show the amount of times a switch did or did not occur for dermatologists when answering part II of the survey.\n\nGroupings Proportion, n (%)[a] 95% CI\n\n**Both incorrect, n=94** 7 (8) 5.5%-10.4%\n\nDermatologist switched, n=7 0 (1) 0%-2.2%\n\nDermatologist stayed, n=7 7 (99) 97.8%-100%\n\n**AI correct, Dermatologist incorrect, n=94** 25 (27) 24.0%-30.1%\n\nDermatologist switched, n=25 11 (46) 33.1%-58.4%\n\nDermatologist stayed, n=25 14 (54) 41.6%-66.9%\n\n**AI incorrect, Dermatologist correct, n=94** 10 (10) 8.4%-11.8%\n\nDermatologist switched, n=10 4 (39) 23.2%-55.6%\n\nDermatologist stayed, n=10 6 (61) 44.4%-76.9%\n\n**Both correct, n=94** 52 (55) 52.4%-57.1%\n\nDermatologist switched, n=52 0 (0) 0%-0.5%\n\nDermatologist stayed, n=52 52 (100) 99.5%-100%\n\naAs the mean of all the participants was taken and every participant ended up rating a varying amount of images due to the quality control step, the reported absolute values are approximations.\n\nOf the 37% (35/94; 95% CI 33.9%-40.6%) of cases in which dermatologists and classifier came to different conclusions, 10% (10/94) of cases were diagnosed correctly by the dermatologists and 27% (25/94) of cases were diagnosed correctly by the classifier.\n\nIn cases of agreement, the mean confidence of the dermatologists increased substantially from the first part of the study to the second part of the study (from 67.0%; 95% CI 60.6%-73.5% to 79.1%; 95% CI 75.0%-83.0%, respectively). In cases of disagreement, it decreased from 61.7% (95% CI 54.6%-68.9%) in the first to 44.3% (95% CI 31.8%-56.7%) in the second part of the study. This was also reflected in the dermatologists\u2019 switching behavior. In cases of agreement, the dermatologists basically never altered their classifications (0/59, 0%; 95% CI 0%-0.8%). In contrast, in cases of disagreement, the dermatologists altered their decision in 43% (15/35; 95% CI 30.8%-56.3%) of those cases. The dermatologists altered their diagnosis less often when they had initially diagnosed the lesion correctly (subsequently changed answer in 4/10, 39% of those cases) than when they had initially diagnosed it incorrectly (subsequently changed answer in 11/25, 46% of those cases, see Table 2). Altogether, this resulted in the observed overall increase in the correct diagnoses by the dermatologists. Out of all the occurring switches, dermatologists showed a higher willingness to switch from benign to malignant (11/15, 72%; 95% CI 53.4%-90.9%) than from malignant to benign (4/15, 28%; 95% CI 9.1%-46.6%).\n\nThe mean confidence levels of the classifier were much higher when the correct conclusion was reached (71.7%; 95% CI 69.3%-74.1%) than when the diagnosis was incorrect (34.8%; 95% CI 30.9%-38.7%). Regarding the dermatologists, a similar trend was less striking for part I (67.3%; 95% CI 61.1%-73.5% correct vs 60.8%; 95% CI 53.3%-68.2% incorrect) but became more pronounced for part II (71.8%; 95% CI 66.4%-77.1% correct vs 52.3%; 95% CI 41.0%-63.5% incorrect). A more detailed breakdown of the dermatologists\u2019 and classifier\u2019s confidence levels is shown in Table 3. Finally, the dermatologists altered their diagnoses in divergent cases more often when the CNN\u2019s confidence levels were high than when they were low (mean CNN confidence levels 63.3%; 95% CI 56.4%-70.1% with subsequent switch vs 53.7%; 95% CI 46.7%-60.7% with no subsequent switch).\n\n**Table 3.** Confidence distribution of the classifier and dermatologists for part I and part II.\n\nGroupings Confidence (95% CI)\n\n**Both incorrect**\n\nAI[a] 35.8% (27.6%-44.1%)\n\nDermatologist part I 62.3% (54.6%-69.9%)\n\nDermatologist part II 69.8% (62.4%-77.3%)\n\n**AI correct, Dermatologist incorrect**\n\nAI 65.4% (60.7%-70.2%)\n\nDermatologist part I 60.8% (53.3%-68.3%)\n\nDermatologist part II 46.0% (33.6%-58.5%)\n\n**AI incorrect, Dermatologist correct**\n\nAI 32.9% (28.1%-37.6%)\n\nDermatologist part I 63.2% (56.3%-70.2%)\n\nDermatologist part II 47.9% (36.4%-59.4%)\n\n**Both correct**\n\nAI 74.5% (72.0%-76.9%)\n\nDermatologist part I 67.9% (61.5%-74.3%)\n\nDermatologist part II 80.3% (76.2%-84.3%)\n\naAI: artificial intelligence.\n\n##### Learning Effect\n\nThe dermatologists received their results immediately after completing the first survey of the second part of the study. Thus, they had the opportunity to implement the lessons learned in the first AI-supported survey during the completion of the second survey in part II. A subanalysis showed that there was no detectable change in the dermatologists\u2019 performance from the first to the second survey in part I of the study (see figure in Multimedia Appendix 2) with no significant difference between sensitivity (P=.50) and specificity (P=.76). In contrast, the dermatologists tended to perform better in the second survey of part II compared to the first survey, suggesting that they had learned how to better incorporate the CNN results into their diagnostic procedures, albeit with differences in sensitivity (P=.21) and specificity (P=.43) being still insignificant. In divergent cases, participants switched their diagnosis in 49% (8/17; 95% CI 36.2%-62.3%) of cases in survey 2, whereas they had only done so in 38% (7/18; 95% CI 21.1%-54.5%) of these cases for survey 1. The tendency to switch at the correct moments in cases of disagreement was reinforced in the second survey wherein dermatologists switched far more often when they were incorrect (5/13, 39%; 95% CI 23.0%-55.7% in the first survey vs 6/12, 52%; 95% CI 39.8%-64.3% in the second survey).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Artificial Intelligence and Its Effect on Dermatologists' Accuracy in Dermoscopic Melanoma Image Classification: Web-Based Survey Study.",
        "abstract": "## BACKGROUND\nEarly detection of melanoma can be lifesaving but this remains a challenge. Recent diagnostic studies have revealed the superiority of artificial intelligence (AI) in classifying dermoscopic images of melanoma and nevi, concluding that these algorithms should assist a dermatologist's diagnoses.\n## OBJECTIVE\nThe aim of this study was to investigate whether AI support improves the accuracy and overall diagnostic performance of dermatologists in the dichotomous image-based discrimination between melanoma and nevus.\n## METHODS\nTwelve board-certified dermatologists were presented disjoint sets of 100 unique dermoscopic images of melanomas and nevi (total of 1200 unique images), and they had to classify the images based on personal experience alone (part I) and with the support of a trained convolutional neural network (CNN, part II). Additionally, dermatologists were asked to rate their confidence in their final decision for each image.\n## RESULTS\nWhile the mean specificity of the dermatologists based on personal experience alone remained almost unchanged (70.6% vs 72.4%; P=.54) with AI support, the mean sensitivity and mean accuracy increased significantly (59.4% vs 74.6%; P=.003 and 65.0% vs 73.6%; P=.002, respectively) with AI support. Out of the 10% (10/94; 95% CI 8.4%-11.8%) of cases where dermatologists were correct and AI was incorrect, dermatologists on average changed to the incorrect answer for 39% (4/10; 95% CI 23.2%-55.6%) of cases. When dermatologists were incorrect and AI was correct (25/94, 27%; 95% CI 24.0%-30.1%), dermatologists changed their answers to the correct answer for 46% (11/25; 95% CI 33.1%-58.4%) of cases. Additionally, the dermatologists' average confidence in their decisions increased when the CNN confirmed their decision and decreased when the CNN disagreed, even when the dermatologists were correct. Reported values are based on the mean of all participants. Whenever absolute values are shown, the denominator and numerator are approximations as every dermatologist ended up rating a varying number of images due to a quality control step.\n## CONCLUSIONS\nThe findings of our study show that AI support can improve the overall accuracy of the dermatologists in the dichotomous image-based discrimination between melanoma and nevus. This supports the argument for AI-based tools to aid clinicians in skin lesion classification and provides a rationale for studies of such classifiers in real-life settings, wherein clinicians can integrate additional information such as patient age and medical history into their decisions.\n",
        "methods": "### Methods\n\n##### Study Design\n\nThis study was conducted from January 10, 2019 (design of study) to September 27, 2019 and was inspired by the study design of Sinz et al [2]. Questionnaires were sent out on June 12, 2019 and were completed by August 12, 2019. Ethical approval was waived by the ethics committee of the University of Heidelberg, Mannheim Faculty of Medicine as all the images were open source and anonymous and all participating dermatologists automatically became part of the study group.\n\nThe set of images used for the evaluation consisted of 1200 unique dermoscopic images split into 12 nonoverlapping individual test sets containing 100 images each (50 melanomas and 50 nevi). Each individual set was randomly assigned to exactly one dermatologist who diagnosed his/her set twice\u2014first without AI support (part I) and afterwards with AI support (part II). Images for each part were presented to the participants in 2 separate surveys containing 50 images each (again 50:50 split) to investigate a possible learning effect. Thus, every participant completed 4 surveys (2 surveys for part I and 2 surveys for part II) where images seen in part I were identical to those seen in part II. Both surveys from part I were carried out before any survey from part II. The participants received detailed instructions by email in which the survey structure and the procedure were discussed and the lesion distribution was disclosed. For part II, participants were made aware of the classifier\u2019s performance, which was established beforehand on a separate validation set.\n\n##### Classifier Training\n\nImages were obtained from the ISIC archive [19], with a large fraction of images coming from the HAM10000 dataset [20]. The archive is publicly accessible and contains anonymous dermoscopic images from multiple sources and various camera systems. From the available pigmented lesions, only images showing a biopsy-verified melanoma (n=1633) or nevus (n=3311) were selected (total n=4944).\n\nAs each of the 12 individual test sets consisted of 2 subsets (with 50 images each), there were 24 test subsets. A designated training set was constructed for each test subset by removing its images from the total image data, resulting in 24 training sets, each containing 4894 images (ie, 4944 \u2013 50). The class distribution across each training set was unevenly distributed with 1608 melanoma images in contrast to 3286 nevus images. To counter the class imbalance, the set of melanoma images was duplicated. Online data augmentation was used to increase the diversity of the training data and to modify the duplicated images. Before balancing, an evenly balanced validation set was removed from each training set for calibration and validation purposes later. A detailed description of the classifier training is given in Multimedia Appendix 1.\n\n##### Classifier Integration\n\nFor AI support to be used in an effective manner, a confidence measure was displayed in addition to the binary decision (melanoma or nevus), which indicated the classifier\u2019s confidence in its decision. The measure was obtained by mapping the classifier\u2019s raw output probability for the predicted class into a percentage range from 0% to 100%.\n\nParticipants were made aware of the classifier\u2019s overall performance by establishing the overall performance of all 24 classifiers on their validation set and by taking the average, which resulted in mean sensitivity, mean specificity, and mean accuracy of 78.0%, 81.0%, and 80.0% respectively.\n\n##### Electronic Survey and Participants\n\nFifteen participants were invited by personal invitation via their university email accounts by the principal investigator (TJB) based on previous cooperation. Three had to be excluded because one reported to not have time to do the second survey properly, and 2 did not attempt or finish the survey in the proposed timeframe, leaving 12 dermatologists (all board-certified and coauthors of this publication) from 9 German university hospitals for analysis. The Google Forms closed web-based survey was not anonymous and each participant had to disclose his or her full name and additional metadata. For part I, the dermatologists were shown images of biopsy-verified skin lesions and asked 3 mandatory questions. First, a personal rating of the image quality had to be given (excellent, good, sufficient, poor, other image problems, no image visible). Second, the participant was asked for a diagnosis of the displayed lesion with the options being benign or malignant. Third, the participant was requested to quantify his or her confidence in the decision on a scale of 0 (=very uncertain) to 10 (=very certain). Part II showed identical images and questions and was additionally labeled with the dermatologist\u2019s previous response and the diagnosis of the CNN together with its confidence level. For details on the participants, see Multimedia Appendix 1.\n\n##### Performance and Statistical Analysis\n\nThe primary endpoint of this study was whether dermatologists\u2019 accuracy (overall proportion of the correct predictions among the total number of examined cases) would increase with AI support with the secondary endpoint investigating how sensitivity and specificity would change (tested for significance using Wilcoxon test). Additional analyses investigated whether dermatologists experienced a learning effect when working with AI support (tested for significance using Pearson\u2019s chi-squared test) and how they would respond to this form of computer aid. Statistical significance was considered at P<.05, corrected to 0.016 (Bonferroni correction for the primary and secondary endpoint) to account for multiple testing. The details on how survey answers were evaluated are listed in Multimedia Appendix 1. As a pair of classifiers was responsible for scoring the test set of one participant (one classifier for survey 1, the other for survey 2), the results of both classifiers were combined for analysis reasons so that every dermatologist had a respective counterpart. Values shown in the result sections are based on the mean of all the participants. If absolute values are shown, the denominator and numerator are approximated. In theory, every dermatologist rated the exact same number of images, but due to the quality control step, not every rating was counted. Therefore, each dermatologist ended up rating a varying number of images.",
        "results": "### Results\n\n##### Classification Results\n\nIn the first part of the study, the dermatologists had an overall mean sensitivity of 59.4% (95% CI 53.3%-65.5%), specificity of 70.6% (95% CI 62.3%-78.9%), and accuracy of 65.0% (95% CI 62.3%-67.6%, see Table 1).\n\nIn the second part of the study, where participants could integrate the results of the CNN-based classifier in their decision-making, their overall sensitivity increased significantly to 74.6% (95% CI 69.9%-79.3%; _P=.003). Their mean specificity also showed a positive trend (72.4%; 95% CI 66.2%-78.6%; P=.54), so that the overall accuracy also increased significantly to 73.6% (95% CI 70.9%-76.3%; P=.002).\n\nThe CNN on its own had an even higher sensitivity (84.7%; 95% CI 81.9%-87.6%), specificity (79.1%; 95% CI 74.8%-83.4%), and accuracy (81.9%; 95% CI 79.7%-84.2%) on the test set, which was comparable to its performance on the validation set. The overall performance of the dermatologists and the CNN are summarized in Table 1 while individual performances are captured in Figure 1 and Figure 2, which show an overview of the various metrics for every participant with and without AI support.\n\n**Table 1.** Overall mean performance of dermatologists without artificial intelligence (AI) support (-AI) compared to that of dermatologists with AI support (+AI) and AI on its own. Performance is split into 3 categories, that is, sensitivity, specificity, and accuracy.\n\nPerformance AI Dermatologist -AI Dermatologist +AI\n\nSensitivity (95% CI) 84.7% (81.9%-87.6%) 59.4% (53.3%-65.5%) 74.6% (69.9%-79.3%)\n\nSpecificity (95% CI) 79.1% (74.8%-83.4%) 70.6% (62.3%-78.9%) 72.4% (66.2%-78.6%)\n\nAccuracy (95% CI) 81.9% (79.7%-84.2%) 65.0% (62.3%-67.6%) 73.6% (70.9%-76.3%)\n\n**Figure 1.** Combined and individual dermatologists\u2019 performance without and with artificial intelligence (AI) support. Every dot represents a single participant. A line between 2 dots connects the participants' metric without AI support to the corresponding metric with AI support. Highlighted dots represent the participants combined. Boxes indicate 25th and 75th percentile while the horizontal line within shows the median (50th percentile). Whiskers indicate the data range (1.5*IQR) where points beyond are considered as outliers.\n\n**Figure 2.** Combined and individual dermatologists\u2019 diagnostic accuracy without and with artificial intelligence (AI) support. Diagnostic accuracy is measured using sensitivity and specificity. Arrows represent the change in the diagnostic accuracy from without AI support to with AI support. Highlighted arrows represent the participants combined. In addition, the black curve denotes the mean receiver operating characteristic curve of the classifier. ROC: receiver operating characteristic; CNN: convolutional neural network; AUC: area under the curve.\n\n##### Effects of the Use of the Classifier on the Decisions of the Dermatologists\n\nThe mean overall confidence of the dermatologists increased only marginally with the use of the classifier (65.0%; 95% CI 58.5%-71.6% without AI to 66.8%; 95% CI 60.2%-73.4% with AI). The confidence of the classifier in its diagnostic decisions was at 65.0% (95% CI 62.8%-67.4%). Upon detailed analysis, one can detect opposite effects on the dermatologists\u2019 confidence and switching behavior depending on agreement or disagreement with the classifier. On average, dermatologists and classifiers agreed in 63% (59/94; 95% CI 59.4%-66.1%) of cases. Among these, 8% (7/94) of cases were diagnosed incorrectly and 55% (52/94) of cases were classified correctly (see Table 2).\n\n**Table 2.** Distribution of the correct and incorrect predictions by classifier and dermatologists without artificial intelligence (AI) support and switching in response to AI support. Percentages displayed below show the amount of times a switch did or did not occur for dermatologists when answering part II of the survey.\n\nGroupings Proportion, n (%)[a] 95% CI\n\n**Both incorrect, n=94** 7 (8) 5.5%-10.4%\n\nDermatologist switched, n=7 0 (1) 0%-2.2%\n\nDermatologist stayed, n=7 7 (99) 97.8%-100%\n\n**AI correct, Dermatologist incorrect, n=94** 25 (27) 24.0%-30.1%\n\nDermatologist switched, n=25 11 (46) 33.1%-58.4%\n\nDermatologist stayed, n=25 14 (54) 41.6%-66.9%\n\n**AI incorrect, Dermatologist correct, n=94** 10 (10) 8.4%-11.8%\n\nDermatologist switched, n=10 4 (39) 23.2%-55.6%\n\nDermatologist stayed, n=10 6 (61) 44.4%-76.9%\n\n**Both correct, n=94** 52 (55) 52.4%-57.1%\n\nDermatologist switched, n=52 0 (0) 0%-0.5%\n\nDermatologist stayed, n=52 52 (100) 99.5%-100%\n\naAs the mean of all the participants was taken and every participant ended up rating a varying amount of images due to the quality control step, the reported absolute values are approximations.\n\nOf the 37% (35/94; 95% CI 33.9%-40.6%) of cases in which dermatologists and classifier came to different conclusions, 10% (10/94) of cases were diagnosed correctly by the dermatologists and 27% (25/94) of cases were diagnosed correctly by the classifier.\n\nIn cases of agreement, the mean confidence of the dermatologists increased substantially from the first part of the study to the second part of the study (from 67.0%; 95% CI 60.6%-73.5% to 79.1%; 95% CI 75.0%-83.0%, respectively). In cases of disagreement, it decreased from 61.7% (95% CI 54.6%-68.9%) in the first to 44.3% (95% CI 31.8%-56.7%) in the second part of the study. This was also reflected in the dermatologists\u2019 switching behavior. In cases of agreement, the dermatologists basically never altered their classifications (0/59, 0%; 95% CI 0%-0.8%). In contrast, in cases of disagreement, the dermatologists altered their decision in 43% (15/35; 95% CI 30.8%-56.3%) of those cases. The dermatologists altered their diagnosis less often when they had initially diagnosed the lesion correctly (subsequently changed answer in 4/10, 39% of those cases) than when they had initially diagnosed it incorrectly (subsequently changed answer in 11/25, 46% of those cases, see Table 2). Altogether, this resulted in the observed overall increase in the correct diagnoses by the dermatologists. Out of all the occurring switches, dermatologists showed a higher willingness to switch from benign to malignant (11/15, 72%; 95% CI 53.4%-90.9%) than from malignant to benign (4/15, 28%; 95% CI 9.1%-46.6%).\n\nThe mean confidence levels of the classifier were much higher when the correct conclusion was reached (71.7%; 95% CI 69.3%-74.1%) than when the diagnosis was incorrect (34.8%; 95% CI 30.9%-38.7%). Regarding the dermatologists, a similar trend was less striking for part I (67.3%; 95% CI 61.1%-73.5% correct vs 60.8%; 95% CI 53.3%-68.2% incorrect) but became more pronounced for part II (71.8%; 95% CI 66.4%-77.1% correct vs 52.3%; 95% CI 41.0%-63.5% incorrect). A more detailed breakdown of the dermatologists\u2019 and classifier\u2019s confidence levels is shown in Table 3. Finally, the dermatologists altered their diagnoses in divergent cases more often when the CNN\u2019s confidence levels were high than when they were low (mean CNN confidence levels 63.3%; 95% CI 56.4%-70.1% with subsequent switch vs 53.7%; 95% CI 46.7%-60.7% with no subsequent switch).\n\n**Table 3.** Confidence distribution of the classifier and dermatologists for part I and part II.\n\nGroupings Confidence (95% CI)\n\n**Both incorrect**\n\nAI[a] 35.8% (27.6%-44.1%)\n\nDermatologist part I 62.3% (54.6%-69.9%)\n\nDermatologist part II 69.8% (62.4%-77.3%)\n\n**AI correct, Dermatologist incorrect**\n\nAI 65.4% (60.7%-70.2%)\n\nDermatologist part I 60.8% (53.3%-68.3%)\n\nDermatologist part II 46.0% (33.6%-58.5%)\n\n**AI incorrect, Dermatologist correct**\n\nAI 32.9% (28.1%-37.6%)\n\nDermatologist part I 63.2% (56.3%-70.2%)\n\nDermatologist part II 47.9% (36.4%-59.4%)\n\n**Both correct**\n\nAI 74.5% (72.0%-76.9%)\n\nDermatologist part I 67.9% (61.5%-74.3%)\n\nDermatologist part II 80.3% (76.2%-84.3%)\n\naAI: artificial intelligence.\n\n##### Learning Effect\n\nThe dermatologists received their results immediately after completing the first survey of the second part of the study. Thus, they had the opportunity to implement the lessons learned in the first AI-supported survey during the completion of the second survey in part II. A subanalysis showed that there was no detectable change in the dermatologists\u2019 performance from the first to the second survey in part I of the study (see figure in Multimedia Appendix 2) with no significant difference between sensitivity (P=.50) and specificity (P=.76). In contrast, the dermatologists tended to perform better in the second survey of part II compared to the first survey, suggesting that they had learned how to better incorporate the CNN results into their diagnostic procedures, albeit with differences in sensitivity (P=.21) and specificity (P=.43) being still insignificant. In divergent cases, participants switched their diagnosis in 49% (8/17; 95% CI 36.2%-62.3%) of cases in survey 2, whereas they had only done so in 38% (7/18; 95% CI 21.1%-54.5%) of these cases for survey 1. The tendency to switch at the correct moments in cases of disagreement was reinforced in the second survey wherein dermatologists switched far more often when they were incorrect (5/13, 39%; 95% CI 23.0%-55.7% in the first survey vs 6/12, 52%; 95% CI 39.8%-64.3% in the second survey).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Artificial Intelligence and Its Effect on Dermatologists' Accuracy in Dermoscopic Melanoma Image Classification: Web-Based Survey Study.",
        "abstract": "## BACKGROUND\nEarly detection of melanoma can be lifesaving but this remains a challenge. Recent diagnostic studies have revealed the superiority of artificial intelligence (AI) in classifying dermoscopic images of melanoma and nevi, concluding that these algorithms should assist a dermatologist's diagnoses.\n## OBJECTIVE\nThe aim of this study was to investigate whether AI support improves the accuracy and overall diagnostic performance of dermatologists in the dichotomous image-based discrimination between melanoma and nevus.\n## METHODS\nTwelve board-certified dermatologists were presented disjoint sets of 100 unique dermoscopic images of melanomas and nevi (total of 1200 unique images), and they had to classify the images based on personal experience alone (part I) and with the support of a trained convolutional neural network (CNN, part II). Additionally, dermatologists were asked to rate their confidence in their final decision for each image.\n## RESULTS\nWhile the mean specificity of the dermatologists based on personal experience alone remained almost unchanged (70.6% vs 72.4%; P=.54) with AI support, the mean sensitivity and mean accuracy increased significantly (59.4% vs 74.6%; P=.003 and 65.0% vs 73.6%; P=.002, respectively) with AI support. Out of the 10% (10/94; 95% CI 8.4%-11.8%) of cases where dermatologists were correct and AI was incorrect, dermatologists on average changed to the incorrect answer for 39% (4/10; 95% CI 23.2%-55.6%) of cases. When dermatologists were incorrect and AI was correct (25/94, 27%; 95% CI 24.0%-30.1%), dermatologists changed their answers to the correct answer for 46% (11/25; 95% CI 33.1%-58.4%) of cases. Additionally, the dermatologists' average confidence in their decisions increased when the CNN confirmed their decision and decreased when the CNN disagreed, even when the dermatologists were correct. Reported values are based on the mean of all participants. Whenever absolute values are shown, the denominator and numerator are approximations as every dermatologist ended up rating a varying number of images due to a quality control step.\n## CONCLUSIONS\nThe findings of our study show that AI support can improve the overall accuracy of the dermatologists in the dichotomous image-based discrimination between melanoma and nevus. This supports the argument for AI-based tools to aid clinicians in skin lesion classification and provides a rationale for studies of such classifiers in real-life settings, wherein clinicians can integrate additional information such as patient age and medical history into their decisions.\n",
        "methods": "### Methods\n\n##### Study Design\n\nThis study was conducted from January 10, 2019 (design of study) to September 27, 2019 and was inspired by the study design of Sinz et al [2]. Questionnaires were sent out on June 12, 2019 and were completed by August 12, 2019. Ethical approval was waived by the ethics committee of the University of Heidelberg, Mannheim Faculty of Medicine as all the images were open source and anonymous and all participating dermatologists automatically became part of the study group.\n\nThe set of images used for the evaluation consisted of 1200 unique dermoscopic images split into 12 nonoverlapping individual test sets containing 100 images each (50 melanomas and 50 nevi). Each individual set was randomly assigned to exactly one dermatologist who diagnosed his/her set twice\u2014first without AI support (part I) and afterwards with AI support (part II). Images for each part were presented to the participants in 2 separate surveys containing 50 images each (again 50:50 split) to investigate a possible learning effect. Thus, every participant completed 4 surveys (2 surveys for part I and 2 surveys for part II) where images seen in part I were identical to those seen in part II. Both surveys from part I were carried out before any survey from part II. The participants received detailed instructions by email in which the survey structure and the procedure were discussed and the lesion distribution was disclosed. For part II, participants were made aware of the classifier\u2019s performance, which was established beforehand on a separate validation set.\n\n##### Classifier Training\n\nImages were obtained from the ISIC archive [19], with a large fraction of images coming from the HAM10000 dataset [20]. The archive is publicly accessible and contains anonymous dermoscopic images from multiple sources and various camera systems. From the available pigmented lesions, only images showing a biopsy-verified melanoma (n=1633) or nevus (n=3311) were selected (total n=4944).\n\nAs each of the 12 individual test sets consisted of 2 subsets (with 50 images each), there were 24 test subsets. A designated training set was constructed for each test subset by removing its images from the total image data, resulting in 24 training sets, each containing 4894 images (ie, 4944 \u2013 50). The class distribution across each training set was unevenly distributed with 1608 melanoma images in contrast to 3286 nevus images. To counter the class imbalance, the set of melanoma images was duplicated. Online data augmentation was used to increase the diversity of the training data and to modify the duplicated images. Before balancing, an evenly balanced validation set was removed from each training set for calibration and validation purposes later. A detailed description of the classifier training is given in Multimedia Appendix 1.\n\n##### Classifier Integration\n\nFor AI support to be used in an effective manner, a confidence measure was displayed in addition to the binary decision (melanoma or nevus), which indicated the classifier\u2019s confidence in its decision. The measure was obtained by mapping the classifier\u2019s raw output probability for the predicted class into a percentage range from 0% to 100%.\n\nParticipants were made aware of the classifier\u2019s overall performance by establishing the overall performance of all 24 classifiers on their validation set and by taking the average, which resulted in mean sensitivity, mean specificity, and mean accuracy of 78.0%, 81.0%, and 80.0% respectively.\n\n##### Electronic Survey and Participants\n\nFifteen participants were invited by personal invitation via their university email accounts by the principal investigator (TJB) based on previous cooperation. Three had to be excluded because one reported to not have time to do the second survey properly, and 2 did not attempt or finish the survey in the proposed timeframe, leaving 12 dermatologists (all board-certified and coauthors of this publication) from 9 German university hospitals for analysis. The Google Forms closed web-based survey was not anonymous and each participant had to disclose his or her full name and additional metadata. For part I, the dermatologists were shown images of biopsy-verified skin lesions and asked 3 mandatory questions. First, a personal rating of the image quality had to be given (excellent, good, sufficient, poor, other image problems, no image visible). Second, the participant was asked for a diagnosis of the displayed lesion with the options being benign or malignant. Third, the participant was requested to quantify his or her confidence in the decision on a scale of 0 (=very uncertain) to 10 (=very certain). Part II showed identical images and questions and was additionally labeled with the dermatologist\u2019s previous response and the diagnosis of the CNN together with its confidence level. For details on the participants, see Multimedia Appendix 1.\n\n##### Performance and Statistical Analysis\n\nThe primary endpoint of this study was whether dermatologists\u2019 accuracy (overall proportion of the correct predictions among the total number of examined cases) would increase with AI support with the secondary endpoint investigating how sensitivity and specificity would change (tested for significance using Wilcoxon test). Additional analyses investigated whether dermatologists experienced a learning effect when working with AI support (tested for significance using Pearson\u2019s chi-squared test) and how they would respond to this form of computer aid. Statistical significance was considered at P<.05, corrected to 0.016 (Bonferroni correction for the primary and secondary endpoint) to account for multiple testing. The details on how survey answers were evaluated are listed in Multimedia Appendix 1. As a pair of classifiers was responsible for scoring the test set of one participant (one classifier for survey 1, the other for survey 2), the results of both classifiers were combined for analysis reasons so that every dermatologist had a respective counterpart. Values shown in the result sections are based on the mean of all the participants. If absolute values are shown, the denominator and numerator are approximated. In theory, every dermatologist rated the exact same number of images, but due to the quality control step, not every rating was counted. Therefore, each dermatologist ended up rating a varying number of images.",
        "results": "### Results\n\n##### Classification Results\n\nIn the first part of the study, the dermatologists had an overall mean sensitivity of 59.4% (95% CI 53.3%-65.5%), specificity of 70.6% (95% CI 62.3%-78.9%), and accuracy of 65.0% (95% CI 62.3%-67.6%, see Table 1).\n\nIn the second part of the study, where participants could integrate the results of the CNN-based classifier in their decision-making, their overall sensitivity increased significantly to 74.6% (95% CI 69.9%-79.3%; _P=.003). Their mean specificity also showed a positive trend (72.4%; 95% CI 66.2%-78.6%; P=.54), so that the overall accuracy also increased significantly to 73.6% (95% CI 70.9%-76.3%; P=.002).\n\nThe CNN on its own had an even higher sensitivity (84.7%; 95% CI 81.9%-87.6%), specificity (79.1%; 95% CI 74.8%-83.4%), and accuracy (81.9%; 95% CI 79.7%-84.2%) on the test set, which was comparable to its performance on the validation set. The overall performance of the dermatologists and the CNN are summarized in Table 1 while individual performances are captured in Figure 1 and Figure 2, which show an overview of the various metrics for every participant with and without AI support.\n\n**Table 1.** Overall mean performance of dermatologists without artificial intelligence (AI) support (-AI) compared to that of dermatologists with AI support (+AI) and AI on its own. Performance is split into 3 categories, that is, sensitivity, specificity, and accuracy.\n\nPerformance AI Dermatologist -AI Dermatologist +AI\n\nSensitivity (95% CI) 84.7% (81.9%-87.6%) 59.4% (53.3%-65.5%) 74.6% (69.9%-79.3%)\n\nSpecificity (95% CI) 79.1% (74.8%-83.4%) 70.6% (62.3%-78.9%) 72.4% (66.2%-78.6%)\n\nAccuracy (95% CI) 81.9% (79.7%-84.2%) 65.0% (62.3%-67.6%) 73.6% (70.9%-76.3%)\n\n**Figure 1.** Combined and individual dermatologists\u2019 performance without and with artificial intelligence (AI) support. Every dot represents a single participant. A line between 2 dots connects the participants' metric without AI support to the corresponding metric with AI support. Highlighted dots represent the participants combined. Boxes indicate 25th and 75th percentile while the horizontal line within shows the median (50th percentile). Whiskers indicate the data range (1.5*IQR) where points beyond are considered as outliers.\n\n**Figure 2.** Combined and individual dermatologists\u2019 diagnostic accuracy without and with artificial intelligence (AI) support. Diagnostic accuracy is measured using sensitivity and specificity. Arrows represent the change in the diagnostic accuracy from without AI support to with AI support. Highlighted arrows represent the participants combined. In addition, the black curve denotes the mean receiver operating characteristic curve of the classifier. ROC: receiver operating characteristic; CNN: convolutional neural network; AUC: area under the curve.\n\n##### Effects of the Use of the Classifier on the Decisions of the Dermatologists\n\nThe mean overall confidence of the dermatologists increased only marginally with the use of the classifier (65.0%; 95% CI 58.5%-71.6% without AI to 66.8%; 95% CI 60.2%-73.4% with AI). The confidence of the classifier in its diagnostic decisions was at 65.0% (95% CI 62.8%-67.4%). Upon detailed analysis, one can detect opposite effects on the dermatologists\u2019 confidence and switching behavior depending on agreement or disagreement with the classifier. On average, dermatologists and classifiers agreed in 63% (59/94; 95% CI 59.4%-66.1%) of cases. Among these, 8% (7/94) of cases were diagnosed incorrectly and 55% (52/94) of cases were classified correctly (see Table 2).\n\n**Table 2.** Distribution of the correct and incorrect predictions by classifier and dermatologists without artificial intelligence (AI) support and switching in response to AI support. Percentages displayed below show the amount of times a switch did or did not occur for dermatologists when answering part II of the survey.\n\nGroupings Proportion, n (%)[a] 95% CI\n\n**Both incorrect, n=94** 7 (8) 5.5%-10.4%\n\nDermatologist switched, n=7 0 (1) 0%-2.2%\n\nDermatologist stayed, n=7 7 (99) 97.8%-100%\n\n**AI correct, Dermatologist incorrect, n=94** 25 (27) 24.0%-30.1%\n\nDermatologist switched, n=25 11 (46) 33.1%-58.4%\n\nDermatologist stayed, n=25 14 (54) 41.6%-66.9%\n\n**AI incorrect, Dermatologist correct, n=94** 10 (10) 8.4%-11.8%\n\nDermatologist switched, n=10 4 (39) 23.2%-55.6%\n\nDermatologist stayed, n=10 6 (61) 44.4%-76.9%\n\n**Both correct, n=94** 52 (55) 52.4%-57.1%\n\nDermatologist switched, n=52 0 (0) 0%-0.5%\n\nDermatologist stayed, n=52 52 (100) 99.5%-100%\n\naAs the mean of all the participants was taken and every participant ended up rating a varying amount of images due to the quality control step, the reported absolute values are approximations.\n\nOf the 37% (35/94; 95% CI 33.9%-40.6%) of cases in which dermatologists and classifier came to different conclusions, 10% (10/94) of cases were diagnosed correctly by the dermatologists and 27% (25/94) of cases were diagnosed correctly by the classifier.\n\nIn cases of agreement, the mean confidence of the dermatologists increased substantially from the first part of the study to the second part of the study (from 67.0%; 95% CI 60.6%-73.5% to 79.1%; 95% CI 75.0%-83.0%, respectively). In cases of disagreement, it decreased from 61.7% (95% CI 54.6%-68.9%) in the first to 44.3% (95% CI 31.8%-56.7%) in the second part of the study. This was also reflected in the dermatologists\u2019 switching behavior. In cases of agreement, the dermatologists basically never altered their classifications (0/59, 0%; 95% CI 0%-0.8%). In contrast, in cases of disagreement, the dermatologists altered their decision in 43% (15/35; 95% CI 30.8%-56.3%) of those cases. The dermatologists altered their diagnosis less often when they had initially diagnosed the lesion correctly (subsequently changed answer in 4/10, 39% of those cases) than when they had initially diagnosed it incorrectly (subsequently changed answer in 11/25, 46% of those cases, see Table 2). Altogether, this resulted in the observed overall increase in the correct diagnoses by the dermatologists. Out of all the occurring switches, dermatologists showed a higher willingness to switch from benign to malignant (11/15, 72%; 95% CI 53.4%-90.9%) than from malignant to benign (4/15, 28%; 95% CI 9.1%-46.6%).\n\nThe mean confidence levels of the classifier were much higher when the correct conclusion was reached (71.7%; 95% CI 69.3%-74.1%) than when the diagnosis was incorrect (34.8%; 95% CI 30.9%-38.7%). Regarding the dermatologists, a similar trend was less striking for part I (67.3%; 95% CI 61.1%-73.5% correct vs 60.8%; 95% CI 53.3%-68.2% incorrect) but became more pronounced for part II (71.8%; 95% CI 66.4%-77.1% correct vs 52.3%; 95% CI 41.0%-63.5% incorrect). A more detailed breakdown of the dermatologists\u2019 and classifier\u2019s confidence levels is shown in Table 3. Finally, the dermatologists altered their diagnoses in divergent cases more often when the CNN\u2019s confidence levels were high than when they were low (mean CNN confidence levels 63.3%; 95% CI 56.4%-70.1% with subsequent switch vs 53.7%; 95% CI 46.7%-60.7% with no subsequent switch).\n\n**Table 3.** Confidence distribution of the classifier and dermatologists for part I and part II.\n\nGroupings Confidence (95% CI)\n\n**Both incorrect**\n\nAI[a] 35.8% (27.6%-44.1%)\n\nDermatologist part I 62.3% (54.6%-69.9%)\n\nDermatologist part II 69.8% (62.4%-77.3%)\n\n**AI correct, Dermatologist incorrect**\n\nAI 65.4% (60.7%-70.2%)\n\nDermatologist part I 60.8% (53.3%-68.3%)\n\nDermatologist part II 46.0% (33.6%-58.5%)\n\n**AI incorrect, Dermatologist correct**\n\nAI 32.9% (28.1%-37.6%)\n\nDermatologist part I 63.2% (56.3%-70.2%)\n\nDermatologist part II 47.9% (36.4%-59.4%)\n\n**Both correct**\n\nAI 74.5% (72.0%-76.9%)\n\nDermatologist part I 67.9% (61.5%-74.3%)\n\nDermatologist part II 80.3% (76.2%-84.3%)\n\naAI: artificial intelligence.\n\n##### Learning Effect\n\nThe dermatologists received their results immediately after completing the first survey of the second part of the study. Thus, they had the opportunity to implement the lessons learned in the first AI-supported survey during the completion of the second survey in part II. A subanalysis showed that there was no detectable change in the dermatologists\u2019 performance from the first to the second survey in part I of the study (see figure in Multimedia Appendix 2) with no significant difference between sensitivity (P=.50) and specificity (P=.76). In contrast, the dermatologists tended to perform better in the second survey of part II compared to the first survey, suggesting that they had learned how to better incorporate the CNN results into their diagnostic procedures, albeit with differences in sensitivity (P=.21) and specificity (P=.43) being still insignificant. In divergent cases, participants switched their diagnosis in 49% (8/17; 95% CI 36.2%-62.3%) of cases in survey 2, whereas they had only done so in 38% (7/18; 95% CI 21.1%-54.5%) of these cases for survey 1. The tendency to switch at the correct moments in cases of disagreement was reinforced in the second survey wherein dermatologists switched far more often when they were incorrect (5/13, 39%; 95% CI 23.0%-55.7% in the first survey vs 6/12, 52%; 95% CI 39.8%-64.3% in the second survey).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions.",
        "abstract": "This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Pigmented Skin Image Dataset_\n\nThis study used images from the anonymous and annotated HAM10000 dataset publicly available through the International Skin Imaging Collaboration (ISIC) archive [21]. All downloaded images were selected using a random generator from the set of available images in the ISIC archive. We stochastically split the master set of 10,015 dermoscopic images into training (n = 8313; 83%) and test (n = 1702; 17%) datasets that were completely disjoint. Images included a representative collection of all-important diagnostic categories across the seven different types of pigmented lesions as detailed in Tschandl et al. [21]. These included melanocytic nevus, vascular skin lesions (including cherry angiomas, angiokeratomas, pyogenic granulomas and hemorrhages), benign keratoses (including seborrheic keratoses, solar lentigo and lichen-planus-like kertoses), dermatofibroma, intraepithelial carcinoma (including actinic keratoses and Bowen\u2019s disease), basal cell carcinoma and melanoma. Examples of images of each lesion type are depicted in Figure 1. The final composition of each dataset is shown in Table 1.\n\n_2.2. Deep Neural Networks_\n\nWe evaluated eight different DNNs, each characterized by a specific architecture. VGG16 and VGG19 contain 16 and 19 convolutional layers, respectively, with very small receptive fields, five max-pooling layers of size for carrying out spatial pooling, followed by three fully connected layers, with the final layer as the soft-max layer [22]. Rectification nonlinearity (ReLu) activation is applied to all hidden layers. The model also uses dropout regularization in the fully connected layers. ResNet34 is a 34-layer residual network while ResNet50 and ResNet101 are 50- and 101-layers deep, respectively. The architecture of all these DNNs is similar to the one found in VGG consisting mostly of 3 \u00d7 3 filters, however, instead, shortcut connections are inserted resulting into a residual network. SEResNet50 architecture is based on ResNet. A squeeze-and-excitation block is applied at the end of each non-identity branch of residual block [23]. Differently, and instead of increasing its size by adding more or deeper layers, EfficientNetB5 scales up the network width, depth and resolution with a set of fixed scaling coefficients [24]. Finally, MobileNet uses depth-wise separable convolutions which significantly reduce the number of parameters when compared to a network based on standard convolutions and the same depth across the structure in the networks [25]. The framework is 54 layers deep. For each DNN, the initial weights of all layers of the network were set up after pretraining with ImageNet. To assess both the performance of the algorithm and the enhanced training techniques as accurately as possible, we retrained each DNN a total of five times (folds), and each training run consisted of 90 epochs. Training using the curated image patches took approximately 6 h to complete, 45k iterations on a 4 GeForce GTX 1080 GPU configuration. Training accuracy for the curated patches reached maximum accuracy (100%) at around epoch 32, whereas the pretrained model only began to converge around epoch 25. All these DNNs were trained and tested with two different image input size: 300 \u00d7 224 RGB and 600 \u00d7 450 RGB. The low-resolution images were obtained by cropping, distorting and linear resizing the original high-resolution images. DNNs were also trained and tested without or with the clinical features (sex, age and location of the lesion) associated to every image in the HAM10000 database.\n\n_2.3. Image Preprocessing_\n\nWhen a deep convolutional neural network overfits, it works extremely well on training data but poorly on data it has never seen before. This is especially important in the field of dermatology because of the variability that exists in the images that the neural network will be analyzing. Two steps were taken to reduce overfitting. First, a dropout layer was added and set to 0.5. This results in 50% of the neurons to be randomly turned off during the training process and therefore reduce the likelihood of overfitting. The second step taken to reduce overfitting was to use data augmentation. In data augmentation, the images are modified to account for some of the variability that exists in image taking. To account for the grid location, the size of the dermatological manifestation and the angle of the image, the training images fed into the model were altered using zoom (25% probability to be increased between \u22121 and 2%, rotated between \u22126\u00ba and 6\u00ba and vertical/horizontal transfer from \u22122 to 2%), rotation (25% probability of being randomly rotated 90\u00ba clockwise), transposing (15% probability of making a random axial symmetry in one of its diagonals) and horizontal and vertical flipping randomly (50% probability of making a horizontal or vertical \u201cmirror\u201d or both) or following optical parameters from the different types of phone cameras such as shear and brightness (30% probability of contrast modification between \u22128 and 8%) or optical (80% probability of being distorted with a range between \u22126 and 6%) or grid (75% probability at step 4 with a range of \u221228\u201328%) distortions. These modifications were applied to 66% of the input images. The model was run once without data augmentation with dropout and once with data augmentation including drop out in both instances.\n\n_2.4. Generation of Synthetic Pigmented Skin Lesion Images Using an infoGAN_\n\nGenerative adversarial networks (GANs) are a type of generative model that attempt to synthesize novel data that are indistinguishable from the training data [26]. They consist of two neural networks, locked in competition: a generator that captures the data distribution and creates synthesized data (e.g., an image), and a discriminator that estimates the probability that a sample came from the training data rather than from the generator. The two networks are sealed in a zero-sum game, where the success of one corresponds to the failure of the other. The training procedure for the generator is to maximize the probability of the discriminator making an error [27]. Thus, this framework is based on a value function that one model seeks to maximize and the other seeks to minimize. Since the two networks are differentiable, the system generates a gradient that can be used to steer both networks to the right direction. We trained an information maximizing generative adversarial network (infoGAN) composed of a generator and a discriminator, on all high resolution images of the HAM10000 dataset [20]. The infoGAN was adapted to the progressive architecture of the model by splitting the structured code into parts and feeding each part to the network by conditioning activation in the corresponding block of the generator (Supplementary Figure S1). To prevent detrimental competition between the discriminator and generator, and to achieve convergence in an efficient way, we followed the recommendations detailed in Chen et al. [20]. Briefly, the discriminator and generator were composed in eight progressive blocks with an input/output of spatial resolution of 4 \u00d7 4 in the initial step up to 512 \u00d7 512 in step 7. The Batch size has gone dynamically from 128 in step 0 to 2 in the last step. Both the generator and discriminator were optimized using Adam with an initial learning rate of 0.0075 and exponential decay of 0.99 and a Wasserstein function with gradient penalty. The training has progressed in phases of progressive increased resolution. More specifically, the model was capable of generating high resolution images with isolated semantic features controlled by a set of real valued variables. Color, age, sex, localization and type of lesion were the most important semantic features discovered during the training in a completely unsupervised fashion without human input. After training, the generator produced novel images, similar to those in the dataset.\n\n_2.5. Contests among General Practitioners_\n\nTo compare the accuracy of DNNs with non-dermatologist practitioners, we conducted two different challenges. The first aimed to establish the accuracy of general practitioners in classifying images from the HAM10000 dataset without time constraint. For this purpose, a group of 22 general practitioners from any given center in Buenos Aires (Argentina) were given access to 163 images of the different skin pigmented lesions through an anonymous website specifically created for this purpose. Physicians could enter and exit the website without limitation. Alongside the image, recorded factors such as the age, sex and localization of the lesion were shown. All physicians were asked to classify every image within the seven different diagnostics. No incentives were offered for participation. To ensure fair comparisons between the results determined by general practitioners and those determined by the DNNs, the same 162 images were run with the DNN framework which had the worse accuracy metrics in low-resolution and without aggregated clinical features of the eight DNNs tested. To determine if physicians could benefit from access to the algorithmic tool during the own classification task, a second evaluation was conducted. A group of 19 general practitioners that voluntarily accepted to participate in the study was first asked to assess 35 images in a simulated exercise with time constraints (physicians had 45 s to classify every image). In a second step, physicians had access to the predictions of the same algorithmic framework used during the first challenge, the same group having classified each image based on both their criteria and the algorithmic output. For this task, a new set of 35 images was shown with the same time constraints. In both contests, the ethics committee waived ethical approval owing to the use of anonymized dermatologic images obtained from the publicly available HAM10000 dataset.\n\n_2.6. Statistical Analysis_\n\nAfter the model had been trained, a test step was performed in which 1702 images of the seven dermatological manifestations were used as input and the results were statistically analyzed. A confusion matrix was constructed based on comparing the frameworks\u2019 prediction with each of the actual labels. All analyses were performed and programmed via a Jupyter notebook in Python. Sensitivity, specificity, geometric mean, accuracy and error rate were calculated for each dermatological manifestation [28]. Sensitivity or true positive rate (TPR) represented the positive and correctly classified samples to the total number of positive samples. The specificity or true negative rate (TNR) was estimated as the ratio of the correctly classified negative samples to the total number of negative samples. Geometric means were calculated by using the product of TPR and TNR. Accuracy was defined as the ratio between the correctly classified samples to the total number of samples [28]. We also calculated the error rate as the complement of accuracy. All these measures are suitable to evaluate the classification performance based on imbalanced data as found in the HAM10000 database. All metric results were calculated with respect to the class labels documented in the HAM10000 database archive. A summary diagram is depicted on Supplementary Figure S2.",
        "results": "**3. Results**\n\n_3.1. Classification Metrics across Eight Different DNNs_\n\nThe results of global accuracy and error rate for each DNN for the classification of seven pigmented skin lesions at low image resolution (300 \u00d7 224 RGB) are shown in Table 2. The average global accuracy for the eight DNNs reached 76.30% \u00b1 2.79, ranging from 74.05% (EfficientNetB5) to 82.47% (MobileNet). As shown in Supplementary Table S1, the TPR, TNR and geometric mean for each disease subtype varied according to the tested DNN. Almost all DNNs showed the highest TPR for melanocytic nevi classification when compared to the other pigmented lesions, with observed geometric mean values equal or lower than 0.65; interestingly, VGG16, VGG19 and MobileNet also showed high TPR for vascular lesion classification (geometric mean values lower than 0.65). In the case of melanoma and benign keratosis classification, all of them showed a TPR of approximately 0.5, with geometric mean values around 0.75. Likewise, basal cell carcinoma classification showed a TPR of approximately 0.5 and geometric mean values higher than 0.75 using ResNet50, ResNet101, SEResNet50 and EfficientNetB5. Similar results were observed for SEResNet5 and MobileNet and intraepithelial carcinoma classification.\n\n**Table 2. Classification metrics of HAM10000 images at two different resolutions and without aggregated clinical features using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 75.32 24.68 76.73 23.27  \nResNet50 74.56 25.44 75.97 24.03  \nResNet101 75.62 24.38 77.02 22.98  \nSEResNet50 77.82 22.22 79.13 20.87  \nVGG16 76.85 23.15 78.25 21.75  \nVGG19 74.21 25.79 75.62 24.38  \nEfficientNetB5 74.05 25.91 75.50 24.5  \nMobileNet 82.47 17.53 83.88 16.12  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network. 3 accuracy and error rates are expressed as percentages.\n\nUsing higher image resolution (600 \u00d7 450 RGB) improved the global accuracies across all DNNs (Table 2). The global accuracy average for the eight DNNs tested was 77.76 \u00b1 2.77, ranging from 75.50 (EfficientNetB5) to 83.88% (MobileNet). Although higher than the mean average observed with low-resolution images, the difference was not statistically significant (p = 0.07; Mann\u2013Whitney U test). The TPR, TNR and geometric mean values for each disease subtype are detailed in Supplementary Table S2. The highest TPR was observed for melanocytic nevi across all DNNs, and in the case of vascular lesions using VGG16 and MobileNet. On the contrary, the lowest TPR was observed for dermatofibrosis and intraepithelial carcinoma with ResNet34, ResNet50, SEResnet50 and EfficientNetB5. When compared with the low-resolution image parallel cases, the TPR, TNR and geometric mean values were quite similar for most of the disease subtypes. However, a drastic improvement of TPR was observed for dermatofibrosis using ResNet34 (from 0.23 to 0.34), ResNet50 (from 0.29 to 0.55), ResNet101 (from 0.26 to 0.42), VGG16 (from 0.17 to 0.4), VGG19 (from 0.26 to 0.40) and MobileNet (from 0.39 to 0.5). Similar TPR improvements were observed for the intraepithelial carcinoma classification using VGG19 (from 0.26 to 0.40). The average cascade framework runtime of the high-resolution classification model was 21.46+/\u22122.3 milliseconds per image, whereas the low-resolution model required only 18.6+/\u22121.22 milliseconds per image. Altogether, these results indicate that the tested DNNs can classify seven different types of pigmented skin lesions with accuracies higher than 0.7. No major differences in runtime were observed between the cascade framework input with low- or high resolution.\n\n_3.2. Classification Metrics of Different DNNs Aggregating Image and Clinical Features_\n\nWe then investigated if adding clinical features to the analysis could improve the classification accuracy of each DNN. As the HAM10000 dataset provides the sex, age and localization of the skin lesion associated to every image, we gathered these clinical features and aggregated them with the corresponding image to be used as input for each of the tested DNNs. Results are shown in Table 3. For low-resolution images, the addition of clinical data improved the global performance of all DNNs but MobileNet. The average global accuracy for the eight DNNs was 78.86% \u00b1 1.81, ranging from 75.73 (EfficientNetB5) to 81.24% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with low-resolution images without aggregated clinical features (p = 0.004; Mann\u2013Whitney U test). The highest increase was observed for VGG19 raising from 74.21% to 79.43 (5.22%). As shown in Supplementary Table S3, classification improvements were observed in almost all pigmented skin lesions for all DNNs. The highest marginal increases in accuracy were observed for dermatofibrosis where the TPR values for ResNet 50, SEResNet50, VGG16, EfficientNetB5 and MobileNet were raised by 15, 51, 22, 28 and 46%, respectively. Similarly, in the case of the intraepithelial carcinoma condition, TPR values increased 13 and 26% for SEResnet50 and VGG19, respectively, and for basal cell carcinoma classification, increased by 11, 14 and 12% for ResNet50, VGG19 and EfficientNetB5, respectively. For vascular lesions, the TPR values increased by 34, 14 and 29% for ResNet101, VGG19 and EfficientNetB5, respectively. Finally, adding clinical features to skin images also improved the melanoma condition accuracy for VGG16 (from 0.47 to 0.61) and VGG19 (from 0.44 to 0.58). Of note, a decreased TPR was observed for ResNet101 in the classification of dermatofibrosis (from 0.26 to 0.16) and intraepithelial carcinoma (from 0.37 to 0.29), and in melanoma for MobileNet (from 0.90 to 0.72).\n\n**Table 3. Classification metrics of HAM10000 low- and high-resolution images with aggregated clinical data using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 77.43 22.57 78.84 21.16  \nResNet50 79.31 20.69 80.72 19.28  \nResNet101 77.55 22.45 78.96 21.04  \nSEResNet50 80.01 19.99 80.72 19.28  \nVGG16 80.25 19.75 81.65 18.35  \nVGG19 79.43 20.57 79.02 20.98  \nEfficientNetB5 75.73 24.27 77.14 22.86  \nMobileNet 81.24 18.76 84.73 15.27  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network; 3 accuracy and error rates are expressed as percentages.\n\nFor high-resolution images, the average global accuracy also increased when clinical features were added to the model (Table 3). The average global accuracy was 80.22% \u00b1 2.30, ranging from 77.14 (EfficientNetB5) to 84.73% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with high-resolution images without aggregated clinical features (p = 0.02; Mann\u2013Whitney U test). This was particularly evident in the case of ResNet50 performance with an increase of 4.75%. The TPR, TNR and geometric mean values are shown in Supplementary Table S4. Major TPR improvements were observed for dermatofibrosis with ResNet34 (26%), ResNet101 (20%), VGG19 (17%) and EfficientNetB5 (28%). TPR increases were also observed for basal cell carcinoma with ResNet50 (12%) and EfficientNetB5 (13%), for vascular lesions with ResNet101 (37%) and EfficientNetB5 (28%), for melanoma with VGG16 (14%), VGG19 (32%) and MobileNet (16%) and intraepithelial carcinoma with VGG16 (24%). Interestingly, the TPR of ResNet50 and SEResNet50 were reduced for dermatofibrosis from 0.55 to 0.20 and from 0.44 to 0.29, respectively. When compared to the DNNs\u2019 performance with low-resolution images and clinical features, no major differences were observed (p = 0.21; Mann\u2013Whitney U test). Of interest, for dermatofibrosis classification, the TPR increased for ResNet34 (35%), ResNet101 (50%), VGG16 (50%) and VGG19 (24%); the exception was ResNet50 with a TPR reduction from 0.44 to 0.2. Altogether, these results indicate that the addition of information related to sex, age and localization of the lesion improves the accuracy of DNNs.\n\n_3.3. Performance across Synthetic Pigmented-Skin Lesion Images_\n\nWe applied an infoGAN to generate synthetic images from the HAM10000 dataset. As shown in Figure 2, the synthetic images seemed realistic and diverse. We then calculated the global accuracy for EfficientNetB5 for the classification of seven pigmented skin lesions. Of the 40 synthetic images analyzed, the network made a single error, so the certainty index was 97.5%; however, this value lacks any relevance since the GAN used for pseudo-labeling deliberately increases the definition limit of each class, inducing an improvement in the certainty of the classifier. Altogether, these data indicate that the synthetic samples are highly realistic and can be used as inputs to train DNNs on pigmented skin lesion classification.\n\n### A B C D\n\n E F G\n\n**Figure 2. Examples of synthetic images generated with the information maximizing generative adversarial network (infoGAN): (A) melanocytic nevus; (B) benign keratosis; (C) vascular lesion; (D) dermatofibroma; (E) intraepithelial carcinoma; (F) basal cell carcinoma; and (G) melanoma.**\n\n_3.4. Performance across General Practitioners with and without Assistance from DNNs Output_\n\nIn the first challenge, 22 general practitioners were asked to classify 162 images without any time constraint. The mean global accuracy and mean error rate were 27.74 and 72.26%, respectively. These results were similar to those previously published for non-dermatologists [9,10]. The best TPR (0.79) was obtained for the melanocytic nevi while the worse metrics were observed for vascular lesions (0.02), dermatofibrosis (0.01) and intraepithelial carcinoma (0.07) (Supplementary Table S5). As EfficientNetB5 was slightly less accurate than the other tested DNNs, we decided to use this framework as a comparator (see Table 1). In the same dataset, this DNN had a mean global accuracy of 78.40% and mean error rate 21.60% (Table 4). Compared to physicians, this was a relevant and significant difference as EfficientNetB5 showed a higher TPR in all disease subtypes (Supplementary Table S5).\n\n**Table 4. Classification metrics of non-dermatologists, general practitioners with or without use of the algorithmic platform and with time constraints.**\n\n**Condition** **Accuracy** **[1]** **Error Rate**\n\nEfficientNetB5 77.14 22.86  \nGPs [2] 17.29 82.71  \nGPs + AI 42.43 57.57  \n\n1 accuracy and error rate are expressed as percentages. 2 GPs: general practitioners; AI: artificial intelligence.\n\nIn the second challenge, 19 general practitioners were asked to classify 35 images with a time constraint of 45 s per image. The global accuracy for this classification was 17.29% (Table 4). In the same dataset, EfficientNetB5 achieved a global accuracy of 77.14%, significantly outperforming physicians. When general practitioners were given the opportunity to access the output of EfficientNetB5 per image, the global accuracy increased to 42.42%. This represented an increase of 25.13%. This result also indicated that, in some cases, physicians did not follow the recommendation of the DNN. Of note, the access of physicians to DNN prediction increased TPR for basal cell carcinoma (from 0.10 to 0.54) and melanoma (from 0.08 to 0.35) (Supplementary Table S6). In contrast, a small decrease in TPR for benign keratosis was observed (see Supplementary Table S6). Altogether, these results show that DNNs have the capability to classify seven different pigmented skin lesions with a level of competence higher to that of the general practitioners participating in these challenges. The access of DNN output by physicians improves their ability to classify pigmented skin lesions, particularly basal cell carcinoma and melanoma.",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Unclear"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions.",
        "abstract": "This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Pigmented Skin Image Dataset_\n\nThis study used images from the anonymous and annotated HAM10000 dataset publicly available through the International Skin Imaging Collaboration (ISIC) archive [21]. All downloaded images were selected using a random generator from the set of available images in the ISIC archive. We stochastically split the master set of 10,015 dermoscopic images into training (n = 8313; 83%) and test (n = 1702; 17%) datasets that were completely disjoint. Images included a representative collection of all-important diagnostic categories across the seven different types of pigmented lesions as detailed in Tschandl et al. [21]. These included melanocytic nevus, vascular skin lesions (including cherry angiomas, angiokeratomas, pyogenic granulomas and hemorrhages), benign keratoses (including seborrheic keratoses, solar lentigo and lichen-planus-like kertoses), dermatofibroma, intraepithelial carcinoma (including actinic keratoses and Bowen\u2019s disease), basal cell carcinoma and melanoma. Examples of images of each lesion type are depicted in Figure 1. The final composition of each dataset is shown in Table 1.\n\n_2.2. Deep Neural Networks_\n\nWe evaluated eight different DNNs, each characterized by a specific architecture. VGG16 and VGG19 contain 16 and 19 convolutional layers, respectively, with very small receptive fields, five max-pooling layers of size for carrying out spatial pooling, followed by three fully connected layers, with the final layer as the soft-max layer [22]. Rectification nonlinearity (ReLu) activation is applied to all hidden layers. The model also uses dropout regularization in the fully connected layers. ResNet34 is a 34-layer residual network while ResNet50 and ResNet101 are 50- and 101-layers deep, respectively. The architecture of all these DNNs is similar to the one found in VGG consisting mostly of 3 \u00d7 3 filters, however, instead, shortcut connections are inserted resulting into a residual network. SEResNet50 architecture is based on ResNet. A squeeze-and-excitation block is applied at the end of each non-identity branch of residual block [23]. Differently, and instead of increasing its size by adding more or deeper layers, EfficientNetB5 scales up the network width, depth and resolution with a set of fixed scaling coefficients [24]. Finally, MobileNet uses depth-wise separable convolutions which significantly reduce the number of parameters when compared to a network based on standard convolutions and the same depth across the structure in the networks [25]. The framework is 54 layers deep. For each DNN, the initial weights of all layers of the network were set up after pretraining with ImageNet. To assess both the performance of the algorithm and the enhanced training techniques as accurately as possible, we retrained each DNN a total of five times (folds), and each training run consisted of 90 epochs. Training using the curated image patches took approximately 6 h to complete, 45k iterations on a 4 GeForce GTX 1080 GPU configuration. Training accuracy for the curated patches reached maximum accuracy (100%) at around epoch 32, whereas the pretrained model only began to converge around epoch 25. All these DNNs were trained and tested with two different image input size: 300 \u00d7 224 RGB and 600 \u00d7 450 RGB. The low-resolution images were obtained by cropping, distorting and linear resizing the original high-resolution images. DNNs were also trained and tested without or with the clinical features (sex, age and location of the lesion) associated to every image in the HAM10000 database.\n\n_2.3. Image Preprocessing_\n\nWhen a deep convolutional neural network overfits, it works extremely well on training data but poorly on data it has never seen before. This is especially important in the field of dermatology because of the variability that exists in the images that the neural network will be analyzing. Two steps were taken to reduce overfitting. First, a dropout layer was added and set to 0.5. This results in 50% of the neurons to be randomly turned off during the training process and therefore reduce the likelihood of overfitting. The second step taken to reduce overfitting was to use data augmentation. In data augmentation, the images are modified to account for some of the variability that exists in image taking. To account for the grid location, the size of the dermatological manifestation and the angle of the image, the training images fed into the model were altered using zoom (25% probability to be increased between \u22121 and 2%, rotated between \u22126\u00ba and 6\u00ba and vertical/horizontal transfer from \u22122 to 2%), rotation (25% probability of being randomly rotated 90\u00ba clockwise), transposing (15% probability of making a random axial symmetry in one of its diagonals) and horizontal and vertical flipping randomly (50% probability of making a horizontal or vertical \u201cmirror\u201d or both) or following optical parameters from the different types of phone cameras such as shear and brightness (30% probability of contrast modification between \u22128 and 8%) or optical (80% probability of being distorted with a range between \u22126 and 6%) or grid (75% probability at step 4 with a range of \u221228\u201328%) distortions. These modifications were applied to 66% of the input images. The model was run once without data augmentation with dropout and once with data augmentation including drop out in both instances.\n\n_2.4. Generation of Synthetic Pigmented Skin Lesion Images Using an infoGAN_\n\nGenerative adversarial networks (GANs) are a type of generative model that attempt to synthesize novel data that are indistinguishable from the training data [26]. They consist of two neural networks, locked in competition: a generator that captures the data distribution and creates synthesized data (e.g., an image), and a discriminator that estimates the probability that a sample came from the training data rather than from the generator. The two networks are sealed in a zero-sum game, where the success of one corresponds to the failure of the other. The training procedure for the generator is to maximize the probability of the discriminator making an error [27]. Thus, this framework is based on a value function that one model seeks to maximize and the other seeks to minimize. Since the two networks are differentiable, the system generates a gradient that can be used to steer both networks to the right direction. We trained an information maximizing generative adversarial network (infoGAN) composed of a generator and a discriminator, on all high resolution images of the HAM10000 dataset [20]. The infoGAN was adapted to the progressive architecture of the model by splitting the structured code into parts and feeding each part to the network by conditioning activation in the corresponding block of the generator (Supplementary Figure S1). To prevent detrimental competition between the discriminator and generator, and to achieve convergence in an efficient way, we followed the recommendations detailed in Chen et al. [20]. Briefly, the discriminator and generator were composed in eight progressive blocks with an input/output of spatial resolution of 4 \u00d7 4 in the initial step up to 512 \u00d7 512 in step 7. The Batch size has gone dynamically from 128 in step 0 to 2 in the last step. Both the generator and discriminator were optimized using Adam with an initial learning rate of 0.0075 and exponential decay of 0.99 and a Wasserstein function with gradient penalty. The training has progressed in phases of progressive increased resolution. More specifically, the model was capable of generating high resolution images with isolated semantic features controlled by a set of real valued variables. Color, age, sex, localization and type of lesion were the most important semantic features discovered during the training in a completely unsupervised fashion without human input. After training, the generator produced novel images, similar to those in the dataset.\n\n_2.5. Contests among General Practitioners_\n\nTo compare the accuracy of DNNs with non-dermatologist practitioners, we conducted two different challenges. The first aimed to establish the accuracy of general practitioners in classifying images from the HAM10000 dataset without time constraint. For this purpose, a group of 22 general practitioners from any given center in Buenos Aires (Argentina) were given access to 163 images of the different skin pigmented lesions through an anonymous website specifically created for this purpose. Physicians could enter and exit the website without limitation. Alongside the image, recorded factors such as the age, sex and localization of the lesion were shown. All physicians were asked to classify every image within the seven different diagnostics. No incentives were offered for participation. To ensure fair comparisons between the results determined by general practitioners and those determined by the DNNs, the same 162 images were run with the DNN framework which had the worse accuracy metrics in low-resolution and without aggregated clinical features of the eight DNNs tested. To determine if physicians could benefit from access to the algorithmic tool during the own classification task, a second evaluation was conducted. A group of 19 general practitioners that voluntarily accepted to participate in the study was first asked to assess 35 images in a simulated exercise with time constraints (physicians had 45 s to classify every image). In a second step, physicians had access to the predictions of the same algorithmic framework used during the first challenge, the same group having classified each image based on both their criteria and the algorithmic output. For this task, a new set of 35 images was shown with the same time constraints. In both contests, the ethics committee waived ethical approval owing to the use of anonymized dermatologic images obtained from the publicly available HAM10000 dataset.\n\n_2.6. Statistical Analysis_\n\nAfter the model had been trained, a test step was performed in which 1702 images of the seven dermatological manifestations were used as input and the results were statistically analyzed. A confusion matrix was constructed based on comparing the frameworks\u2019 prediction with each of the actual labels. All analyses were performed and programmed via a Jupyter notebook in Python. Sensitivity, specificity, geometric mean, accuracy and error rate were calculated for each dermatological manifestation [28]. Sensitivity or true positive rate (TPR) represented the positive and correctly classified samples to the total number of positive samples. The specificity or true negative rate (TNR) was estimated as the ratio of the correctly classified negative samples to the total number of negative samples. Geometric means were calculated by using the product of TPR and TNR. Accuracy was defined as the ratio between the correctly classified samples to the total number of samples [28]. We also calculated the error rate as the complement of accuracy. All these measures are suitable to evaluate the classification performance based on imbalanced data as found in the HAM10000 database. All metric results were calculated with respect to the class labels documented in the HAM10000 database archive. A summary diagram is depicted on Supplementary Figure S2.",
        "results": "**3. Results**\n\n_3.1. Classification Metrics across Eight Different DNNs_\n\nThe results of global accuracy and error rate for each DNN for the classification of seven pigmented skin lesions at low image resolution (300 \u00d7 224 RGB) are shown in Table 2. The average global accuracy for the eight DNNs reached 76.30% \u00b1 2.79, ranging from 74.05% (EfficientNetB5) to 82.47% (MobileNet). As shown in Supplementary Table S1, the TPR, TNR and geometric mean for each disease subtype varied according to the tested DNN. Almost all DNNs showed the highest TPR for melanocytic nevi classification when compared to the other pigmented lesions, with observed geometric mean values equal or lower than 0.65; interestingly, VGG16, VGG19 and MobileNet also showed high TPR for vascular lesion classification (geometric mean values lower than 0.65). In the case of melanoma and benign keratosis classification, all of them showed a TPR of approximately 0.5, with geometric mean values around 0.75. Likewise, basal cell carcinoma classification showed a TPR of approximately 0.5 and geometric mean values higher than 0.75 using ResNet50, ResNet101, SEResNet50 and EfficientNetB5. Similar results were observed for SEResNet5 and MobileNet and intraepithelial carcinoma classification.\n\n**Table 2. Classification metrics of HAM10000 images at two different resolutions and without aggregated clinical features using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 75.32 24.68 76.73 23.27  \nResNet50 74.56 25.44 75.97 24.03  \nResNet101 75.62 24.38 77.02 22.98  \nSEResNet50 77.82 22.22 79.13 20.87  \nVGG16 76.85 23.15 78.25 21.75  \nVGG19 74.21 25.79 75.62 24.38  \nEfficientNetB5 74.05 25.91 75.50 24.5  \nMobileNet 82.47 17.53 83.88 16.12  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network. 3 accuracy and error rates are expressed as percentages.\n\nUsing higher image resolution (600 \u00d7 450 RGB) improved the global accuracies across all DNNs (Table 2). The global accuracy average for the eight DNNs tested was 77.76 \u00b1 2.77, ranging from 75.50 (EfficientNetB5) to 83.88% (MobileNet). Although higher than the mean average observed with low-resolution images, the difference was not statistically significant (p = 0.07; Mann\u2013Whitney U test). The TPR, TNR and geometric mean values for each disease subtype are detailed in Supplementary Table S2. The highest TPR was observed for melanocytic nevi across all DNNs, and in the case of vascular lesions using VGG16 and MobileNet. On the contrary, the lowest TPR was observed for dermatofibrosis and intraepithelial carcinoma with ResNet34, ResNet50, SEResnet50 and EfficientNetB5. When compared with the low-resolution image parallel cases, the TPR, TNR and geometric mean values were quite similar for most of the disease subtypes. However, a drastic improvement of TPR was observed for dermatofibrosis using ResNet34 (from 0.23 to 0.34), ResNet50 (from 0.29 to 0.55), ResNet101 (from 0.26 to 0.42), VGG16 (from 0.17 to 0.4), VGG19 (from 0.26 to 0.40) and MobileNet (from 0.39 to 0.5). Similar TPR improvements were observed for the intraepithelial carcinoma classification using VGG19 (from 0.26 to 0.40). The average cascade framework runtime of the high-resolution classification model was 21.46+/\u22122.3 milliseconds per image, whereas the low-resolution model required only 18.6+/\u22121.22 milliseconds per image. Altogether, these results indicate that the tested DNNs can classify seven different types of pigmented skin lesions with accuracies higher than 0.7. No major differences in runtime were observed between the cascade framework input with low- or high resolution.\n\n_3.2. Classification Metrics of Different DNNs Aggregating Image and Clinical Features_\n\nWe then investigated if adding clinical features to the analysis could improve the classification accuracy of each DNN. As the HAM10000 dataset provides the sex, age and localization of the skin lesion associated to every image, we gathered these clinical features and aggregated them with the corresponding image to be used as input for each of the tested DNNs. Results are shown in Table 3. For low-resolution images, the addition of clinical data improved the global performance of all DNNs but MobileNet. The average global accuracy for the eight DNNs was 78.86% \u00b1 1.81, ranging from 75.73 (EfficientNetB5) to 81.24% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with low-resolution images without aggregated clinical features (p = 0.004; Mann\u2013Whitney U test). The highest increase was observed for VGG19 raising from 74.21% to 79.43 (5.22%). As shown in Supplementary Table S3, classification improvements were observed in almost all pigmented skin lesions for all DNNs. The highest marginal increases in accuracy were observed for dermatofibrosis where the TPR values for ResNet 50, SEResNet50, VGG16, EfficientNetB5 and MobileNet were raised by 15, 51, 22, 28 and 46%, respectively. Similarly, in the case of the intraepithelial carcinoma condition, TPR values increased 13 and 26% for SEResnet50 and VGG19, respectively, and for basal cell carcinoma classification, increased by 11, 14 and 12% for ResNet50, VGG19 and EfficientNetB5, respectively. For vascular lesions, the TPR values increased by 34, 14 and 29% for ResNet101, VGG19 and EfficientNetB5, respectively. Finally, adding clinical features to skin images also improved the melanoma condition accuracy for VGG16 (from 0.47 to 0.61) and VGG19 (from 0.44 to 0.58). Of note, a decreased TPR was observed for ResNet101 in the classification of dermatofibrosis (from 0.26 to 0.16) and intraepithelial carcinoma (from 0.37 to 0.29), and in melanoma for MobileNet (from 0.90 to 0.72).\n\n**Table 3. Classification metrics of HAM10000 low- and high-resolution images with aggregated clinical data using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 77.43 22.57 78.84 21.16  \nResNet50 79.31 20.69 80.72 19.28  \nResNet101 77.55 22.45 78.96 21.04  \nSEResNet50 80.01 19.99 80.72 19.28  \nVGG16 80.25 19.75 81.65 18.35  \nVGG19 79.43 20.57 79.02 20.98  \nEfficientNetB5 75.73 24.27 77.14 22.86  \nMobileNet 81.24 18.76 84.73 15.27  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network; 3 accuracy and error rates are expressed as percentages.\n\nFor high-resolution images, the average global accuracy also increased when clinical features were added to the model (Table 3). The average global accuracy was 80.22% \u00b1 2.30, ranging from 77.14 (EfficientNetB5) to 84.73% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with high-resolution images without aggregated clinical features (p = 0.02; Mann\u2013Whitney U test). This was particularly evident in the case of ResNet50 performance with an increase of 4.75%. The TPR, TNR and geometric mean values are shown in Supplementary Table S4. Major TPR improvements were observed for dermatofibrosis with ResNet34 (26%), ResNet101 (20%), VGG19 (17%) and EfficientNetB5 (28%). TPR increases were also observed for basal cell carcinoma with ResNet50 (12%) and EfficientNetB5 (13%), for vascular lesions with ResNet101 (37%) and EfficientNetB5 (28%), for melanoma with VGG16 (14%), VGG19 (32%) and MobileNet (16%) and intraepithelial carcinoma with VGG16 (24%). Interestingly, the TPR of ResNet50 and SEResNet50 were reduced for dermatofibrosis from 0.55 to 0.20 and from 0.44 to 0.29, respectively. When compared to the DNNs\u2019 performance with low-resolution images and clinical features, no major differences were observed (p = 0.21; Mann\u2013Whitney U test). Of interest, for dermatofibrosis classification, the TPR increased for ResNet34 (35%), ResNet101 (50%), VGG16 (50%) and VGG19 (24%); the exception was ResNet50 with a TPR reduction from 0.44 to 0.2. Altogether, these results indicate that the addition of information related to sex, age and localization of the lesion improves the accuracy of DNNs.\n\n_3.3. Performance across Synthetic Pigmented-Skin Lesion Images_\n\nWe applied an infoGAN to generate synthetic images from the HAM10000 dataset. As shown in Figure 2, the synthetic images seemed realistic and diverse. We then calculated the global accuracy for EfficientNetB5 for the classification of seven pigmented skin lesions. Of the 40 synthetic images analyzed, the network made a single error, so the certainty index was 97.5%; however, this value lacks any relevance since the GAN used for pseudo-labeling deliberately increases the definition limit of each class, inducing an improvement in the certainty of the classifier. Altogether, these data indicate that the synthetic samples are highly realistic and can be used as inputs to train DNNs on pigmented skin lesion classification.\n\n### A B C D\n\n E F G\n\n**Figure 2. Examples of synthetic images generated with the information maximizing generative adversarial network (infoGAN): (A) melanocytic nevus; (B) benign keratosis; (C) vascular lesion; (D) dermatofibroma; (E) intraepithelial carcinoma; (F) basal cell carcinoma; and (G) melanoma.**\n\n_3.4. Performance across General Practitioners with and without Assistance from DNNs Output_\n\nIn the first challenge, 22 general practitioners were asked to classify 162 images without any time constraint. The mean global accuracy and mean error rate were 27.74 and 72.26%, respectively. These results were similar to those previously published for non-dermatologists [9,10]. The best TPR (0.79) was obtained for the melanocytic nevi while the worse metrics were observed for vascular lesions (0.02), dermatofibrosis (0.01) and intraepithelial carcinoma (0.07) (Supplementary Table S5). As EfficientNetB5 was slightly less accurate than the other tested DNNs, we decided to use this framework as a comparator (see Table 1). In the same dataset, this DNN had a mean global accuracy of 78.40% and mean error rate 21.60% (Table 4). Compared to physicians, this was a relevant and significant difference as EfficientNetB5 showed a higher TPR in all disease subtypes (Supplementary Table S5).\n\n**Table 4. Classification metrics of non-dermatologists, general practitioners with or without use of the algorithmic platform and with time constraints.**\n\n**Condition** **Accuracy** **[1]** **Error Rate**\n\nEfficientNetB5 77.14 22.86  \nGPs [2] 17.29 82.71  \nGPs + AI 42.43 57.57  \n\n1 accuracy and error rate are expressed as percentages. 2 GPs: general practitioners; AI: artificial intelligence.\n\nIn the second challenge, 19 general practitioners were asked to classify 35 images with a time constraint of 45 s per image. The global accuracy for this classification was 17.29% (Table 4). In the same dataset, EfficientNetB5 achieved a global accuracy of 77.14%, significantly outperforming physicians. When general practitioners were given the opportunity to access the output of EfficientNetB5 per image, the global accuracy increased to 42.42%. This represented an increase of 25.13%. This result also indicated that, in some cases, physicians did not follow the recommendation of the DNN. Of note, the access of physicians to DNN prediction increased TPR for basal cell carcinoma (from 0.10 to 0.54) and melanoma (from 0.08 to 0.35) (Supplementary Table S6). In contrast, a small decrease in TPR for benign keratosis was observed (see Supplementary Table S6). Altogether, these results show that DNNs have the capability to classify seven different pigmented skin lesions with a level of competence higher to that of the general practitioners participating in these challenges. The access of DNN output by physicians improves their ability to classify pigmented skin lesions, particularly basal cell carcinoma and melanoma.",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions.",
        "abstract": "This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Pigmented Skin Image Dataset_\n\nThis study used images from the anonymous and annotated HAM10000 dataset publicly available through the International Skin Imaging Collaboration (ISIC) archive [21]. All downloaded images were selected using a random generator from the set of available images in the ISIC archive. We stochastically split the master set of 10,015 dermoscopic images into training (n = 8313; 83%) and test (n = 1702; 17%) datasets that were completely disjoint. Images included a representative collection of all-important diagnostic categories across the seven different types of pigmented lesions as detailed in Tschandl et al. [21]. These included melanocytic nevus, vascular skin lesions (including cherry angiomas, angiokeratomas, pyogenic granulomas and hemorrhages), benign keratoses (including seborrheic keratoses, solar lentigo and lichen-planus-like kertoses), dermatofibroma, intraepithelial carcinoma (including actinic keratoses and Bowen\u2019s disease), basal cell carcinoma and melanoma. Examples of images of each lesion type are depicted in Figure 1. The final composition of each dataset is shown in Table 1.\n\n_2.2. Deep Neural Networks_\n\nWe evaluated eight different DNNs, each characterized by a specific architecture. VGG16 and VGG19 contain 16 and 19 convolutional layers, respectively, with very small receptive fields, five max-pooling layers of size for carrying out spatial pooling, followed by three fully connected layers, with the final layer as the soft-max layer [22]. Rectification nonlinearity (ReLu) activation is applied to all hidden layers. The model also uses dropout regularization in the fully connected layers. ResNet34 is a 34-layer residual network while ResNet50 and ResNet101 are 50- and 101-layers deep, respectively. The architecture of all these DNNs is similar to the one found in VGG consisting mostly of 3 \u00d7 3 filters, however, instead, shortcut connections are inserted resulting into a residual network. SEResNet50 architecture is based on ResNet. A squeeze-and-excitation block is applied at the end of each non-identity branch of residual block [23]. Differently, and instead of increasing its size by adding more or deeper layers, EfficientNetB5 scales up the network width, depth and resolution with a set of fixed scaling coefficients [24]. Finally, MobileNet uses depth-wise separable convolutions which significantly reduce the number of parameters when compared to a network based on standard convolutions and the same depth across the structure in the networks [25]. The framework is 54 layers deep. For each DNN, the initial weights of all layers of the network were set up after pretraining with ImageNet. To assess both the performance of the algorithm and the enhanced training techniques as accurately as possible, we retrained each DNN a total of five times (folds), and each training run consisted of 90 epochs. Training using the curated image patches took approximately 6 h to complete, 45k iterations on a 4 GeForce GTX 1080 GPU configuration. Training accuracy for the curated patches reached maximum accuracy (100%) at around epoch 32, whereas the pretrained model only began to converge around epoch 25. All these DNNs were trained and tested with two different image input size: 300 \u00d7 224 RGB and 600 \u00d7 450 RGB. The low-resolution images were obtained by cropping, distorting and linear resizing the original high-resolution images. DNNs were also trained and tested without or with the clinical features (sex, age and location of the lesion) associated to every image in the HAM10000 database.\n\n_2.3. Image Preprocessing_\n\nWhen a deep convolutional neural network overfits, it works extremely well on training data but poorly on data it has never seen before. This is especially important in the field of dermatology because of the variability that exists in the images that the neural network will be analyzing. Two steps were taken to reduce overfitting. First, a dropout layer was added and set to 0.5. This results in 50% of the neurons to be randomly turned off during the training process and therefore reduce the likelihood of overfitting. The second step taken to reduce overfitting was to use data augmentation. In data augmentation, the images are modified to account for some of the variability that exists in image taking. To account for the grid location, the size of the dermatological manifestation and the angle of the image, the training images fed into the model were altered using zoom (25% probability to be increased between \u22121 and 2%, rotated between \u22126\u00ba and 6\u00ba and vertical/horizontal transfer from \u22122 to 2%), rotation (25% probability of being randomly rotated 90\u00ba clockwise), transposing (15% probability of making a random axial symmetry in one of its diagonals) and horizontal and vertical flipping randomly (50% probability of making a horizontal or vertical \u201cmirror\u201d or both) or following optical parameters from the different types of phone cameras such as shear and brightness (30% probability of contrast modification between \u22128 and 8%) or optical (80% probability of being distorted with a range between \u22126 and 6%) or grid (75% probability at step 4 with a range of \u221228\u201328%) distortions. These modifications were applied to 66% of the input images. The model was run once without data augmentation with dropout and once with data augmentation including drop out in both instances.\n\n_2.4. Generation of Synthetic Pigmented Skin Lesion Images Using an infoGAN_\n\nGenerative adversarial networks (GANs) are a type of generative model that attempt to synthesize novel data that are indistinguishable from the training data [26]. They consist of two neural networks, locked in competition: a generator that captures the data distribution and creates synthesized data (e.g., an image), and a discriminator that estimates the probability that a sample came from the training data rather than from the generator. The two networks are sealed in a zero-sum game, where the success of one corresponds to the failure of the other. The training procedure for the generator is to maximize the probability of the discriminator making an error [27]. Thus, this framework is based on a value function that one model seeks to maximize and the other seeks to minimize. Since the two networks are differentiable, the system generates a gradient that can be used to steer both networks to the right direction. We trained an information maximizing generative adversarial network (infoGAN) composed of a generator and a discriminator, on all high resolution images of the HAM10000 dataset [20]. The infoGAN was adapted to the progressive architecture of the model by splitting the structured code into parts and feeding each part to the network by conditioning activation in the corresponding block of the generator (Supplementary Figure S1). To prevent detrimental competition between the discriminator and generator, and to achieve convergence in an efficient way, we followed the recommendations detailed in Chen et al. [20]. Briefly, the discriminator and generator were composed in eight progressive blocks with an input/output of spatial resolution of 4 \u00d7 4 in the initial step up to 512 \u00d7 512 in step 7. The Batch size has gone dynamically from 128 in step 0 to 2 in the last step. Both the generator and discriminator were optimized using Adam with an initial learning rate of 0.0075 and exponential decay of 0.99 and a Wasserstein function with gradient penalty. The training has progressed in phases of progressive increased resolution. More specifically, the model was capable of generating high resolution images with isolated semantic features controlled by a set of real valued variables. Color, age, sex, localization and type of lesion were the most important semantic features discovered during the training in a completely unsupervised fashion without human input. After training, the generator produced novel images, similar to those in the dataset.\n\n_2.5. Contests among General Practitioners_\n\nTo compare the accuracy of DNNs with non-dermatologist practitioners, we conducted two different challenges. The first aimed to establish the accuracy of general practitioners in classifying images from the HAM10000 dataset without time constraint. For this purpose, a group of 22 general practitioners from any given center in Buenos Aires (Argentina) were given access to 163 images of the different skin pigmented lesions through an anonymous website specifically created for this purpose. Physicians could enter and exit the website without limitation. Alongside the image, recorded factors such as the age, sex and localization of the lesion were shown. All physicians were asked to classify every image within the seven different diagnostics. No incentives were offered for participation. To ensure fair comparisons between the results determined by general practitioners and those determined by the DNNs, the same 162 images were run with the DNN framework which had the worse accuracy metrics in low-resolution and without aggregated clinical features of the eight DNNs tested. To determine if physicians could benefit from access to the algorithmic tool during the own classification task, a second evaluation was conducted. A group of 19 general practitioners that voluntarily accepted to participate in the study was first asked to assess 35 images in a simulated exercise with time constraints (physicians had 45 s to classify every image). In a second step, physicians had access to the predictions of the same algorithmic framework used during the first challenge, the same group having classified each image based on both their criteria and the algorithmic output. For this task, a new set of 35 images was shown with the same time constraints. In both contests, the ethics committee waived ethical approval owing to the use of anonymized dermatologic images obtained from the publicly available HAM10000 dataset.\n\n_2.6. Statistical Analysis_\n\nAfter the model had been trained, a test step was performed in which 1702 images of the seven dermatological manifestations were used as input and the results were statistically analyzed. A confusion matrix was constructed based on comparing the frameworks\u2019 prediction with each of the actual labels. All analyses were performed and programmed via a Jupyter notebook in Python. Sensitivity, specificity, geometric mean, accuracy and error rate were calculated for each dermatological manifestation [28]. Sensitivity or true positive rate (TPR) represented the positive and correctly classified samples to the total number of positive samples. The specificity or true negative rate (TNR) was estimated as the ratio of the correctly classified negative samples to the total number of negative samples. Geometric means were calculated by using the product of TPR and TNR. Accuracy was defined as the ratio between the correctly classified samples to the total number of samples [28]. We also calculated the error rate as the complement of accuracy. All these measures are suitable to evaluate the classification performance based on imbalanced data as found in the HAM10000 database. All metric results were calculated with respect to the class labels documented in the HAM10000 database archive. A summary diagram is depicted on Supplementary Figure S2.",
        "results": "**3. Results**\n\n_3.1. Classification Metrics across Eight Different DNNs_\n\nThe results of global accuracy and error rate for each DNN for the classification of seven pigmented skin lesions at low image resolution (300 \u00d7 224 RGB) are shown in Table 2. The average global accuracy for the eight DNNs reached 76.30% \u00b1 2.79, ranging from 74.05% (EfficientNetB5) to 82.47% (MobileNet). As shown in Supplementary Table S1, the TPR, TNR and geometric mean for each disease subtype varied according to the tested DNN. Almost all DNNs showed the highest TPR for melanocytic nevi classification when compared to the other pigmented lesions, with observed geometric mean values equal or lower than 0.65; interestingly, VGG16, VGG19 and MobileNet also showed high TPR for vascular lesion classification (geometric mean values lower than 0.65). In the case of melanoma and benign keratosis classification, all of them showed a TPR of approximately 0.5, with geometric mean values around 0.75. Likewise, basal cell carcinoma classification showed a TPR of approximately 0.5 and geometric mean values higher than 0.75 using ResNet50, ResNet101, SEResNet50 and EfficientNetB5. Similar results were observed for SEResNet5 and MobileNet and intraepithelial carcinoma classification.\n\n**Table 2. Classification metrics of HAM10000 images at two different resolutions and without aggregated clinical features using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 75.32 24.68 76.73 23.27  \nResNet50 74.56 25.44 75.97 24.03  \nResNet101 75.62 24.38 77.02 22.98  \nSEResNet50 77.82 22.22 79.13 20.87  \nVGG16 76.85 23.15 78.25 21.75  \nVGG19 74.21 25.79 75.62 24.38  \nEfficientNetB5 74.05 25.91 75.50 24.5  \nMobileNet 82.47 17.53 83.88 16.12  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network. 3 accuracy and error rates are expressed as percentages.\n\nUsing higher image resolution (600 \u00d7 450 RGB) improved the global accuracies across all DNNs (Table 2). The global accuracy average for the eight DNNs tested was 77.76 \u00b1 2.77, ranging from 75.50 (EfficientNetB5) to 83.88% (MobileNet). Although higher than the mean average observed with low-resolution images, the difference was not statistically significant (p = 0.07; Mann\u2013Whitney U test). The TPR, TNR and geometric mean values for each disease subtype are detailed in Supplementary Table S2. The highest TPR was observed for melanocytic nevi across all DNNs, and in the case of vascular lesions using VGG16 and MobileNet. On the contrary, the lowest TPR was observed for dermatofibrosis and intraepithelial carcinoma with ResNet34, ResNet50, SEResnet50 and EfficientNetB5. When compared with the low-resolution image parallel cases, the TPR, TNR and geometric mean values were quite similar for most of the disease subtypes. However, a drastic improvement of TPR was observed for dermatofibrosis using ResNet34 (from 0.23 to 0.34), ResNet50 (from 0.29 to 0.55), ResNet101 (from 0.26 to 0.42), VGG16 (from 0.17 to 0.4), VGG19 (from 0.26 to 0.40) and MobileNet (from 0.39 to 0.5). Similar TPR improvements were observed for the intraepithelial carcinoma classification using VGG19 (from 0.26 to 0.40). The average cascade framework runtime of the high-resolution classification model was 21.46+/\u22122.3 milliseconds per image, whereas the low-resolution model required only 18.6+/\u22121.22 milliseconds per image. Altogether, these results indicate that the tested DNNs can classify seven different types of pigmented skin lesions with accuracies higher than 0.7. No major differences in runtime were observed between the cascade framework input with low- or high resolution.\n\n_3.2. Classification Metrics of Different DNNs Aggregating Image and Clinical Features_\n\nWe then investigated if adding clinical features to the analysis could improve the classification accuracy of each DNN. As the HAM10000 dataset provides the sex, age and localization of the skin lesion associated to every image, we gathered these clinical features and aggregated them with the corresponding image to be used as input for each of the tested DNNs. Results are shown in Table 3. For low-resolution images, the addition of clinical data improved the global performance of all DNNs but MobileNet. The average global accuracy for the eight DNNs was 78.86% \u00b1 1.81, ranging from 75.73 (EfficientNetB5) to 81.24% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with low-resolution images without aggregated clinical features (p = 0.004; Mann\u2013Whitney U test). The highest increase was observed for VGG19 raising from 74.21% to 79.43 (5.22%). As shown in Supplementary Table S3, classification improvements were observed in almost all pigmented skin lesions for all DNNs. The highest marginal increases in accuracy were observed for dermatofibrosis where the TPR values for ResNet 50, SEResNet50, VGG16, EfficientNetB5 and MobileNet were raised by 15, 51, 22, 28 and 46%, respectively. Similarly, in the case of the intraepithelial carcinoma condition, TPR values increased 13 and 26% for SEResnet50 and VGG19, respectively, and for basal cell carcinoma classification, increased by 11, 14 and 12% for ResNet50, VGG19 and EfficientNetB5, respectively. For vascular lesions, the TPR values increased by 34, 14 and 29% for ResNet101, VGG19 and EfficientNetB5, respectively. Finally, adding clinical features to skin images also improved the melanoma condition accuracy for VGG16 (from 0.47 to 0.61) and VGG19 (from 0.44 to 0.58). Of note, a decreased TPR was observed for ResNet101 in the classification of dermatofibrosis (from 0.26 to 0.16) and intraepithelial carcinoma (from 0.37 to 0.29), and in melanoma for MobileNet (from 0.90 to 0.72).\n\n**Table 3. Classification metrics of HAM10000 low- and high-resolution images with aggregated clinical data using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 77.43 22.57 78.84 21.16  \nResNet50 79.31 20.69 80.72 19.28  \nResNet101 77.55 22.45 78.96 21.04  \nSEResNet50 80.01 19.99 80.72 19.28  \nVGG16 80.25 19.75 81.65 18.35  \nVGG19 79.43 20.57 79.02 20.98  \nEfficientNetB5 75.73 24.27 77.14 22.86  \nMobileNet 81.24 18.76 84.73 15.27  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network; 3 accuracy and error rates are expressed as percentages.\n\nFor high-resolution images, the average global accuracy also increased when clinical features were added to the model (Table 3). The average global accuracy was 80.22% \u00b1 2.30, ranging from 77.14 (EfficientNetB5) to 84.73% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with high-resolution images without aggregated clinical features (p = 0.02; Mann\u2013Whitney U test). This was particularly evident in the case of ResNet50 performance with an increase of 4.75%. The TPR, TNR and geometric mean values are shown in Supplementary Table S4. Major TPR improvements were observed for dermatofibrosis with ResNet34 (26%), ResNet101 (20%), VGG19 (17%) and EfficientNetB5 (28%). TPR increases were also observed for basal cell carcinoma with ResNet50 (12%) and EfficientNetB5 (13%), for vascular lesions with ResNet101 (37%) and EfficientNetB5 (28%), for melanoma with VGG16 (14%), VGG19 (32%) and MobileNet (16%) and intraepithelial carcinoma with VGG16 (24%). Interestingly, the TPR of ResNet50 and SEResNet50 were reduced for dermatofibrosis from 0.55 to 0.20 and from 0.44 to 0.29, respectively. When compared to the DNNs\u2019 performance with low-resolution images and clinical features, no major differences were observed (p = 0.21; Mann\u2013Whitney U test). Of interest, for dermatofibrosis classification, the TPR increased for ResNet34 (35%), ResNet101 (50%), VGG16 (50%) and VGG19 (24%); the exception was ResNet50 with a TPR reduction from 0.44 to 0.2. Altogether, these results indicate that the addition of information related to sex, age and localization of the lesion improves the accuracy of DNNs.\n\n_3.3. Performance across Synthetic Pigmented-Skin Lesion Images_\n\nWe applied an infoGAN to generate synthetic images from the HAM10000 dataset. As shown in Figure 2, the synthetic images seemed realistic and diverse. We then calculated the global accuracy for EfficientNetB5 for the classification of seven pigmented skin lesions. Of the 40 synthetic images analyzed, the network made a single error, so the certainty index was 97.5%; however, this value lacks any relevance since the GAN used for pseudo-labeling deliberately increases the definition limit of each class, inducing an improvement in the certainty of the classifier. Altogether, these data indicate that the synthetic samples are highly realistic and can be used as inputs to train DNNs on pigmented skin lesion classification.\n\n### A B C D\n\n E F G\n\n**Figure 2. Examples of synthetic images generated with the information maximizing generative adversarial network (infoGAN): (A) melanocytic nevus; (B) benign keratosis; (C) vascular lesion; (D) dermatofibroma; (E) intraepithelial carcinoma; (F) basal cell carcinoma; and (G) melanoma.**\n\n_3.4. Performance across General Practitioners with and without Assistance from DNNs Output_\n\nIn the first challenge, 22 general practitioners were asked to classify 162 images without any time constraint. The mean global accuracy and mean error rate were 27.74 and 72.26%, respectively. These results were similar to those previously published for non-dermatologists [9,10]. The best TPR (0.79) was obtained for the melanocytic nevi while the worse metrics were observed for vascular lesions (0.02), dermatofibrosis (0.01) and intraepithelial carcinoma (0.07) (Supplementary Table S5). As EfficientNetB5 was slightly less accurate than the other tested DNNs, we decided to use this framework as a comparator (see Table 1). In the same dataset, this DNN had a mean global accuracy of 78.40% and mean error rate 21.60% (Table 4). Compared to physicians, this was a relevant and significant difference as EfficientNetB5 showed a higher TPR in all disease subtypes (Supplementary Table S5).\n\n**Table 4. Classification metrics of non-dermatologists, general practitioners with or without use of the algorithmic platform and with time constraints.**\n\n**Condition** **Accuracy** **[1]** **Error Rate**\n\nEfficientNetB5 77.14 22.86  \nGPs [2] 17.29 82.71  \nGPs + AI 42.43 57.57  \n\n1 accuracy and error rate are expressed as percentages. 2 GPs: general practitioners; AI: artificial intelligence.\n\nIn the second challenge, 19 general practitioners were asked to classify 35 images with a time constraint of 45 s per image. The global accuracy for this classification was 17.29% (Table 4). In the same dataset, EfficientNetB5 achieved a global accuracy of 77.14%, significantly outperforming physicians. When general practitioners were given the opportunity to access the output of EfficientNetB5 per image, the global accuracy increased to 42.42%. This represented an increase of 25.13%. This result also indicated that, in some cases, physicians did not follow the recommendation of the DNN. Of note, the access of physicians to DNN prediction increased TPR for basal cell carcinoma (from 0.10 to 0.54) and melanoma (from 0.08 to 0.35) (Supplementary Table S6). In contrast, a small decrease in TPR for benign keratosis was observed (see Supplementary Table S6). Altogether, these results show that DNNs have the capability to classify seven different pigmented skin lesions with a level of competence higher to that of the general practitioners participating in these challenges. The access of DNN output by physicians improves their ability to classify pigmented skin lesions, particularly basal cell carcinoma and melanoma.",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions.",
        "abstract": "This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Pigmented Skin Image Dataset_\n\nThis study used images from the anonymous and annotated HAM10000 dataset publicly available through the International Skin Imaging Collaboration (ISIC) archive [21]. All downloaded images were selected using a random generator from the set of available images in the ISIC archive. We stochastically split the master set of 10,015 dermoscopic images into training (n = 8313; 83%) and test (n = 1702; 17%) datasets that were completely disjoint. Images included a representative collection of all-important diagnostic categories across the seven different types of pigmented lesions as detailed in Tschandl et al. [21]. These included melanocytic nevus, vascular skin lesions (including cherry angiomas, angiokeratomas, pyogenic granulomas and hemorrhages), benign keratoses (including seborrheic keratoses, solar lentigo and lichen-planus-like kertoses), dermatofibroma, intraepithelial carcinoma (including actinic keratoses and Bowen\u2019s disease), basal cell carcinoma and melanoma. Examples of images of each lesion type are depicted in Figure 1. The final composition of each dataset is shown in Table 1.\n\n_2.2. Deep Neural Networks_\n\nWe evaluated eight different DNNs, each characterized by a specific architecture. VGG16 and VGG19 contain 16 and 19 convolutional layers, respectively, with very small receptive fields, five max-pooling layers of size for carrying out spatial pooling, followed by three fully connected layers, with the final layer as the soft-max layer [22]. Rectification nonlinearity (ReLu) activation is applied to all hidden layers. The model also uses dropout regularization in the fully connected layers. ResNet34 is a 34-layer residual network while ResNet50 and ResNet101 are 50- and 101-layers deep, respectively. The architecture of all these DNNs is similar to the one found in VGG consisting mostly of 3 \u00d7 3 filters, however, instead, shortcut connections are inserted resulting into a residual network. SEResNet50 architecture is based on ResNet. A squeeze-and-excitation block is applied at the end of each non-identity branch of residual block [23]. Differently, and instead of increasing its size by adding more or deeper layers, EfficientNetB5 scales up the network width, depth and resolution with a set of fixed scaling coefficients [24]. Finally, MobileNet uses depth-wise separable convolutions which significantly reduce the number of parameters when compared to a network based on standard convolutions and the same depth across the structure in the networks [25]. The framework is 54 layers deep. For each DNN, the initial weights of all layers of the network were set up after pretraining with ImageNet. To assess both the performance of the algorithm and the enhanced training techniques as accurately as possible, we retrained each DNN a total of five times (folds), and each training run consisted of 90 epochs. Training using the curated image patches took approximately 6 h to complete, 45k iterations on a 4 GeForce GTX 1080 GPU configuration. Training accuracy for the curated patches reached maximum accuracy (100%) at around epoch 32, whereas the pretrained model only began to converge around epoch 25. All these DNNs were trained and tested with two different image input size: 300 \u00d7 224 RGB and 600 \u00d7 450 RGB. The low-resolution images were obtained by cropping, distorting and linear resizing the original high-resolution images. DNNs were also trained and tested without or with the clinical features (sex, age and location of the lesion) associated to every image in the HAM10000 database.\n\n_2.3. Image Preprocessing_\n\nWhen a deep convolutional neural network overfits, it works extremely well on training data but poorly on data it has never seen before. This is especially important in the field of dermatology because of the variability that exists in the images that the neural network will be analyzing. Two steps were taken to reduce overfitting. First, a dropout layer was added and set to 0.5. This results in 50% of the neurons to be randomly turned off during the training process and therefore reduce the likelihood of overfitting. The second step taken to reduce overfitting was to use data augmentation. In data augmentation, the images are modified to account for some of the variability that exists in image taking. To account for the grid location, the size of the dermatological manifestation and the angle of the image, the training images fed into the model were altered using zoom (25% probability to be increased between \u22121 and 2%, rotated between \u22126\u00ba and 6\u00ba and vertical/horizontal transfer from \u22122 to 2%), rotation (25% probability of being randomly rotated 90\u00ba clockwise), transposing (15% probability of making a random axial symmetry in one of its diagonals) and horizontal and vertical flipping randomly (50% probability of making a horizontal or vertical \u201cmirror\u201d or both) or following optical parameters from the different types of phone cameras such as shear and brightness (30% probability of contrast modification between \u22128 and 8%) or optical (80% probability of being distorted with a range between \u22126 and 6%) or grid (75% probability at step 4 with a range of \u221228\u201328%) distortions. These modifications were applied to 66% of the input images. The model was run once without data augmentation with dropout and once with data augmentation including drop out in both instances.\n\n_2.4. Generation of Synthetic Pigmented Skin Lesion Images Using an infoGAN_\n\nGenerative adversarial networks (GANs) are a type of generative model that attempt to synthesize novel data that are indistinguishable from the training data [26]. They consist of two neural networks, locked in competition: a generator that captures the data distribution and creates synthesized data (e.g., an image), and a discriminator that estimates the probability that a sample came from the training data rather than from the generator. The two networks are sealed in a zero-sum game, where the success of one corresponds to the failure of the other. The training procedure for the generator is to maximize the probability of the discriminator making an error [27]. Thus, this framework is based on a value function that one model seeks to maximize and the other seeks to minimize. Since the two networks are differentiable, the system generates a gradient that can be used to steer both networks to the right direction. We trained an information maximizing generative adversarial network (infoGAN) composed of a generator and a discriminator, on all high resolution images of the HAM10000 dataset [20]. The infoGAN was adapted to the progressive architecture of the model by splitting the structured code into parts and feeding each part to the network by conditioning activation in the corresponding block of the generator (Supplementary Figure S1). To prevent detrimental competition between the discriminator and generator, and to achieve convergence in an efficient way, we followed the recommendations detailed in Chen et al. [20]. Briefly, the discriminator and generator were composed in eight progressive blocks with an input/output of spatial resolution of 4 \u00d7 4 in the initial step up to 512 \u00d7 512 in step 7. The Batch size has gone dynamically from 128 in step 0 to 2 in the last step. Both the generator and discriminator were optimized using Adam with an initial learning rate of 0.0075 and exponential decay of 0.99 and a Wasserstein function with gradient penalty. The training has progressed in phases of progressive increased resolution. More specifically, the model was capable of generating high resolution images with isolated semantic features controlled by a set of real valued variables. Color, age, sex, localization and type of lesion were the most important semantic features discovered during the training in a completely unsupervised fashion without human input. After training, the generator produced novel images, similar to those in the dataset.\n\n_2.5. Contests among General Practitioners_\n\nTo compare the accuracy of DNNs with non-dermatologist practitioners, we conducted two different challenges. The first aimed to establish the accuracy of general practitioners in classifying images from the HAM10000 dataset without time constraint. For this purpose, a group of 22 general practitioners from any given center in Buenos Aires (Argentina) were given access to 163 images of the different skin pigmented lesions through an anonymous website specifically created for this purpose. Physicians could enter and exit the website without limitation. Alongside the image, recorded factors such as the age, sex and localization of the lesion were shown. All physicians were asked to classify every image within the seven different diagnostics. No incentives were offered for participation. To ensure fair comparisons between the results determined by general practitioners and those determined by the DNNs, the same 162 images were run with the DNN framework which had the worse accuracy metrics in low-resolution and without aggregated clinical features of the eight DNNs tested. To determine if physicians could benefit from access to the algorithmic tool during the own classification task, a second evaluation was conducted. A group of 19 general practitioners that voluntarily accepted to participate in the study was first asked to assess 35 images in a simulated exercise with time constraints (physicians had 45 s to classify every image). In a second step, physicians had access to the predictions of the same algorithmic framework used during the first challenge, the same group having classified each image based on both their criteria and the algorithmic output. For this task, a new set of 35 images was shown with the same time constraints. In both contests, the ethics committee waived ethical approval owing to the use of anonymized dermatologic images obtained from the publicly available HAM10000 dataset.\n\n_2.6. Statistical Analysis_\n\nAfter the model had been trained, a test step was performed in which 1702 images of the seven dermatological manifestations were used as input and the results were statistically analyzed. A confusion matrix was constructed based on comparing the frameworks\u2019 prediction with each of the actual labels. All analyses were performed and programmed via a Jupyter notebook in Python. Sensitivity, specificity, geometric mean, accuracy and error rate were calculated for each dermatological manifestation [28]. Sensitivity or true positive rate (TPR) represented the positive and correctly classified samples to the total number of positive samples. The specificity or true negative rate (TNR) was estimated as the ratio of the correctly classified negative samples to the total number of negative samples. Geometric means were calculated by using the product of TPR and TNR. Accuracy was defined as the ratio between the correctly classified samples to the total number of samples [28]. We also calculated the error rate as the complement of accuracy. All these measures are suitable to evaluate the classification performance based on imbalanced data as found in the HAM10000 database. All metric results were calculated with respect to the class labels documented in the HAM10000 database archive. A summary diagram is depicted on Supplementary Figure S2.",
        "results": "**3. Results**\n\n_3.1. Classification Metrics across Eight Different DNNs_\n\nThe results of global accuracy and error rate for each DNN for the classification of seven pigmented skin lesions at low image resolution (300 \u00d7 224 RGB) are shown in Table 2. The average global accuracy for the eight DNNs reached 76.30% \u00b1 2.79, ranging from 74.05% (EfficientNetB5) to 82.47% (MobileNet). As shown in Supplementary Table S1, the TPR, TNR and geometric mean for each disease subtype varied according to the tested DNN. Almost all DNNs showed the highest TPR for melanocytic nevi classification when compared to the other pigmented lesions, with observed geometric mean values equal or lower than 0.65; interestingly, VGG16, VGG19 and MobileNet also showed high TPR for vascular lesion classification (geometric mean values lower than 0.65). In the case of melanoma and benign keratosis classification, all of them showed a TPR of approximately 0.5, with geometric mean values around 0.75. Likewise, basal cell carcinoma classification showed a TPR of approximately 0.5 and geometric mean values higher than 0.75 using ResNet50, ResNet101, SEResNet50 and EfficientNetB5. Similar results were observed for SEResNet5 and MobileNet and intraepithelial carcinoma classification.\n\n**Table 2. Classification metrics of HAM10000 images at two different resolutions and without aggregated clinical features using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 75.32 24.68 76.73 23.27  \nResNet50 74.56 25.44 75.97 24.03  \nResNet101 75.62 24.38 77.02 22.98  \nSEResNet50 77.82 22.22 79.13 20.87  \nVGG16 76.85 23.15 78.25 21.75  \nVGG19 74.21 25.79 75.62 24.38  \nEfficientNetB5 74.05 25.91 75.50 24.5  \nMobileNet 82.47 17.53 83.88 16.12  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network. 3 accuracy and error rates are expressed as percentages.\n\nUsing higher image resolution (600 \u00d7 450 RGB) improved the global accuracies across all DNNs (Table 2). The global accuracy average for the eight DNNs tested was 77.76 \u00b1 2.77, ranging from 75.50 (EfficientNetB5) to 83.88% (MobileNet). Although higher than the mean average observed with low-resolution images, the difference was not statistically significant (p = 0.07; Mann\u2013Whitney U test). The TPR, TNR and geometric mean values for each disease subtype are detailed in Supplementary Table S2. The highest TPR was observed for melanocytic nevi across all DNNs, and in the case of vascular lesions using VGG16 and MobileNet. On the contrary, the lowest TPR was observed for dermatofibrosis and intraepithelial carcinoma with ResNet34, ResNet50, SEResnet50 and EfficientNetB5. When compared with the low-resolution image parallel cases, the TPR, TNR and geometric mean values were quite similar for most of the disease subtypes. However, a drastic improvement of TPR was observed for dermatofibrosis using ResNet34 (from 0.23 to 0.34), ResNet50 (from 0.29 to 0.55), ResNet101 (from 0.26 to 0.42), VGG16 (from 0.17 to 0.4), VGG19 (from 0.26 to 0.40) and MobileNet (from 0.39 to 0.5). Similar TPR improvements were observed for the intraepithelial carcinoma classification using VGG19 (from 0.26 to 0.40). The average cascade framework runtime of the high-resolution classification model was 21.46+/\u22122.3 milliseconds per image, whereas the low-resolution model required only 18.6+/\u22121.22 milliseconds per image. Altogether, these results indicate that the tested DNNs can classify seven different types of pigmented skin lesions with accuracies higher than 0.7. No major differences in runtime were observed between the cascade framework input with low- or high resolution.\n\n_3.2. Classification Metrics of Different DNNs Aggregating Image and Clinical Features_\n\nWe then investigated if adding clinical features to the analysis could improve the classification accuracy of each DNN. As the HAM10000 dataset provides the sex, age and localization of the skin lesion associated to every image, we gathered these clinical features and aggregated them with the corresponding image to be used as input for each of the tested DNNs. Results are shown in Table 3. For low-resolution images, the addition of clinical data improved the global performance of all DNNs but MobileNet. The average global accuracy for the eight DNNs was 78.86% \u00b1 1.81, ranging from 75.73 (EfficientNetB5) to 81.24% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with low-resolution images without aggregated clinical features (p = 0.004; Mann\u2013Whitney U test). The highest increase was observed for VGG19 raising from 74.21% to 79.43 (5.22%). As shown in Supplementary Table S3, classification improvements were observed in almost all pigmented skin lesions for all DNNs. The highest marginal increases in accuracy were observed for dermatofibrosis where the TPR values for ResNet 50, SEResNet50, VGG16, EfficientNetB5 and MobileNet were raised by 15, 51, 22, 28 and 46%, respectively. Similarly, in the case of the intraepithelial carcinoma condition, TPR values increased 13 and 26% for SEResnet50 and VGG19, respectively, and for basal cell carcinoma classification, increased by 11, 14 and 12% for ResNet50, VGG19 and EfficientNetB5, respectively. For vascular lesions, the TPR values increased by 34, 14 and 29% for ResNet101, VGG19 and EfficientNetB5, respectively. Finally, adding clinical features to skin images also improved the melanoma condition accuracy for VGG16 (from 0.47 to 0.61) and VGG19 (from 0.44 to 0.58). Of note, a decreased TPR was observed for ResNet101 in the classification of dermatofibrosis (from 0.26 to 0.16) and intraepithelial carcinoma (from 0.37 to 0.29), and in melanoma for MobileNet (from 0.90 to 0.72).\n\n**Table 3. Classification metrics of HAM10000 low- and high-resolution images with aggregated clinical data using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 77.43 22.57 78.84 21.16  \nResNet50 79.31 20.69 80.72 19.28  \nResNet101 77.55 22.45 78.96 21.04  \nSEResNet50 80.01 19.99 80.72 19.28  \nVGG16 80.25 19.75 81.65 18.35  \nVGG19 79.43 20.57 79.02 20.98  \nEfficientNetB5 75.73 24.27 77.14 22.86  \nMobileNet 81.24 18.76 84.73 15.27  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network; 3 accuracy and error rates are expressed as percentages.\n\nFor high-resolution images, the average global accuracy also increased when clinical features were added to the model (Table 3). The average global accuracy was 80.22% \u00b1 2.30, ranging from 77.14 (EfficientNetB5) to 84.73% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with high-resolution images without aggregated clinical features (p = 0.02; Mann\u2013Whitney U test). This was particularly evident in the case of ResNet50 performance with an increase of 4.75%. The TPR, TNR and geometric mean values are shown in Supplementary Table S4. Major TPR improvements were observed for dermatofibrosis with ResNet34 (26%), ResNet101 (20%), VGG19 (17%) and EfficientNetB5 (28%). TPR increases were also observed for basal cell carcinoma with ResNet50 (12%) and EfficientNetB5 (13%), for vascular lesions with ResNet101 (37%) and EfficientNetB5 (28%), for melanoma with VGG16 (14%), VGG19 (32%) and MobileNet (16%) and intraepithelial carcinoma with VGG16 (24%). Interestingly, the TPR of ResNet50 and SEResNet50 were reduced for dermatofibrosis from 0.55 to 0.20 and from 0.44 to 0.29, respectively. When compared to the DNNs\u2019 performance with low-resolution images and clinical features, no major differences were observed (p = 0.21; Mann\u2013Whitney U test). Of interest, for dermatofibrosis classification, the TPR increased for ResNet34 (35%), ResNet101 (50%), VGG16 (50%) and VGG19 (24%); the exception was ResNet50 with a TPR reduction from 0.44 to 0.2. Altogether, these results indicate that the addition of information related to sex, age and localization of the lesion improves the accuracy of DNNs.\n\n_3.3. Performance across Synthetic Pigmented-Skin Lesion Images_\n\nWe applied an infoGAN to generate synthetic images from the HAM10000 dataset. As shown in Figure 2, the synthetic images seemed realistic and diverse. We then calculated the global accuracy for EfficientNetB5 for the classification of seven pigmented skin lesions. Of the 40 synthetic images analyzed, the network made a single error, so the certainty index was 97.5%; however, this value lacks any relevance since the GAN used for pseudo-labeling deliberately increases the definition limit of each class, inducing an improvement in the certainty of the classifier. Altogether, these data indicate that the synthetic samples are highly realistic and can be used as inputs to train DNNs on pigmented skin lesion classification.\n\n### A B C D\n\n E F G\n\n**Figure 2. Examples of synthetic images generated with the information maximizing generative adversarial network (infoGAN): (A) melanocytic nevus; (B) benign keratosis; (C) vascular lesion; (D) dermatofibroma; (E) intraepithelial carcinoma; (F) basal cell carcinoma; and (G) melanoma.**\n\n_3.4. Performance across General Practitioners with and without Assistance from DNNs Output_\n\nIn the first challenge, 22 general practitioners were asked to classify 162 images without any time constraint. The mean global accuracy and mean error rate were 27.74 and 72.26%, respectively. These results were similar to those previously published for non-dermatologists [9,10]. The best TPR (0.79) was obtained for the melanocytic nevi while the worse metrics were observed for vascular lesions (0.02), dermatofibrosis (0.01) and intraepithelial carcinoma (0.07) (Supplementary Table S5). As EfficientNetB5 was slightly less accurate than the other tested DNNs, we decided to use this framework as a comparator (see Table 1). In the same dataset, this DNN had a mean global accuracy of 78.40% and mean error rate 21.60% (Table 4). Compared to physicians, this was a relevant and significant difference as EfficientNetB5 showed a higher TPR in all disease subtypes (Supplementary Table S5).\n\n**Table 4. Classification metrics of non-dermatologists, general practitioners with or without use of the algorithmic platform and with time constraints.**\n\n**Condition** **Accuracy** **[1]** **Error Rate**\n\nEfficientNetB5 77.14 22.86  \nGPs [2] 17.29 82.71  \nGPs + AI 42.43 57.57  \n\n1 accuracy and error rate are expressed as percentages. 2 GPs: general practitioners; AI: artificial intelligence.\n\nIn the second challenge, 19 general practitioners were asked to classify 35 images with a time constraint of 45 s per image. The global accuracy for this classification was 17.29% (Table 4). In the same dataset, EfficientNetB5 achieved a global accuracy of 77.14%, significantly outperforming physicians. When general practitioners were given the opportunity to access the output of EfficientNetB5 per image, the global accuracy increased to 42.42%. This represented an increase of 25.13%. This result also indicated that, in some cases, physicians did not follow the recommendation of the DNN. Of note, the access of physicians to DNN prediction increased TPR for basal cell carcinoma (from 0.10 to 0.54) and melanoma (from 0.08 to 0.35) (Supplementary Table S6). In contrast, a small decrease in TPR for benign keratosis was observed (see Supplementary Table S6). Altogether, these results show that DNNs have the capability to classify seven different pigmented skin lesions with a level of competence higher to that of the general practitioners participating in these challenges. The access of DNN output by physicians improves their ability to classify pigmented skin lesions, particularly basal cell carcinoma and melanoma.",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions.",
        "abstract": "This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Pigmented Skin Image Dataset_\n\nThis study used images from the anonymous and annotated HAM10000 dataset publicly available through the International Skin Imaging Collaboration (ISIC) archive [21]. All downloaded images were selected using a random generator from the set of available images in the ISIC archive. We stochastically split the master set of 10,015 dermoscopic images into training (n = 8313; 83%) and test (n = 1702; 17%) datasets that were completely disjoint. Images included a representative collection of all-important diagnostic categories across the seven different types of pigmented lesions as detailed in Tschandl et al. [21]. These included melanocytic nevus, vascular skin lesions (including cherry angiomas, angiokeratomas, pyogenic granulomas and hemorrhages), benign keratoses (including seborrheic keratoses, solar lentigo and lichen-planus-like kertoses), dermatofibroma, intraepithelial carcinoma (including actinic keratoses and Bowen\u2019s disease), basal cell carcinoma and melanoma. Examples of images of each lesion type are depicted in Figure 1. The final composition of each dataset is shown in Table 1.\n\n_2.2. Deep Neural Networks_\n\nWe evaluated eight different DNNs, each characterized by a specific architecture. VGG16 and VGG19 contain 16 and 19 convolutional layers, respectively, with very small receptive fields, five max-pooling layers of size for carrying out spatial pooling, followed by three fully connected layers, with the final layer as the soft-max layer [22]. Rectification nonlinearity (ReLu) activation is applied to all hidden layers. The model also uses dropout regularization in the fully connected layers. ResNet34 is a 34-layer residual network while ResNet50 and ResNet101 are 50- and 101-layers deep, respectively. The architecture of all these DNNs is similar to the one found in VGG consisting mostly of 3 \u00d7 3 filters, however, instead, shortcut connections are inserted resulting into a residual network. SEResNet50 architecture is based on ResNet. A squeeze-and-excitation block is applied at the end of each non-identity branch of residual block [23]. Differently, and instead of increasing its size by adding more or deeper layers, EfficientNetB5 scales up the network width, depth and resolution with a set of fixed scaling coefficients [24]. Finally, MobileNet uses depth-wise separable convolutions which significantly reduce the number of parameters when compared to a network based on standard convolutions and the same depth across the structure in the networks [25]. The framework is 54 layers deep. For each DNN, the initial weights of all layers of the network were set up after pretraining with ImageNet. To assess both the performance of the algorithm and the enhanced training techniques as accurately as possible, we retrained each DNN a total of five times (folds), and each training run consisted of 90 epochs. Training using the curated image patches took approximately 6 h to complete, 45k iterations on a 4 GeForce GTX 1080 GPU configuration. Training accuracy for the curated patches reached maximum accuracy (100%) at around epoch 32, whereas the pretrained model only began to converge around epoch 25. All these DNNs were trained and tested with two different image input size: 300 \u00d7 224 RGB and 600 \u00d7 450 RGB. The low-resolution images were obtained by cropping, distorting and linear resizing the original high-resolution images. DNNs were also trained and tested without or with the clinical features (sex, age and location of the lesion) associated to every image in the HAM10000 database.\n\n_2.3. Image Preprocessing_\n\nWhen a deep convolutional neural network overfits, it works extremely well on training data but poorly on data it has never seen before. This is especially important in the field of dermatology because of the variability that exists in the images that the neural network will be analyzing. Two steps were taken to reduce overfitting. First, a dropout layer was added and set to 0.5. This results in 50% of the neurons to be randomly turned off during the training process and therefore reduce the likelihood of overfitting. The second step taken to reduce overfitting was to use data augmentation. In data augmentation, the images are modified to account for some of the variability that exists in image taking. To account for the grid location, the size of the dermatological manifestation and the angle of the image, the training images fed into the model were altered using zoom (25% probability to be increased between \u22121 and 2%, rotated between \u22126\u00ba and 6\u00ba and vertical/horizontal transfer from \u22122 to 2%), rotation (25% probability of being randomly rotated 90\u00ba clockwise), transposing (15% probability of making a random axial symmetry in one of its diagonals) and horizontal and vertical flipping randomly (50% probability of making a horizontal or vertical \u201cmirror\u201d or both) or following optical parameters from the different types of phone cameras such as shear and brightness (30% probability of contrast modification between \u22128 and 8%) or optical (80% probability of being distorted with a range between \u22126 and 6%) or grid (75% probability at step 4 with a range of \u221228\u201328%) distortions. These modifications were applied to 66% of the input images. The model was run once without data augmentation with dropout and once with data augmentation including drop out in both instances.\n\n_2.4. Generation of Synthetic Pigmented Skin Lesion Images Using an infoGAN_\n\nGenerative adversarial networks (GANs) are a type of generative model that attempt to synthesize novel data that are indistinguishable from the training data [26]. They consist of two neural networks, locked in competition: a generator that captures the data distribution and creates synthesized data (e.g., an image), and a discriminator that estimates the probability that a sample came from the training data rather than from the generator. The two networks are sealed in a zero-sum game, where the success of one corresponds to the failure of the other. The training procedure for the generator is to maximize the probability of the discriminator making an error [27]. Thus, this framework is based on a value function that one model seeks to maximize and the other seeks to minimize. Since the two networks are differentiable, the system generates a gradient that can be used to steer both networks to the right direction. We trained an information maximizing generative adversarial network (infoGAN) composed of a generator and a discriminator, on all high resolution images of the HAM10000 dataset [20]. The infoGAN was adapted to the progressive architecture of the model by splitting the structured code into parts and feeding each part to the network by conditioning activation in the corresponding block of the generator (Supplementary Figure S1). To prevent detrimental competition between the discriminator and generator, and to achieve convergence in an efficient way, we followed the recommendations detailed in Chen et al. [20]. Briefly, the discriminator and generator were composed in eight progressive blocks with an input/output of spatial resolution of 4 \u00d7 4 in the initial step up to 512 \u00d7 512 in step 7. The Batch size has gone dynamically from 128 in step 0 to 2 in the last step. Both the generator and discriminator were optimized using Adam with an initial learning rate of 0.0075 and exponential decay of 0.99 and a Wasserstein function with gradient penalty. The training has progressed in phases of progressive increased resolution. More specifically, the model was capable of generating high resolution images with isolated semantic features controlled by a set of real valued variables. Color, age, sex, localization and type of lesion were the most important semantic features discovered during the training in a completely unsupervised fashion without human input. After training, the generator produced novel images, similar to those in the dataset.\n\n_2.5. Contests among General Practitioners_\n\nTo compare the accuracy of DNNs with non-dermatologist practitioners, we conducted two different challenges. The first aimed to establish the accuracy of general practitioners in classifying images from the HAM10000 dataset without time constraint. For this purpose, a group of 22 general practitioners from any given center in Buenos Aires (Argentina) were given access to 163 images of the different skin pigmented lesions through an anonymous website specifically created for this purpose. Physicians could enter and exit the website without limitation. Alongside the image, recorded factors such as the age, sex and localization of the lesion were shown. All physicians were asked to classify every image within the seven different diagnostics. No incentives were offered for participation. To ensure fair comparisons between the results determined by general practitioners and those determined by the DNNs, the same 162 images were run with the DNN framework which had the worse accuracy metrics in low-resolution and without aggregated clinical features of the eight DNNs tested. To determine if physicians could benefit from access to the algorithmic tool during the own classification task, a second evaluation was conducted. A group of 19 general practitioners that voluntarily accepted to participate in the study was first asked to assess 35 images in a simulated exercise with time constraints (physicians had 45 s to classify every image). In a second step, physicians had access to the predictions of the same algorithmic framework used during the first challenge, the same group having classified each image based on both their criteria and the algorithmic output. For this task, a new set of 35 images was shown with the same time constraints. In both contests, the ethics committee waived ethical approval owing to the use of anonymized dermatologic images obtained from the publicly available HAM10000 dataset.\n\n_2.6. Statistical Analysis_\n\nAfter the model had been trained, a test step was performed in which 1702 images of the seven dermatological manifestations were used as input and the results were statistically analyzed. A confusion matrix was constructed based on comparing the frameworks\u2019 prediction with each of the actual labels. All analyses were performed and programmed via a Jupyter notebook in Python. Sensitivity, specificity, geometric mean, accuracy and error rate were calculated for each dermatological manifestation [28]. Sensitivity or true positive rate (TPR) represented the positive and correctly classified samples to the total number of positive samples. The specificity or true negative rate (TNR) was estimated as the ratio of the correctly classified negative samples to the total number of negative samples. Geometric means were calculated by using the product of TPR and TNR. Accuracy was defined as the ratio between the correctly classified samples to the total number of samples [28]. We also calculated the error rate as the complement of accuracy. All these measures are suitable to evaluate the classification performance based on imbalanced data as found in the HAM10000 database. All metric results were calculated with respect to the class labels documented in the HAM10000 database archive. A summary diagram is depicted on Supplementary Figure S2.",
        "results": "**3. Results**\n\n_3.1. Classification Metrics across Eight Different DNNs_\n\nThe results of global accuracy and error rate for each DNN for the classification of seven pigmented skin lesions at low image resolution (300 \u00d7 224 RGB) are shown in Table 2. The average global accuracy for the eight DNNs reached 76.30% \u00b1 2.79, ranging from 74.05% (EfficientNetB5) to 82.47% (MobileNet). As shown in Supplementary Table S1, the TPR, TNR and geometric mean for each disease subtype varied according to the tested DNN. Almost all DNNs showed the highest TPR for melanocytic nevi classification when compared to the other pigmented lesions, with observed geometric mean values equal or lower than 0.65; interestingly, VGG16, VGG19 and MobileNet also showed high TPR for vascular lesion classification (geometric mean values lower than 0.65). In the case of melanoma and benign keratosis classification, all of them showed a TPR of approximately 0.5, with geometric mean values around 0.75. Likewise, basal cell carcinoma classification showed a TPR of approximately 0.5 and geometric mean values higher than 0.75 using ResNet50, ResNet101, SEResNet50 and EfficientNetB5. Similar results were observed for SEResNet5 and MobileNet and intraepithelial carcinoma classification.\n\n**Table 2. Classification metrics of HAM10000 images at two different resolutions and without aggregated clinical features using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 75.32 24.68 76.73 23.27  \nResNet50 74.56 25.44 75.97 24.03  \nResNet101 75.62 24.38 77.02 22.98  \nSEResNet50 77.82 22.22 79.13 20.87  \nVGG16 76.85 23.15 78.25 21.75  \nVGG19 74.21 25.79 75.62 24.38  \nEfficientNetB5 74.05 25.91 75.50 24.5  \nMobileNet 82.47 17.53 83.88 16.12  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network. 3 accuracy and error rates are expressed as percentages.\n\nUsing higher image resolution (600 \u00d7 450 RGB) improved the global accuracies across all DNNs (Table 2). The global accuracy average for the eight DNNs tested was 77.76 \u00b1 2.77, ranging from 75.50 (EfficientNetB5) to 83.88% (MobileNet). Although higher than the mean average observed with low-resolution images, the difference was not statistically significant (p = 0.07; Mann\u2013Whitney U test). The TPR, TNR and geometric mean values for each disease subtype are detailed in Supplementary Table S2. The highest TPR was observed for melanocytic nevi across all DNNs, and in the case of vascular lesions using VGG16 and MobileNet. On the contrary, the lowest TPR was observed for dermatofibrosis and intraepithelial carcinoma with ResNet34, ResNet50, SEResnet50 and EfficientNetB5. When compared with the low-resolution image parallel cases, the TPR, TNR and geometric mean values were quite similar for most of the disease subtypes. However, a drastic improvement of TPR was observed for dermatofibrosis using ResNet34 (from 0.23 to 0.34), ResNet50 (from 0.29 to 0.55), ResNet101 (from 0.26 to 0.42), VGG16 (from 0.17 to 0.4), VGG19 (from 0.26 to 0.40) and MobileNet (from 0.39 to 0.5). Similar TPR improvements were observed for the intraepithelial carcinoma classification using VGG19 (from 0.26 to 0.40). The average cascade framework runtime of the high-resolution classification model was 21.46+/\u22122.3 milliseconds per image, whereas the low-resolution model required only 18.6+/\u22121.22 milliseconds per image. Altogether, these results indicate that the tested DNNs can classify seven different types of pigmented skin lesions with accuracies higher than 0.7. No major differences in runtime were observed between the cascade framework input with low- or high resolution.\n\n_3.2. Classification Metrics of Different DNNs Aggregating Image and Clinical Features_\n\nWe then investigated if adding clinical features to the analysis could improve the classification accuracy of each DNN. As the HAM10000 dataset provides the sex, age and localization of the skin lesion associated to every image, we gathered these clinical features and aggregated them with the corresponding image to be used as input for each of the tested DNNs. Results are shown in Table 3. For low-resolution images, the addition of clinical data improved the global performance of all DNNs but MobileNet. The average global accuracy for the eight DNNs was 78.86% \u00b1 1.81, ranging from 75.73 (EfficientNetB5) to 81.24% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with low-resolution images without aggregated clinical features (p = 0.004; Mann\u2013Whitney U test). The highest increase was observed for VGG19 raising from 74.21% to 79.43 (5.22%). As shown in Supplementary Table S3, classification improvements were observed in almost all pigmented skin lesions for all DNNs. The highest marginal increases in accuracy were observed for dermatofibrosis where the TPR values for ResNet 50, SEResNet50, VGG16, EfficientNetB5 and MobileNet were raised by 15, 51, 22, 28 and 46%, respectively. Similarly, in the case of the intraepithelial carcinoma condition, TPR values increased 13 and 26% for SEResnet50 and VGG19, respectively, and for basal cell carcinoma classification, increased by 11, 14 and 12% for ResNet50, VGG19 and EfficientNetB5, respectively. For vascular lesions, the TPR values increased by 34, 14 and 29% for ResNet101, VGG19 and EfficientNetB5, respectively. Finally, adding clinical features to skin images also improved the melanoma condition accuracy for VGG16 (from 0.47 to 0.61) and VGG19 (from 0.44 to 0.58). Of note, a decreased TPR was observed for ResNet101 in the classification of dermatofibrosis (from 0.26 to 0.16) and intraepithelial carcinoma (from 0.37 to 0.29), and in melanoma for MobileNet (from 0.90 to 0.72).\n\n**Table 3. Classification metrics of HAM10000 low- and high-resolution images with aggregated clinical data using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 77.43 22.57 78.84 21.16  \nResNet50 79.31 20.69 80.72 19.28  \nResNet101 77.55 22.45 78.96 21.04  \nSEResNet50 80.01 19.99 80.72 19.28  \nVGG16 80.25 19.75 81.65 18.35  \nVGG19 79.43 20.57 79.02 20.98  \nEfficientNetB5 75.73 24.27 77.14 22.86  \nMobileNet 81.24 18.76 84.73 15.27  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network; 3 accuracy and error rates are expressed as percentages.\n\nFor high-resolution images, the average global accuracy also increased when clinical features were added to the model (Table 3). The average global accuracy was 80.22% \u00b1 2.30, ranging from 77.14 (EfficientNetB5) to 84.73% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with high-resolution images without aggregated clinical features (p = 0.02; Mann\u2013Whitney U test). This was particularly evident in the case of ResNet50 performance with an increase of 4.75%. The TPR, TNR and geometric mean values are shown in Supplementary Table S4. Major TPR improvements were observed for dermatofibrosis with ResNet34 (26%), ResNet101 (20%), VGG19 (17%) and EfficientNetB5 (28%). TPR increases were also observed for basal cell carcinoma with ResNet50 (12%) and EfficientNetB5 (13%), for vascular lesions with ResNet101 (37%) and EfficientNetB5 (28%), for melanoma with VGG16 (14%), VGG19 (32%) and MobileNet (16%) and intraepithelial carcinoma with VGG16 (24%). Interestingly, the TPR of ResNet50 and SEResNet50 were reduced for dermatofibrosis from 0.55 to 0.20 and from 0.44 to 0.29, respectively. When compared to the DNNs\u2019 performance with low-resolution images and clinical features, no major differences were observed (p = 0.21; Mann\u2013Whitney U test). Of interest, for dermatofibrosis classification, the TPR increased for ResNet34 (35%), ResNet101 (50%), VGG16 (50%) and VGG19 (24%); the exception was ResNet50 with a TPR reduction from 0.44 to 0.2. Altogether, these results indicate that the addition of information related to sex, age and localization of the lesion improves the accuracy of DNNs.\n\n_3.3. Performance across Synthetic Pigmented-Skin Lesion Images_\n\nWe applied an infoGAN to generate synthetic images from the HAM10000 dataset. As shown in Figure 2, the synthetic images seemed realistic and diverse. We then calculated the global accuracy for EfficientNetB5 for the classification of seven pigmented skin lesions. Of the 40 synthetic images analyzed, the network made a single error, so the certainty index was 97.5%; however, this value lacks any relevance since the GAN used for pseudo-labeling deliberately increases the definition limit of each class, inducing an improvement in the certainty of the classifier. Altogether, these data indicate that the synthetic samples are highly realistic and can be used as inputs to train DNNs on pigmented skin lesion classification.\n\n### A B C D\n\n E F G\n\n**Figure 2. Examples of synthetic images generated with the information maximizing generative adversarial network (infoGAN): (A) melanocytic nevus; (B) benign keratosis; (C) vascular lesion; (D) dermatofibroma; (E) intraepithelial carcinoma; (F) basal cell carcinoma; and (G) melanoma.**\n\n_3.4. Performance across General Practitioners with and without Assistance from DNNs Output_\n\nIn the first challenge, 22 general practitioners were asked to classify 162 images without any time constraint. The mean global accuracy and mean error rate were 27.74 and 72.26%, respectively. These results were similar to those previously published for non-dermatologists [9,10]. The best TPR (0.79) was obtained for the melanocytic nevi while the worse metrics were observed for vascular lesions (0.02), dermatofibrosis (0.01) and intraepithelial carcinoma (0.07) (Supplementary Table S5). As EfficientNetB5 was slightly less accurate than the other tested DNNs, we decided to use this framework as a comparator (see Table 1). In the same dataset, this DNN had a mean global accuracy of 78.40% and mean error rate 21.60% (Table 4). Compared to physicians, this was a relevant and significant difference as EfficientNetB5 showed a higher TPR in all disease subtypes (Supplementary Table S5).\n\n**Table 4. Classification metrics of non-dermatologists, general practitioners with or without use of the algorithmic platform and with time constraints.**\n\n**Condition** **Accuracy** **[1]** **Error Rate**\n\nEfficientNetB5 77.14 22.86  \nGPs [2] 17.29 82.71  \nGPs + AI 42.43 57.57  \n\n1 accuracy and error rate are expressed as percentages. 2 GPs: general practitioners; AI: artificial intelligence.\n\nIn the second challenge, 19 general practitioners were asked to classify 35 images with a time constraint of 45 s per image. The global accuracy for this classification was 17.29% (Table 4). In the same dataset, EfficientNetB5 achieved a global accuracy of 77.14%, significantly outperforming physicians. When general practitioners were given the opportunity to access the output of EfficientNetB5 per image, the global accuracy increased to 42.42%. This represented an increase of 25.13%. This result also indicated that, in some cases, physicians did not follow the recommendation of the DNN. Of note, the access of physicians to DNN prediction increased TPR for basal cell carcinoma (from 0.10 to 0.54) and melanoma (from 0.08 to 0.35) (Supplementary Table S6). In contrast, a small decrease in TPR for benign keratosis was observed (see Supplementary Table S6). Altogether, these results show that DNNs have the capability to classify seven different pigmented skin lesions with a level of competence higher to that of the general practitioners participating in these challenges. The access of DNN output by physicians improves their ability to classify pigmented skin lesions, particularly basal cell carcinoma and melanoma.",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Unclear"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions.",
        "abstract": "This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Pigmented Skin Image Dataset_\n\nThis study used images from the anonymous and annotated HAM10000 dataset publicly available through the International Skin Imaging Collaboration (ISIC) archive [21]. All downloaded images were selected using a random generator from the set of available images in the ISIC archive. We stochastically split the master set of 10,015 dermoscopic images into training (n = 8313; 83%) and test (n = 1702; 17%) datasets that were completely disjoint. Images included a representative collection of all-important diagnostic categories across the seven different types of pigmented lesions as detailed in Tschandl et al. [21]. These included melanocytic nevus, vascular skin lesions (including cherry angiomas, angiokeratomas, pyogenic granulomas and hemorrhages), benign keratoses (including seborrheic keratoses, solar lentigo and lichen-planus-like kertoses), dermatofibroma, intraepithelial carcinoma (including actinic keratoses and Bowen\u2019s disease), basal cell carcinoma and melanoma. Examples of images of each lesion type are depicted in Figure 1. The final composition of each dataset is shown in Table 1.\n\n_2.2. Deep Neural Networks_\n\nWe evaluated eight different DNNs, each characterized by a specific architecture. VGG16 and VGG19 contain 16 and 19 convolutional layers, respectively, with very small receptive fields, five max-pooling layers of size for carrying out spatial pooling, followed by three fully connected layers, with the final layer as the soft-max layer [22]. Rectification nonlinearity (ReLu) activation is applied to all hidden layers. The model also uses dropout regularization in the fully connected layers. ResNet34 is a 34-layer residual network while ResNet50 and ResNet101 are 50- and 101-layers deep, respectively. The architecture of all these DNNs is similar to the one found in VGG consisting mostly of 3 \u00d7 3 filters, however, instead, shortcut connections are inserted resulting into a residual network. SEResNet50 architecture is based on ResNet. A squeeze-and-excitation block is applied at the end of each non-identity branch of residual block [23]. Differently, and instead of increasing its size by adding more or deeper layers, EfficientNetB5 scales up the network width, depth and resolution with a set of fixed scaling coefficients [24]. Finally, MobileNet uses depth-wise separable convolutions which significantly reduce the number of parameters when compared to a network based on standard convolutions and the same depth across the structure in the networks [25]. The framework is 54 layers deep. For each DNN, the initial weights of all layers of the network were set up after pretraining with ImageNet. To assess both the performance of the algorithm and the enhanced training techniques as accurately as possible, we retrained each DNN a total of five times (folds), and each training run consisted of 90 epochs. Training using the curated image patches took approximately 6 h to complete, 45k iterations on a 4 GeForce GTX 1080 GPU configuration. Training accuracy for the curated patches reached maximum accuracy (100%) at around epoch 32, whereas the pretrained model only began to converge around epoch 25. All these DNNs were trained and tested with two different image input size: 300 \u00d7 224 RGB and 600 \u00d7 450 RGB. The low-resolution images were obtained by cropping, distorting and linear resizing the original high-resolution images. DNNs were also trained and tested without or with the clinical features (sex, age and location of the lesion) associated to every image in the HAM10000 database.\n\n_2.3. Image Preprocessing_\n\nWhen a deep convolutional neural network overfits, it works extremely well on training data but poorly on data it has never seen before. This is especially important in the field of dermatology because of the variability that exists in the images that the neural network will be analyzing. Two steps were taken to reduce overfitting. First, a dropout layer was added and set to 0.5. This results in 50% of the neurons to be randomly turned off during the training process and therefore reduce the likelihood of overfitting. The second step taken to reduce overfitting was to use data augmentation. In data augmentation, the images are modified to account for some of the variability that exists in image taking. To account for the grid location, the size of the dermatological manifestation and the angle of the image, the training images fed into the model were altered using zoom (25% probability to be increased between \u22121 and 2%, rotated between \u22126\u00ba and 6\u00ba and vertical/horizontal transfer from \u22122 to 2%), rotation (25% probability of being randomly rotated 90\u00ba clockwise), transposing (15% probability of making a random axial symmetry in one of its diagonals) and horizontal and vertical flipping randomly (50% probability of making a horizontal or vertical \u201cmirror\u201d or both) or following optical parameters from the different types of phone cameras such as shear and brightness (30% probability of contrast modification between \u22128 and 8%) or optical (80% probability of being distorted with a range between \u22126 and 6%) or grid (75% probability at step 4 with a range of \u221228\u201328%) distortions. These modifications were applied to 66% of the input images. The model was run once without data augmentation with dropout and once with data augmentation including drop out in both instances.\n\n_2.4. Generation of Synthetic Pigmented Skin Lesion Images Using an infoGAN_\n\nGenerative adversarial networks (GANs) are a type of generative model that attempt to synthesize novel data that are indistinguishable from the training data [26]. They consist of two neural networks, locked in competition: a generator that captures the data distribution and creates synthesized data (e.g., an image), and a discriminator that estimates the probability that a sample came from the training data rather than from the generator. The two networks are sealed in a zero-sum game, where the success of one corresponds to the failure of the other. The training procedure for the generator is to maximize the probability of the discriminator making an error [27]. Thus, this framework is based on a value function that one model seeks to maximize and the other seeks to minimize. Since the two networks are differentiable, the system generates a gradient that can be used to steer both networks to the right direction. We trained an information maximizing generative adversarial network (infoGAN) composed of a generator and a discriminator, on all high resolution images of the HAM10000 dataset [20]. The infoGAN was adapted to the progressive architecture of the model by splitting the structured code into parts and feeding each part to the network by conditioning activation in the corresponding block of the generator (Supplementary Figure S1). To prevent detrimental competition between the discriminator and generator, and to achieve convergence in an efficient way, we followed the recommendations detailed in Chen et al. [20]. Briefly, the discriminator and generator were composed in eight progressive blocks with an input/output of spatial resolution of 4 \u00d7 4 in the initial step up to 512 \u00d7 512 in step 7. The Batch size has gone dynamically from 128 in step 0 to 2 in the last step. Both the generator and discriminator were optimized using Adam with an initial learning rate of 0.0075 and exponential decay of 0.99 and a Wasserstein function with gradient penalty. The training has progressed in phases of progressive increased resolution. More specifically, the model was capable of generating high resolution images with isolated semantic features controlled by a set of real valued variables. Color, age, sex, localization and type of lesion were the most important semantic features discovered during the training in a completely unsupervised fashion without human input. After training, the generator produced novel images, similar to those in the dataset.\n\n_2.5. Contests among General Practitioners_\n\nTo compare the accuracy of DNNs with non-dermatologist practitioners, we conducted two different challenges. The first aimed to establish the accuracy of general practitioners in classifying images from the HAM10000 dataset without time constraint. For this purpose, a group of 22 general practitioners from any given center in Buenos Aires (Argentina) were given access to 163 images of the different skin pigmented lesions through an anonymous website specifically created for this purpose. Physicians could enter and exit the website without limitation. Alongside the image, recorded factors such as the age, sex and localization of the lesion were shown. All physicians were asked to classify every image within the seven different diagnostics. No incentives were offered for participation. To ensure fair comparisons between the results determined by general practitioners and those determined by the DNNs, the same 162 images were run with the DNN framework which had the worse accuracy metrics in low-resolution and without aggregated clinical features of the eight DNNs tested. To determine if physicians could benefit from access to the algorithmic tool during the own classification task, a second evaluation was conducted. A group of 19 general practitioners that voluntarily accepted to participate in the study was first asked to assess 35 images in a simulated exercise with time constraints (physicians had 45 s to classify every image). In a second step, physicians had access to the predictions of the same algorithmic framework used during the first challenge, the same group having classified each image based on both their criteria and the algorithmic output. For this task, a new set of 35 images was shown with the same time constraints. In both contests, the ethics committee waived ethical approval owing to the use of anonymized dermatologic images obtained from the publicly available HAM10000 dataset.\n\n_2.6. Statistical Analysis_\n\nAfter the model had been trained, a test step was performed in which 1702 images of the seven dermatological manifestations were used as input and the results were statistically analyzed. A confusion matrix was constructed based on comparing the frameworks\u2019 prediction with each of the actual labels. All analyses were performed and programmed via a Jupyter notebook in Python. Sensitivity, specificity, geometric mean, accuracy and error rate were calculated for each dermatological manifestation [28]. Sensitivity or true positive rate (TPR) represented the positive and correctly classified samples to the total number of positive samples. The specificity or true negative rate (TNR) was estimated as the ratio of the correctly classified negative samples to the total number of negative samples. Geometric means were calculated by using the product of TPR and TNR. Accuracy was defined as the ratio between the correctly classified samples to the total number of samples [28]. We also calculated the error rate as the complement of accuracy. All these measures are suitable to evaluate the classification performance based on imbalanced data as found in the HAM10000 database. All metric results were calculated with respect to the class labels documented in the HAM10000 database archive. A summary diagram is depicted on Supplementary Figure S2.",
        "results": "**3. Results**\n\n_3.1. Classification Metrics across Eight Different DNNs_\n\nThe results of global accuracy and error rate for each DNN for the classification of seven pigmented skin lesions at low image resolution (300 \u00d7 224 RGB) are shown in Table 2. The average global accuracy for the eight DNNs reached 76.30% \u00b1 2.79, ranging from 74.05% (EfficientNetB5) to 82.47% (MobileNet). As shown in Supplementary Table S1, the TPR, TNR and geometric mean for each disease subtype varied according to the tested DNN. Almost all DNNs showed the highest TPR for melanocytic nevi classification when compared to the other pigmented lesions, with observed geometric mean values equal or lower than 0.65; interestingly, VGG16, VGG19 and MobileNet also showed high TPR for vascular lesion classification (geometric mean values lower than 0.65). In the case of melanoma and benign keratosis classification, all of them showed a TPR of approximately 0.5, with geometric mean values around 0.75. Likewise, basal cell carcinoma classification showed a TPR of approximately 0.5 and geometric mean values higher than 0.75 using ResNet50, ResNet101, SEResNet50 and EfficientNetB5. Similar results were observed for SEResNet5 and MobileNet and intraepithelial carcinoma classification.\n\n**Table 2. Classification metrics of HAM10000 images at two different resolutions and without aggregated clinical features using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 75.32 24.68 76.73 23.27  \nResNet50 74.56 25.44 75.97 24.03  \nResNet101 75.62 24.38 77.02 22.98  \nSEResNet50 77.82 22.22 79.13 20.87  \nVGG16 76.85 23.15 78.25 21.75  \nVGG19 74.21 25.79 75.62 24.38  \nEfficientNetB5 74.05 25.91 75.50 24.5  \nMobileNet 82.47 17.53 83.88 16.12  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network. 3 accuracy and error rates are expressed as percentages.\n\nUsing higher image resolution (600 \u00d7 450 RGB) improved the global accuracies across all DNNs (Table 2). The global accuracy average for the eight DNNs tested was 77.76 \u00b1 2.77, ranging from 75.50 (EfficientNetB5) to 83.88% (MobileNet). Although higher than the mean average observed with low-resolution images, the difference was not statistically significant (p = 0.07; Mann\u2013Whitney U test). The TPR, TNR and geometric mean values for each disease subtype are detailed in Supplementary Table S2. The highest TPR was observed for melanocytic nevi across all DNNs, and in the case of vascular lesions using VGG16 and MobileNet. On the contrary, the lowest TPR was observed for dermatofibrosis and intraepithelial carcinoma with ResNet34, ResNet50, SEResnet50 and EfficientNetB5. When compared with the low-resolution image parallel cases, the TPR, TNR and geometric mean values were quite similar for most of the disease subtypes. However, a drastic improvement of TPR was observed for dermatofibrosis using ResNet34 (from 0.23 to 0.34), ResNet50 (from 0.29 to 0.55), ResNet101 (from 0.26 to 0.42), VGG16 (from 0.17 to 0.4), VGG19 (from 0.26 to 0.40) and MobileNet (from 0.39 to 0.5). Similar TPR improvements were observed for the intraepithelial carcinoma classification using VGG19 (from 0.26 to 0.40). The average cascade framework runtime of the high-resolution classification model was 21.46+/\u22122.3 milliseconds per image, whereas the low-resolution model required only 18.6+/\u22121.22 milliseconds per image. Altogether, these results indicate that the tested DNNs can classify seven different types of pigmented skin lesions with accuracies higher than 0.7. No major differences in runtime were observed between the cascade framework input with low- or high resolution.\n\n_3.2. Classification Metrics of Different DNNs Aggregating Image and Clinical Features_\n\nWe then investigated if adding clinical features to the analysis could improve the classification accuracy of each DNN. As the HAM10000 dataset provides the sex, age and localization of the skin lesion associated to every image, we gathered these clinical features and aggregated them with the corresponding image to be used as input for each of the tested DNNs. Results are shown in Table 3. For low-resolution images, the addition of clinical data improved the global performance of all DNNs but MobileNet. The average global accuracy for the eight DNNs was 78.86% \u00b1 1.81, ranging from 75.73 (EfficientNetB5) to 81.24% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with low-resolution images without aggregated clinical features (p = 0.004; Mann\u2013Whitney U test). The highest increase was observed for VGG19 raising from 74.21% to 79.43 (5.22%). As shown in Supplementary Table S3, classification improvements were observed in almost all pigmented skin lesions for all DNNs. The highest marginal increases in accuracy were observed for dermatofibrosis where the TPR values for ResNet 50, SEResNet50, VGG16, EfficientNetB5 and MobileNet were raised by 15, 51, 22, 28 and 46%, respectively. Similarly, in the case of the intraepithelial carcinoma condition, TPR values increased 13 and 26% for SEResnet50 and VGG19, respectively, and for basal cell carcinoma classification, increased by 11, 14 and 12% for ResNet50, VGG19 and EfficientNetB5, respectively. For vascular lesions, the TPR values increased by 34, 14 and 29% for ResNet101, VGG19 and EfficientNetB5, respectively. Finally, adding clinical features to skin images also improved the melanoma condition accuracy for VGG16 (from 0.47 to 0.61) and VGG19 (from 0.44 to 0.58). Of note, a decreased TPR was observed for ResNet101 in the classification of dermatofibrosis (from 0.26 to 0.16) and intraepithelial carcinoma (from 0.37 to 0.29), and in melanoma for MobileNet (from 0.90 to 0.72).\n\n**Table 3. Classification metrics of HAM10000 low- and high-resolution images with aggregated clinical data using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 77.43 22.57 78.84 21.16  \nResNet50 79.31 20.69 80.72 19.28  \nResNet101 77.55 22.45 78.96 21.04  \nSEResNet50 80.01 19.99 80.72 19.28  \nVGG16 80.25 19.75 81.65 18.35  \nVGG19 79.43 20.57 79.02 20.98  \nEfficientNetB5 75.73 24.27 77.14 22.86  \nMobileNet 81.24 18.76 84.73 15.27  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network; 3 accuracy and error rates are expressed as percentages.\n\nFor high-resolution images, the average global accuracy also increased when clinical features were added to the model (Table 3). The average global accuracy was 80.22% \u00b1 2.30, ranging from 77.14 (EfficientNetB5) to 84.73% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with high-resolution images without aggregated clinical features (p = 0.02; Mann\u2013Whitney U test). This was particularly evident in the case of ResNet50 performance with an increase of 4.75%. The TPR, TNR and geometric mean values are shown in Supplementary Table S4. Major TPR improvements were observed for dermatofibrosis with ResNet34 (26%), ResNet101 (20%), VGG19 (17%) and EfficientNetB5 (28%). TPR increases were also observed for basal cell carcinoma with ResNet50 (12%) and EfficientNetB5 (13%), for vascular lesions with ResNet101 (37%) and EfficientNetB5 (28%), for melanoma with VGG16 (14%), VGG19 (32%) and MobileNet (16%) and intraepithelial carcinoma with VGG16 (24%). Interestingly, the TPR of ResNet50 and SEResNet50 were reduced for dermatofibrosis from 0.55 to 0.20 and from 0.44 to 0.29, respectively. When compared to the DNNs\u2019 performance with low-resolution images and clinical features, no major differences were observed (p = 0.21; Mann\u2013Whitney U test). Of interest, for dermatofibrosis classification, the TPR increased for ResNet34 (35%), ResNet101 (50%), VGG16 (50%) and VGG19 (24%); the exception was ResNet50 with a TPR reduction from 0.44 to 0.2. Altogether, these results indicate that the addition of information related to sex, age and localization of the lesion improves the accuracy of DNNs.\n\n_3.3. Performance across Synthetic Pigmented-Skin Lesion Images_\n\nWe applied an infoGAN to generate synthetic images from the HAM10000 dataset. As shown in Figure 2, the synthetic images seemed realistic and diverse. We then calculated the global accuracy for EfficientNetB5 for the classification of seven pigmented skin lesions. Of the 40 synthetic images analyzed, the network made a single error, so the certainty index was 97.5%; however, this value lacks any relevance since the GAN used for pseudo-labeling deliberately increases the definition limit of each class, inducing an improvement in the certainty of the classifier. Altogether, these data indicate that the synthetic samples are highly realistic and can be used as inputs to train DNNs on pigmented skin lesion classification.\n\n### A B C D\n\n E F G\n\n**Figure 2. Examples of synthetic images generated with the information maximizing generative adversarial network (infoGAN): (A) melanocytic nevus; (B) benign keratosis; (C) vascular lesion; (D) dermatofibroma; (E) intraepithelial carcinoma; (F) basal cell carcinoma; and (G) melanoma.**\n\n_3.4. Performance across General Practitioners with and without Assistance from DNNs Output_\n\nIn the first challenge, 22 general practitioners were asked to classify 162 images without any time constraint. The mean global accuracy and mean error rate were 27.74 and 72.26%, respectively. These results were similar to those previously published for non-dermatologists [9,10]. The best TPR (0.79) was obtained for the melanocytic nevi while the worse metrics were observed for vascular lesions (0.02), dermatofibrosis (0.01) and intraepithelial carcinoma (0.07) (Supplementary Table S5). As EfficientNetB5 was slightly less accurate than the other tested DNNs, we decided to use this framework as a comparator (see Table 1). In the same dataset, this DNN had a mean global accuracy of 78.40% and mean error rate 21.60% (Table 4). Compared to physicians, this was a relevant and significant difference as EfficientNetB5 showed a higher TPR in all disease subtypes (Supplementary Table S5).\n\n**Table 4. Classification metrics of non-dermatologists, general practitioners with or without use of the algorithmic platform and with time constraints.**\n\n**Condition** **Accuracy** **[1]** **Error Rate**\n\nEfficientNetB5 77.14 22.86  \nGPs [2] 17.29 82.71  \nGPs + AI 42.43 57.57  \n\n1 accuracy and error rate are expressed as percentages. 2 GPs: general practitioners; AI: artificial intelligence.\n\nIn the second challenge, 19 general practitioners were asked to classify 35 images with a time constraint of 45 s per image. The global accuracy for this classification was 17.29% (Table 4). In the same dataset, EfficientNetB5 achieved a global accuracy of 77.14%, significantly outperforming physicians. When general practitioners were given the opportunity to access the output of EfficientNetB5 per image, the global accuracy increased to 42.42%. This represented an increase of 25.13%. This result also indicated that, in some cases, physicians did not follow the recommendation of the DNN. Of note, the access of physicians to DNN prediction increased TPR for basal cell carcinoma (from 0.10 to 0.54) and melanoma (from 0.08 to 0.35) (Supplementary Table S6). In contrast, a small decrease in TPR for benign keratosis was observed (see Supplementary Table S6). Altogether, these results show that DNNs have the capability to classify seven different pigmented skin lesions with a level of competence higher to that of the general practitioners participating in these challenges. The access of DNN output by physicians improves their ability to classify pigmented skin lesions, particularly basal cell carcinoma and melanoma.",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions.",
        "abstract": "This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (\n",
        "methods": "**2. Materials and Methods**\n\n_2.1. Pigmented Skin Image Dataset_\n\nThis study used images from the anonymous and annotated HAM10000 dataset publicly available through the International Skin Imaging Collaboration (ISIC) archive [21]. All downloaded images were selected using a random generator from the set of available images in the ISIC archive. We stochastically split the master set of 10,015 dermoscopic images into training (n = 8313; 83%) and test (n = 1702; 17%) datasets that were completely disjoint. Images included a representative collection of all-important diagnostic categories across the seven different types of pigmented lesions as detailed in Tschandl et al. [21]. These included melanocytic nevus, vascular skin lesions (including cherry angiomas, angiokeratomas, pyogenic granulomas and hemorrhages), benign keratoses (including seborrheic keratoses, solar lentigo and lichen-planus-like kertoses), dermatofibroma, intraepithelial carcinoma (including actinic keratoses and Bowen\u2019s disease), basal cell carcinoma and melanoma. Examples of images of each lesion type are depicted in Figure 1. The final composition of each dataset is shown in Table 1.\n\n_2.2. Deep Neural Networks_\n\nWe evaluated eight different DNNs, each characterized by a specific architecture. VGG16 and VGG19 contain 16 and 19 convolutional layers, respectively, with very small receptive fields, five max-pooling layers of size for carrying out spatial pooling, followed by three fully connected layers, with the final layer as the soft-max layer [22]. Rectification nonlinearity (ReLu) activation is applied to all hidden layers. The model also uses dropout regularization in the fully connected layers. ResNet34 is a 34-layer residual network while ResNet50 and ResNet101 are 50- and 101-layers deep, respectively. The architecture of all these DNNs is similar to the one found in VGG consisting mostly of 3 \u00d7 3 filters, however, instead, shortcut connections are inserted resulting into a residual network. SEResNet50 architecture is based on ResNet. A squeeze-and-excitation block is applied at the end of each non-identity branch of residual block [23]. Differently, and instead of increasing its size by adding more or deeper layers, EfficientNetB5 scales up the network width, depth and resolution with a set of fixed scaling coefficients [24]. Finally, MobileNet uses depth-wise separable convolutions which significantly reduce the number of parameters when compared to a network based on standard convolutions and the same depth across the structure in the networks [25]. The framework is 54 layers deep. For each DNN, the initial weights of all layers of the network were set up after pretraining with ImageNet. To assess both the performance of the algorithm and the enhanced training techniques as accurately as possible, we retrained each DNN a total of five times (folds), and each training run consisted of 90 epochs. Training using the curated image patches took approximately 6 h to complete, 45k iterations on a 4 GeForce GTX 1080 GPU configuration. Training accuracy for the curated patches reached maximum accuracy (100%) at around epoch 32, whereas the pretrained model only began to converge around epoch 25. All these DNNs were trained and tested with two different image input size: 300 \u00d7 224 RGB and 600 \u00d7 450 RGB. The low-resolution images were obtained by cropping, distorting and linear resizing the original high-resolution images. DNNs were also trained and tested without or with the clinical features (sex, age and location of the lesion) associated to every image in the HAM10000 database.\n\n_2.3. Image Preprocessing_\n\nWhen a deep convolutional neural network overfits, it works extremely well on training data but poorly on data it has never seen before. This is especially important in the field of dermatology because of the variability that exists in the images that the neural network will be analyzing. Two steps were taken to reduce overfitting. First, a dropout layer was added and set to 0.5. This results in 50% of the neurons to be randomly turned off during the training process and therefore reduce the likelihood of overfitting. The second step taken to reduce overfitting was to use data augmentation. In data augmentation, the images are modified to account for some of the variability that exists in image taking. To account for the grid location, the size of the dermatological manifestation and the angle of the image, the training images fed into the model were altered using zoom (25% probability to be increased between \u22121 and 2%, rotated between \u22126\u00ba and 6\u00ba and vertical/horizontal transfer from \u22122 to 2%), rotation (25% probability of being randomly rotated 90\u00ba clockwise), transposing (15% probability of making a random axial symmetry in one of its diagonals) and horizontal and vertical flipping randomly (50% probability of making a horizontal or vertical \u201cmirror\u201d or both) or following optical parameters from the different types of phone cameras such as shear and brightness (30% probability of contrast modification between \u22128 and 8%) or optical (80% probability of being distorted with a range between \u22126 and 6%) or grid (75% probability at step 4 with a range of \u221228\u201328%) distortions. These modifications were applied to 66% of the input images. The model was run once without data augmentation with dropout and once with data augmentation including drop out in both instances.\n\n_2.4. Generation of Synthetic Pigmented Skin Lesion Images Using an infoGAN_\n\nGenerative adversarial networks (GANs) are a type of generative model that attempt to synthesize novel data that are indistinguishable from the training data [26]. They consist of two neural networks, locked in competition: a generator that captures the data distribution and creates synthesized data (e.g., an image), and a discriminator that estimates the probability that a sample came from the training data rather than from the generator. The two networks are sealed in a zero-sum game, where the success of one corresponds to the failure of the other. The training procedure for the generator is to maximize the probability of the discriminator making an error [27]. Thus, this framework is based on a value function that one model seeks to maximize and the other seeks to minimize. Since the two networks are differentiable, the system generates a gradient that can be used to steer both networks to the right direction. We trained an information maximizing generative adversarial network (infoGAN) composed of a generator and a discriminator, on all high resolution images of the HAM10000 dataset [20]. The infoGAN was adapted to the progressive architecture of the model by splitting the structured code into parts and feeding each part to the network by conditioning activation in the corresponding block of the generator (Supplementary Figure S1). To prevent detrimental competition between the discriminator and generator, and to achieve convergence in an efficient way, we followed the recommendations detailed in Chen et al. [20]. Briefly, the discriminator and generator were composed in eight progressive blocks with an input/output of spatial resolution of 4 \u00d7 4 in the initial step up to 512 \u00d7 512 in step 7. The Batch size has gone dynamically from 128 in step 0 to 2 in the last step. Both the generator and discriminator were optimized using Adam with an initial learning rate of 0.0075 and exponential decay of 0.99 and a Wasserstein function with gradient penalty. The training has progressed in phases of progressive increased resolution. More specifically, the model was capable of generating high resolution images with isolated semantic features controlled by a set of real valued variables. Color, age, sex, localization and type of lesion were the most important semantic features discovered during the training in a completely unsupervised fashion without human input. After training, the generator produced novel images, similar to those in the dataset.\n\n_2.5. Contests among General Practitioners_\n\nTo compare the accuracy of DNNs with non-dermatologist practitioners, we conducted two different challenges. The first aimed to establish the accuracy of general practitioners in classifying images from the HAM10000 dataset without time constraint. For this purpose, a group of 22 general practitioners from any given center in Buenos Aires (Argentina) were given access to 163 images of the different skin pigmented lesions through an anonymous website specifically created for this purpose. Physicians could enter and exit the website without limitation. Alongside the image, recorded factors such as the age, sex and localization of the lesion were shown. All physicians were asked to classify every image within the seven different diagnostics. No incentives were offered for participation. To ensure fair comparisons between the results determined by general practitioners and those determined by the DNNs, the same 162 images were run with the DNN framework which had the worse accuracy metrics in low-resolution and without aggregated clinical features of the eight DNNs tested. To determine if physicians could benefit from access to the algorithmic tool during the own classification task, a second evaluation was conducted. A group of 19 general practitioners that voluntarily accepted to participate in the study was first asked to assess 35 images in a simulated exercise with time constraints (physicians had 45 s to classify every image). In a second step, physicians had access to the predictions of the same algorithmic framework used during the first challenge, the same group having classified each image based on both their criteria and the algorithmic output. For this task, a new set of 35 images was shown with the same time constraints. In both contests, the ethics committee waived ethical approval owing to the use of anonymized dermatologic images obtained from the publicly available HAM10000 dataset.\n\n_2.6. Statistical Analysis_\n\nAfter the model had been trained, a test step was performed in which 1702 images of the seven dermatological manifestations were used as input and the results were statistically analyzed. A confusion matrix was constructed based on comparing the frameworks\u2019 prediction with each of the actual labels. All analyses were performed and programmed via a Jupyter notebook in Python. Sensitivity, specificity, geometric mean, accuracy and error rate were calculated for each dermatological manifestation [28]. Sensitivity or true positive rate (TPR) represented the positive and correctly classified samples to the total number of positive samples. The specificity or true negative rate (TNR) was estimated as the ratio of the correctly classified negative samples to the total number of negative samples. Geometric means were calculated by using the product of TPR and TNR. Accuracy was defined as the ratio between the correctly classified samples to the total number of samples [28]. We also calculated the error rate as the complement of accuracy. All these measures are suitable to evaluate the classification performance based on imbalanced data as found in the HAM10000 database. All metric results were calculated with respect to the class labels documented in the HAM10000 database archive. A summary diagram is depicted on Supplementary Figure S2.",
        "results": "**3. Results**\n\n_3.1. Classification Metrics across Eight Different DNNs_\n\nThe results of global accuracy and error rate for each DNN for the classification of seven pigmented skin lesions at low image resolution (300 \u00d7 224 RGB) are shown in Table 2. The average global accuracy for the eight DNNs reached 76.30% \u00b1 2.79, ranging from 74.05% (EfficientNetB5) to 82.47% (MobileNet). As shown in Supplementary Table S1, the TPR, TNR and geometric mean for each disease subtype varied according to the tested DNN. Almost all DNNs showed the highest TPR for melanocytic nevi classification when compared to the other pigmented lesions, with observed geometric mean values equal or lower than 0.65; interestingly, VGG16, VGG19 and MobileNet also showed high TPR for vascular lesion classification (geometric mean values lower than 0.65). In the case of melanoma and benign keratosis classification, all of them showed a TPR of approximately 0.5, with geometric mean values around 0.75. Likewise, basal cell carcinoma classification showed a TPR of approximately 0.5 and geometric mean values higher than 0.75 using ResNet50, ResNet101, SEResNet50 and EfficientNetB5. Similar results were observed for SEResNet5 and MobileNet and intraepithelial carcinoma classification.\n\n**Table 2. Classification metrics of HAM10000 images at two different resolutions and without aggregated clinical features using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 75.32 24.68 76.73 23.27  \nResNet50 74.56 25.44 75.97 24.03  \nResNet101 75.62 24.38 77.02 22.98  \nSEResNet50 77.82 22.22 79.13 20.87  \nVGG16 76.85 23.15 78.25 21.75  \nVGG19 74.21 25.79 75.62 24.38  \nEfficientNetB5 74.05 25.91 75.50 24.5  \nMobileNet 82.47 17.53 83.88 16.12  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network. 3 accuracy and error rates are expressed as percentages.\n\nUsing higher image resolution (600 \u00d7 450 RGB) improved the global accuracies across all DNNs (Table 2). The global accuracy average for the eight DNNs tested was 77.76 \u00b1 2.77, ranging from 75.50 (EfficientNetB5) to 83.88% (MobileNet). Although higher than the mean average observed with low-resolution images, the difference was not statistically significant (p = 0.07; Mann\u2013Whitney U test). The TPR, TNR and geometric mean values for each disease subtype are detailed in Supplementary Table S2. The highest TPR was observed for melanocytic nevi across all DNNs, and in the case of vascular lesions using VGG16 and MobileNet. On the contrary, the lowest TPR was observed for dermatofibrosis and intraepithelial carcinoma with ResNet34, ResNet50, SEResnet50 and EfficientNetB5. When compared with the low-resolution image parallel cases, the TPR, TNR and geometric mean values were quite similar for most of the disease subtypes. However, a drastic improvement of TPR was observed for dermatofibrosis using ResNet34 (from 0.23 to 0.34), ResNet50 (from 0.29 to 0.55), ResNet101 (from 0.26 to 0.42), VGG16 (from 0.17 to 0.4), VGG19 (from 0.26 to 0.40) and MobileNet (from 0.39 to 0.5). Similar TPR improvements were observed for the intraepithelial carcinoma classification using VGG19 (from 0.26 to 0.40). The average cascade framework runtime of the high-resolution classification model was 21.46+/\u22122.3 milliseconds per image, whereas the low-resolution model required only 18.6+/\u22121.22 milliseconds per image. Altogether, these results indicate that the tested DNNs can classify seven different types of pigmented skin lesions with accuracies higher than 0.7. No major differences in runtime were observed between the cascade framework input with low- or high resolution.\n\n_3.2. Classification Metrics of Different DNNs Aggregating Image and Clinical Features_\n\nWe then investigated if adding clinical features to the analysis could improve the classification accuracy of each DNN. As the HAM10000 dataset provides the sex, age and localization of the skin lesion associated to every image, we gathered these clinical features and aggregated them with the corresponding image to be used as input for each of the tested DNNs. Results are shown in Table 3. For low-resolution images, the addition of clinical data improved the global performance of all DNNs but MobileNet. The average global accuracy for the eight DNNs was 78.86% \u00b1 1.81, ranging from 75.73 (EfficientNetB5) to 81.24% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with low-resolution images without aggregated clinical features (p = 0.004; Mann\u2013Whitney U test). The highest increase was observed for VGG19 raising from 74.21% to 79.43 (5.22%). As shown in Supplementary Table S3, classification improvements were observed in almost all pigmented skin lesions for all DNNs. The highest marginal increases in accuracy were observed for dermatofibrosis where the TPR values for ResNet 50, SEResNet50, VGG16, EfficientNetB5 and MobileNet were raised by 15, 51, 22, 28 and 46%, respectively. Similarly, in the case of the intraepithelial carcinoma condition, TPR values increased 13 and 26% for SEResnet50 and VGG19, respectively, and for basal cell carcinoma classification, increased by 11, 14 and 12% for ResNet50, VGG19 and EfficientNetB5, respectively. For vascular lesions, the TPR values increased by 34, 14 and 29% for ResNet101, VGG19 and EfficientNetB5, respectively. Finally, adding clinical features to skin images also improved the melanoma condition accuracy for VGG16 (from 0.47 to 0.61) and VGG19 (from 0.44 to 0.58). Of note, a decreased TPR was observed for ResNet101 in the classification of dermatofibrosis (from 0.26 to 0.16) and intraepithelial carcinoma (from 0.37 to 0.29), and in melanoma for MobileNet (from 0.90 to 0.72).\n\n**Table 3. Classification metrics of HAM10000 low- and high-resolution images with aggregated clinical data using eight different DNNs.**\n\n**Low-Resolution Images** **[1]** **High-Resolution Images**\n\n**DNN** **[2]** **Accuracy** **[3]** **Error Rate** **Accuracy** **Error Rate**\n\nResNet34 77.43 22.57 78.84 21.16  \nResNet50 79.31 20.69 80.72 19.28  \nResNet101 77.55 22.45 78.96 21.04  \nSEResNet50 80.01 19.99 80.72 19.28  \nVGG16 80.25 19.75 81.65 18.35  \nVGG19 79.43 20.57 79.02 20.98  \nEfficientNetB5 75.73 24.27 77.14 22.86  \nMobileNet 81.24 18.76 84.73 15.27  \n\n1 low-resolution image: 300 \u00d7 224 RGB; high-resolution image: 300 \u00d7 224 RGB. 2 Deep neural network; 3 accuracy and error rates are expressed as percentages.\n\nFor high-resolution images, the average global accuracy also increased when clinical features were added to the model (Table 3). The average global accuracy was 80.22% \u00b1 2.30, ranging from 77.14 (EfficientNetB5) to 84.73% (MobileNet). This represented a statistically significant increase in comparison to the global accuracy of DNNs tested with high-resolution images without aggregated clinical features (p = 0.02; Mann\u2013Whitney U test). This was particularly evident in the case of ResNet50 performance with an increase of 4.75%. The TPR, TNR and geometric mean values are shown in Supplementary Table S4. Major TPR improvements were observed for dermatofibrosis with ResNet34 (26%), ResNet101 (20%), VGG19 (17%) and EfficientNetB5 (28%). TPR increases were also observed for basal cell carcinoma with ResNet50 (12%) and EfficientNetB5 (13%), for vascular lesions with ResNet101 (37%) and EfficientNetB5 (28%), for melanoma with VGG16 (14%), VGG19 (32%) and MobileNet (16%) and intraepithelial carcinoma with VGG16 (24%). Interestingly, the TPR of ResNet50 and SEResNet50 were reduced for dermatofibrosis from 0.55 to 0.20 and from 0.44 to 0.29, respectively. When compared to the DNNs\u2019 performance with low-resolution images and clinical features, no major differences were observed (p = 0.21; Mann\u2013Whitney U test). Of interest, for dermatofibrosis classification, the TPR increased for ResNet34 (35%), ResNet101 (50%), VGG16 (50%) and VGG19 (24%); the exception was ResNet50 with a TPR reduction from 0.44 to 0.2. Altogether, these results indicate that the addition of information related to sex, age and localization of the lesion improves the accuracy of DNNs.\n\n_3.3. Performance across Synthetic Pigmented-Skin Lesion Images_\n\nWe applied an infoGAN to generate synthetic images from the HAM10000 dataset. As shown in Figure 2, the synthetic images seemed realistic and diverse. We then calculated the global accuracy for EfficientNetB5 for the classification of seven pigmented skin lesions. Of the 40 synthetic images analyzed, the network made a single error, so the certainty index was 97.5%; however, this value lacks any relevance since the GAN used for pseudo-labeling deliberately increases the definition limit of each class, inducing an improvement in the certainty of the classifier. Altogether, these data indicate that the synthetic samples are highly realistic and can be used as inputs to train DNNs on pigmented skin lesion classification.\n\n### A B C D\n\n E F G\n\n**Figure 2. Examples of synthetic images generated with the information maximizing generative adversarial network (infoGAN): (A) melanocytic nevus; (B) benign keratosis; (C) vascular lesion; (D) dermatofibroma; (E) intraepithelial carcinoma; (F) basal cell carcinoma; and (G) melanoma.**\n\n_3.4. Performance across General Practitioners with and without Assistance from DNNs Output_\n\nIn the first challenge, 22 general practitioners were asked to classify 162 images without any time constraint. The mean global accuracy and mean error rate were 27.74 and 72.26%, respectively. These results were similar to those previously published for non-dermatologists [9,10]. The best TPR (0.79) was obtained for the melanocytic nevi while the worse metrics were observed for vascular lesions (0.02), dermatofibrosis (0.01) and intraepithelial carcinoma (0.07) (Supplementary Table S5). As EfficientNetB5 was slightly less accurate than the other tested DNNs, we decided to use this framework as a comparator (see Table 1). In the same dataset, this DNN had a mean global accuracy of 78.40% and mean error rate 21.60% (Table 4). Compared to physicians, this was a relevant and significant difference as EfficientNetB5 showed a higher TPR in all disease subtypes (Supplementary Table S5).\n\n**Table 4. Classification metrics of non-dermatologists, general practitioners with or without use of the algorithmic platform and with time constraints.**\n\n**Condition** **Accuracy** **[1]** **Error Rate**\n\nEfficientNetB5 77.14 22.86  \nGPs [2] 17.29 82.71  \nGPs + AI 42.43 57.57  \n\n1 accuracy and error rate are expressed as percentages. 2 GPs: general practitioners; AI: artificial intelligence.\n\nIn the second challenge, 19 general practitioners were asked to classify 35 images with a time constraint of 45 s per image. The global accuracy for this classification was 17.29% (Table 4). In the same dataset, EfficientNetB5 achieved a global accuracy of 77.14%, significantly outperforming physicians. When general practitioners were given the opportunity to access the output of EfficientNetB5 per image, the global accuracy increased to 42.42%. This represented an increase of 25.13%. This result also indicated that, in some cases, physicians did not follow the recommendation of the DNN. Of note, the access of physicians to DNN prediction increased TPR for basal cell carcinoma (from 0.10 to 0.54) and melanoma (from 0.08 to 0.35) (Supplementary Table S6). In contrast, a small decrease in TPR for benign keratosis was observed (see Supplementary Table S6). Altogether, these results show that DNNs have the capability to classify seven different pigmented skin lesions with a level of competence higher to that of the general practitioners participating in these challenges. The access of DNN output by physicians improves their ability to classify pigmented skin lesions, particularly basal cell carcinoma and melanoma.",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Convolutional neural network assistance significantly improves dermatologists' diagnosis of cutaneous tumours using clinical images.",
        "abstract": "## BACKGROUND\nConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation.\n## METHODS\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18\u00a0\u2264\u00a010 years; 5/18\uff1e10 years) in interpreting the 400 clinical images with or without CNN assistance.\n## RESULTS\nThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P\u00a0<\u00a00.001) and kappa (0.74 vs. 0.59, P\u00a0<\u00a00.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P\u00a0<\u00a00.001) and specificity (87.90% vs. 80.92%, P\u00a0<\u00a00.001) of dermatologists with CNN assistance were also significantly improved than those without.\n## CONCLUSIONS\nCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.\n",
        "methods": "## Methods\n\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN\u2019s performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists in interpreting the 400 clinical images with or without CNN assistance.\n\n### 2.1. Dataset\n\nThe study was approved by the institutional review board of the Chinese PLA General Hospital & Medical School (Approval no. 95S2019-123-01). Data on patient demographics and clinical images between 2009 and 2020 were collected via a retrospective chart review, and patient consent was waived by the institutional review board. Basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma (MM) are the most common types of skin cancer in populations. These malignant entities show increasing incidence rates worldwide. And mimicking entities regularly create diagnostic dilemmas in clinical practice settings. Clinical images with the following diagnoses were included: BCC, SCC including keratoacanthoma, MM, Bowen disease (BD), actinic keratosis (AK), melanocytic nevus (MN), seborrhoeic keratosis (SK), haemangioma including pyogenic granuloma, cherry haemangioma, sinusoidal haemangioma and angiokeratoma, dermatofibroma (DF) and wart. All of the diagnoses were based on pathological examinations. The clinical images were taken by cameras (Canon EOS 550D, Nikon D610 and Canon EOS 70D) and smartphones (iPhone 6 s Plus, 7 Plus, 11 Pro, 11 Pro Max and HUAWEI P30, P40, Mate 30Pro).\n\nThe overall data collection was split into training/validation and testing datasets. Of the 29,280 images, 25,773 images were allocated for training and validation. In the remaining 3507 images, 2107 eligible images (excluding 1400 unqualified images) were used for testing. The excluded images contained two components. (1) Blurry and distant images were deleted from the test set but were still used for training. (2) In the test set, each image corresponds to one lesion (no overlap derived from multiple viewpoints) to maintain the diversity of the test set and to ensure that the test set reflected the real performance of the CNN. For a number of the patients, multiple images were from the same lesion, including images taken from different distances and angles. These images were divided into training or test dataset according to the specific lesion to prevent images from the same lesion being used in both training and testing.\n\n### 2.2. Multi-class CNN development\n\nWe trained our CNN based on the EfficientNet-B3 architecture, which was pre-trained on the ImageNet dataset. The model has 27 layers and 12 million parameters, which is suitable for real-time calculation with high accuracy using mobile phones. A total of 25,773 clinical images were used to train the model. An 80/20 training/validation split was conducted to perform cross-validation. The data splitting was conducted according to the lesions to avoid images from the same lesion being used in both training and validation. During training, each image was trimmed to 720 720 pixels to set the lesion at the centre of the image. For each image, we used a 300 300 pixel sliding window to sample ordered, same-sized patches as the CNN input. In the test phase, the CNN produced 10 outputs for a given patch. The image prediction was obtained by averaging all the patches\u2019 probabilities. Colour constancy was performed on all images during training and validation using the Shades of Grey method with Minkowski norm p Z 6.\n\nData augmentation in our study included random rotations and random flips (horizontal and vertical). To address the images unbalanced in different tumours, we trained CNN with weighted loss functions where weights were determined using the inverse frequency of classes in the training data. Specifically, the weight of the class with the largest images (Nmax) was 1.0, and the image number of class k was Nk. Then, the weight of the image from class k was represented as Nmax/Nk. The developed CNN can be accessed at the following URL (http://ai.skinreader.cn/) and has an online prediction function.\n\n### 2.3. Test set enrolment for evaluation\n\nTo make a statistically adequate comparison for each category of cutaneous tumour, with or without CNN assistance, while keeping a feasible total number of clinical images, a total of 400 images (40 images per category) were randomly selected from the test dataset.\n\n### 2.4. Dermatologists\n\nA total of 18 board-certified dermatologists from nine different hospitals participated as readers in this study. Their years of clinical work experience ranged from 1 to 20 years. The majority of them were not experts (experience: 13/18 10 years; 5/18\uff1e10 years). All dermatologists voluntarily participated, understood and agreed with the basic principles and objectives of this study.\n\n### 2.5. Study design\n\nTo evaluate dermatologists\u2019 performance in both assisted and unassisted conditions, the study was designed as a fully crossed, self-control, MRMC study. Dermatologists interpreted all the clinical images in both modalities (with or without CNN assistance), which was separated by a washout period of 2 months. In each modality, dermatologists were only offered one clinical image for their diagnosis and no additional information was supplied. To reduce possible performance differences at the beginning versus the end of interpreting the test set, 400 clinical images were divided into blocks of 40 images with each block containing a roughly similar distribution of tumour categories with a random order. Eighteen dermatologists were randomised into two groups, either of which began with (mode 1) or without (mode 2) CNN assistance first. In either mode, the order and images interpreted were identical; the difference was solely in modality (assisted or not assisted). The dermatologists independently assessed 400 clinical images on a 24\u201d LED monitor (Lenovo Q24i-1L) at a self-controlled pace. When the dermatologists reviewed the images, modalities (with or without CNN assistance) switched every 40 image intervals. If the image had CNN assistance, the top 3 predictions of the CNN appeared in the upper right corner of the corresponding image. If the image was without CNN assistance, only a clinical image was displayed. The participants gave the most likely diagnosis to the corresponding image.\n\n### 2.6. Statistical analysis\n\nThe predicted results of the multi-class CNN on 2107 test images were shown with a confusion matrix. The performance of the CNN model was evaluated by accuracy and kappa coefficient. In the performance evaluation of dermatologists with or without CNN assistance, the primary endpoint was the overall accuracy of dermatologists with pathological diagnosis as the gold standard. The sensitivity and specificity of binary classification was the secondary evaluation endpoint, and kappa consistency analysis was used to evaluate the consistency between dermatologist\u2019s diagnoses (with or without CNN assisted) and pathological results. Dermatologists\u2019 accuracy and kappa with or without assistance were compared by a two-sided paired t test or Wilcoxon signed-rank test, as appropriate. On the binary classification level, the sensitivity and specificity between assisted and unassisted modes were also analysed using a two-sided paired t test. And P < 0.05 was regarded as statistically significant. Statistical analyses were performed in R (version 4.1) software.",
        "results": "Results: The CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without.",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Convolutional neural network assistance significantly improves dermatologists' diagnosis of cutaneous tumours using clinical images.",
        "abstract": "## BACKGROUND\nConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation.\n## METHODS\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18\u00a0\u2264\u00a010 years; 5/18\uff1e10 years) in interpreting the 400 clinical images with or without CNN assistance.\n## RESULTS\nThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P\u00a0<\u00a00.001) and kappa (0.74 vs. 0.59, P\u00a0<\u00a00.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P\u00a0<\u00a00.001) and specificity (87.90% vs. 80.92%, P\u00a0<\u00a00.001) of dermatologists with CNN assistance were also significantly improved than those without.\n## CONCLUSIONS\nCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.\n",
        "methods": "## Methods\n\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN\u2019s performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists in interpreting the 400 clinical images with or without CNN assistance.\n\n### 2.1. Dataset\n\nThe study was approved by the institutional review board of the Chinese PLA General Hospital & Medical School (Approval no. 95S2019-123-01). Data on patient demographics and clinical images between 2009 and 2020 were collected via a retrospective chart review, and patient consent was waived by the institutional review board. Basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma (MM) are the most common types of skin cancer in populations. These malignant entities show increasing incidence rates worldwide. And mimicking entities regularly create diagnostic dilemmas in clinical practice settings. Clinical images with the following diagnoses were included: BCC, SCC including keratoacanthoma, MM, Bowen disease (BD), actinic keratosis (AK), melanocytic nevus (MN), seborrhoeic keratosis (SK), haemangioma including pyogenic granuloma, cherry haemangioma, sinusoidal haemangioma and angiokeratoma, dermatofibroma (DF) and wart. All of the diagnoses were based on pathological examinations. The clinical images were taken by cameras (Canon EOS 550D, Nikon D610 and Canon EOS 70D) and smartphones (iPhone 6 s Plus, 7 Plus, 11 Pro, 11 Pro Max and HUAWEI P30, P40, Mate 30Pro).\n\nThe overall data collection was split into training/validation and testing datasets. Of the 29,280 images, 25,773 images were allocated for training and validation. In the remaining 3507 images, 2107 eligible images (excluding 1400 unqualified images) were used for testing. The excluded images contained two components. (1) Blurry and distant images were deleted from the test set but were still used for training. (2) In the test set, each image corresponds to one lesion (no overlap derived from multiple viewpoints) to maintain the diversity of the test set and to ensure that the test set reflected the real performance of the CNN. For a number of the patients, multiple images were from the same lesion, including images taken from different distances and angles. These images were divided into training or test dataset according to the specific lesion to prevent images from the same lesion being used in both training and testing.\n\n### 2.2. Multi-class CNN development\n\nWe trained our CNN based on the EfficientNet-B3 architecture, which was pre-trained on the ImageNet dataset. The model has 27 layers and 12 million parameters, which is suitable for real-time calculation with high accuracy using mobile phones. A total of 25,773 clinical images were used to train the model. An 80/20 training/validation split was conducted to perform cross-validation. The data splitting was conducted according to the lesions to avoid images from the same lesion being used in both training and validation. During training, each image was trimmed to 720 720 pixels to set the lesion at the centre of the image. For each image, we used a 300 300 pixel sliding window to sample ordered, same-sized patches as the CNN input. In the test phase, the CNN produced 10 outputs for a given patch. The image prediction was obtained by averaging all the patches\u2019 probabilities. Colour constancy was performed on all images during training and validation using the Shades of Grey method with Minkowski norm p Z 6.\n\nData augmentation in our study included random rotations and random flips (horizontal and vertical). To address the images unbalanced in different tumours, we trained CNN with weighted loss functions where weights were determined using the inverse frequency of classes in the training data. Specifically, the weight of the class with the largest images (Nmax) was 1.0, and the image number of class k was Nk. Then, the weight of the image from class k was represented as Nmax/Nk. The developed CNN can be accessed at the following URL (http://ai.skinreader.cn/) and has an online prediction function.\n\n### 2.3. Test set enrolment for evaluation\n\nTo make a statistically adequate comparison for each category of cutaneous tumour, with or without CNN assistance, while keeping a feasible total number of clinical images, a total of 400 images (40 images per category) were randomly selected from the test dataset.\n\n### 2.4. Dermatologists\n\nA total of 18 board-certified dermatologists from nine different hospitals participated as readers in this study. Their years of clinical work experience ranged from 1 to 20 years. The majority of them were not experts (experience: 13/18 10 years; 5/18\uff1e10 years). All dermatologists voluntarily participated, understood and agreed with the basic principles and objectives of this study.\n\n### 2.5. Study design\n\nTo evaluate dermatologists\u2019 performance in both assisted and unassisted conditions, the study was designed as a fully crossed, self-control, MRMC study. Dermatologists interpreted all the clinical images in both modalities (with or without CNN assistance), which was separated by a washout period of 2 months. In each modality, dermatologists were only offered one clinical image for their diagnosis and no additional information was supplied. To reduce possible performance differences at the beginning versus the end of interpreting the test set, 400 clinical images were divided into blocks of 40 images with each block containing a roughly similar distribution of tumour categories with a random order. Eighteen dermatologists were randomised into two groups, either of which began with (mode 1) or without (mode 2) CNN assistance first. In either mode, the order and images interpreted were identical; the difference was solely in modality (assisted or not assisted). The dermatologists independently assessed 400 clinical images on a 24\u201d LED monitor (Lenovo Q24i-1L) at a self-controlled pace. When the dermatologists reviewed the images, modalities (with or without CNN assistance) switched every 40 image intervals. If the image had CNN assistance, the top 3 predictions of the CNN appeared in the upper right corner of the corresponding image. If the image was without CNN assistance, only a clinical image was displayed. The participants gave the most likely diagnosis to the corresponding image.\n\n### 2.6. Statistical analysis\n\nThe predicted results of the multi-class CNN on 2107 test images were shown with a confusion matrix. The performance of the CNN model was evaluated by accuracy and kappa coefficient. In the performance evaluation of dermatologists with or without CNN assistance, the primary endpoint was the overall accuracy of dermatologists with pathological diagnosis as the gold standard. The sensitivity and specificity of binary classification was the secondary evaluation endpoint, and kappa consistency analysis was used to evaluate the consistency between dermatologist\u2019s diagnoses (with or without CNN assisted) and pathological results. Dermatologists\u2019 accuracy and kappa with or without assistance were compared by a two-sided paired t test or Wilcoxon signed-rank test, as appropriate. On the binary classification level, the sensitivity and specificity between assisted and unassisted modes were also analysed using a two-sided paired t test. And P < 0.05 was regarded as statistically significant. Statistical analyses were performed in R (version 4.1) software.",
        "results": "Results: The CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without.",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Convolutional neural network assistance significantly improves dermatologists' diagnosis of cutaneous tumours using clinical images.",
        "abstract": "## BACKGROUND\nConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation.\n## METHODS\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18\u00a0\u2264\u00a010 years; 5/18\uff1e10 years) in interpreting the 400 clinical images with or without CNN assistance.\n## RESULTS\nThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P\u00a0<\u00a00.001) and kappa (0.74 vs. 0.59, P\u00a0<\u00a00.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P\u00a0<\u00a00.001) and specificity (87.90% vs. 80.92%, P\u00a0<\u00a00.001) of dermatologists with CNN assistance were also significantly improved than those without.\n## CONCLUSIONS\nCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.\n",
        "methods": "## Methods\n\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN\u2019s performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists in interpreting the 400 clinical images with or without CNN assistance.\n\n### 2.1. Dataset\n\nThe study was approved by the institutional review board of the Chinese PLA General Hospital & Medical School (Approval no. 95S2019-123-01). Data on patient demographics and clinical images between 2009 and 2020 were collected via a retrospective chart review, and patient consent was waived by the institutional review board. Basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma (MM) are the most common types of skin cancer in populations. These malignant entities show increasing incidence rates worldwide. And mimicking entities regularly create diagnostic dilemmas in clinical practice settings. Clinical images with the following diagnoses were included: BCC, SCC including keratoacanthoma, MM, Bowen disease (BD), actinic keratosis (AK), melanocytic nevus (MN), seborrhoeic keratosis (SK), haemangioma including pyogenic granuloma, cherry haemangioma, sinusoidal haemangioma and angiokeratoma, dermatofibroma (DF) and wart. All of the diagnoses were based on pathological examinations. The clinical images were taken by cameras (Canon EOS 550D, Nikon D610 and Canon EOS 70D) and smartphones (iPhone 6 s Plus, 7 Plus, 11 Pro, 11 Pro Max and HUAWEI P30, P40, Mate 30Pro).\n\nThe overall data collection was split into training/validation and testing datasets. Of the 29,280 images, 25,773 images were allocated for training and validation. In the remaining 3507 images, 2107 eligible images (excluding 1400 unqualified images) were used for testing. The excluded images contained two components. (1) Blurry and distant images were deleted from the test set but were still used for training. (2) In the test set, each image corresponds to one lesion (no overlap derived from multiple viewpoints) to maintain the diversity of the test set and to ensure that the test set reflected the real performance of the CNN. For a number of the patients, multiple images were from the same lesion, including images taken from different distances and angles. These images were divided into training or test dataset according to the specific lesion to prevent images from the same lesion being used in both training and testing.\n\n### 2.2. Multi-class CNN development\n\nWe trained our CNN based on the EfficientNet-B3 architecture, which was pre-trained on the ImageNet dataset. The model has 27 layers and 12 million parameters, which is suitable for real-time calculation with high accuracy using mobile phones. A total of 25,773 clinical images were used to train the model. An 80/20 training/validation split was conducted to perform cross-validation. The data splitting was conducted according to the lesions to avoid images from the same lesion being used in both training and validation. During training, each image was trimmed to 720 720 pixels to set the lesion at the centre of the image. For each image, we used a 300 300 pixel sliding window to sample ordered, same-sized patches as the CNN input. In the test phase, the CNN produced 10 outputs for a given patch. The image prediction was obtained by averaging all the patches\u2019 probabilities. Colour constancy was performed on all images during training and validation using the Shades of Grey method with Minkowski norm p Z 6.\n\nData augmentation in our study included random rotations and random flips (horizontal and vertical). To address the images unbalanced in different tumours, we trained CNN with weighted loss functions where weights were determined using the inverse frequency of classes in the training data. Specifically, the weight of the class with the largest images (Nmax) was 1.0, and the image number of class k was Nk. Then, the weight of the image from class k was represented as Nmax/Nk. The developed CNN can be accessed at the following URL (http://ai.skinreader.cn/) and has an online prediction function.\n\n### 2.3. Test set enrolment for evaluation\n\nTo make a statistically adequate comparison for each category of cutaneous tumour, with or without CNN assistance, while keeping a feasible total number of clinical images, a total of 400 images (40 images per category) were randomly selected from the test dataset.\n\n### 2.4. Dermatologists\n\nA total of 18 board-certified dermatologists from nine different hospitals participated as readers in this study. Their years of clinical work experience ranged from 1 to 20 years. The majority of them were not experts (experience: 13/18 10 years; 5/18\uff1e10 years). All dermatologists voluntarily participated, understood and agreed with the basic principles and objectives of this study.\n\n### 2.5. Study design\n\nTo evaluate dermatologists\u2019 performance in both assisted and unassisted conditions, the study was designed as a fully crossed, self-control, MRMC study. Dermatologists interpreted all the clinical images in both modalities (with or without CNN assistance), which was separated by a washout period of 2 months. In each modality, dermatologists were only offered one clinical image for their diagnosis and no additional information was supplied. To reduce possible performance differences at the beginning versus the end of interpreting the test set, 400 clinical images were divided into blocks of 40 images with each block containing a roughly similar distribution of tumour categories with a random order. Eighteen dermatologists were randomised into two groups, either of which began with (mode 1) or without (mode 2) CNN assistance first. In either mode, the order and images interpreted were identical; the difference was solely in modality (assisted or not assisted). The dermatologists independently assessed 400 clinical images on a 24\u201d LED monitor (Lenovo Q24i-1L) at a self-controlled pace. When the dermatologists reviewed the images, modalities (with or without CNN assistance) switched every 40 image intervals. If the image had CNN assistance, the top 3 predictions of the CNN appeared in the upper right corner of the corresponding image. If the image was without CNN assistance, only a clinical image was displayed. The participants gave the most likely diagnosis to the corresponding image.\n\n### 2.6. Statistical analysis\n\nThe predicted results of the multi-class CNN on 2107 test images were shown with a confusion matrix. The performance of the CNN model was evaluated by accuracy and kappa coefficient. In the performance evaluation of dermatologists with or without CNN assistance, the primary endpoint was the overall accuracy of dermatologists with pathological diagnosis as the gold standard. The sensitivity and specificity of binary classification was the secondary evaluation endpoint, and kappa consistency analysis was used to evaluate the consistency between dermatologist\u2019s diagnoses (with or without CNN assisted) and pathological results. Dermatologists\u2019 accuracy and kappa with or without assistance were compared by a two-sided paired t test or Wilcoxon signed-rank test, as appropriate. On the binary classification level, the sensitivity and specificity between assisted and unassisted modes were also analysed using a two-sided paired t test. And P < 0.05 was regarded as statistically significant. Statistical analyses were performed in R (version 4.1) software.",
        "results": "Results: The CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without.",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Convolutional neural network assistance significantly improves dermatologists' diagnosis of cutaneous tumours using clinical images.",
        "abstract": "## BACKGROUND\nConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation.\n## METHODS\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18\u00a0\u2264\u00a010 years; 5/18\uff1e10 years) in interpreting the 400 clinical images with or without CNN assistance.\n## RESULTS\nThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P\u00a0<\u00a00.001) and kappa (0.74 vs. 0.59, P\u00a0<\u00a00.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P\u00a0<\u00a00.001) and specificity (87.90% vs. 80.92%, P\u00a0<\u00a00.001) of dermatologists with CNN assistance were also significantly improved than those without.\n## CONCLUSIONS\nCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.\n",
        "methods": "## Methods\n\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN\u2019s performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists in interpreting the 400 clinical images with or without CNN assistance.\n\n### 2.1. Dataset\n\nThe study was approved by the institutional review board of the Chinese PLA General Hospital & Medical School (Approval no. 95S2019-123-01). Data on patient demographics and clinical images between 2009 and 2020 were collected via a retrospective chart review, and patient consent was waived by the institutional review board. Basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma (MM) are the most common types of skin cancer in populations. These malignant entities show increasing incidence rates worldwide. And mimicking entities regularly create diagnostic dilemmas in clinical practice settings. Clinical images with the following diagnoses were included: BCC, SCC including keratoacanthoma, MM, Bowen disease (BD), actinic keratosis (AK), melanocytic nevus (MN), seborrhoeic keratosis (SK), haemangioma including pyogenic granuloma, cherry haemangioma, sinusoidal haemangioma and angiokeratoma, dermatofibroma (DF) and wart. All of the diagnoses were based on pathological examinations. The clinical images were taken by cameras (Canon EOS 550D, Nikon D610 and Canon EOS 70D) and smartphones (iPhone 6 s Plus, 7 Plus, 11 Pro, 11 Pro Max and HUAWEI P30, P40, Mate 30Pro).\n\nThe overall data collection was split into training/validation and testing datasets. Of the 29,280 images, 25,773 images were allocated for training and validation. In the remaining 3507 images, 2107 eligible images (excluding 1400 unqualified images) were used for testing. The excluded images contained two components. (1) Blurry and distant images were deleted from the test set but were still used for training. (2) In the test set, each image corresponds to one lesion (no overlap derived from multiple viewpoints) to maintain the diversity of the test set and to ensure that the test set reflected the real performance of the CNN. For a number of the patients, multiple images were from the same lesion, including images taken from different distances and angles. These images were divided into training or test dataset according to the specific lesion to prevent images from the same lesion being used in both training and testing.\n\n### 2.2. Multi-class CNN development\n\nWe trained our CNN based on the EfficientNet-B3 architecture, which was pre-trained on the ImageNet dataset. The model has 27 layers and 12 million parameters, which is suitable for real-time calculation with high accuracy using mobile phones. A total of 25,773 clinical images were used to train the model. An 80/20 training/validation split was conducted to perform cross-validation. The data splitting was conducted according to the lesions to avoid images from the same lesion being used in both training and validation. During training, each image was trimmed to 720 720 pixels to set the lesion at the centre of the image. For each image, we used a 300 300 pixel sliding window to sample ordered, same-sized patches as the CNN input. In the test phase, the CNN produced 10 outputs for a given patch. The image prediction was obtained by averaging all the patches\u2019 probabilities. Colour constancy was performed on all images during training and validation using the Shades of Grey method with Minkowski norm p Z 6.\n\nData augmentation in our study included random rotations and random flips (horizontal and vertical). To address the images unbalanced in different tumours, we trained CNN with weighted loss functions where weights were determined using the inverse frequency of classes in the training data. Specifically, the weight of the class with the largest images (Nmax) was 1.0, and the image number of class k was Nk. Then, the weight of the image from class k was represented as Nmax/Nk. The developed CNN can be accessed at the following URL (http://ai.skinreader.cn/) and has an online prediction function.\n\n### 2.3. Test set enrolment for evaluation\n\nTo make a statistically adequate comparison for each category of cutaneous tumour, with or without CNN assistance, while keeping a feasible total number of clinical images, a total of 400 images (40 images per category) were randomly selected from the test dataset.\n\n### 2.4. Dermatologists\n\nA total of 18 board-certified dermatologists from nine different hospitals participated as readers in this study. Their years of clinical work experience ranged from 1 to 20 years. The majority of them were not experts (experience: 13/18 10 years; 5/18\uff1e10 years). All dermatologists voluntarily participated, understood and agreed with the basic principles and objectives of this study.\n\n### 2.5. Study design\n\nTo evaluate dermatologists\u2019 performance in both assisted and unassisted conditions, the study was designed as a fully crossed, self-control, MRMC study. Dermatologists interpreted all the clinical images in both modalities (with or without CNN assistance), which was separated by a washout period of 2 months. In each modality, dermatologists were only offered one clinical image for their diagnosis and no additional information was supplied. To reduce possible performance differences at the beginning versus the end of interpreting the test set, 400 clinical images were divided into blocks of 40 images with each block containing a roughly similar distribution of tumour categories with a random order. Eighteen dermatologists were randomised into two groups, either of which began with (mode 1) or without (mode 2) CNN assistance first. In either mode, the order and images interpreted were identical; the difference was solely in modality (assisted or not assisted). The dermatologists independently assessed 400 clinical images on a 24\u201d LED monitor (Lenovo Q24i-1L) at a self-controlled pace. When the dermatologists reviewed the images, modalities (with or without CNN assistance) switched every 40 image intervals. If the image had CNN assistance, the top 3 predictions of the CNN appeared in the upper right corner of the corresponding image. If the image was without CNN assistance, only a clinical image was displayed. The participants gave the most likely diagnosis to the corresponding image.\n\n### 2.6. Statistical analysis\n\nThe predicted results of the multi-class CNN on 2107 test images were shown with a confusion matrix. The performance of the CNN model was evaluated by accuracy and kappa coefficient. In the performance evaluation of dermatologists with or without CNN assistance, the primary endpoint was the overall accuracy of dermatologists with pathological diagnosis as the gold standard. The sensitivity and specificity of binary classification was the secondary evaluation endpoint, and kappa consistency analysis was used to evaluate the consistency between dermatologist\u2019s diagnoses (with or without CNN assisted) and pathological results. Dermatologists\u2019 accuracy and kappa with or without assistance were compared by a two-sided paired t test or Wilcoxon signed-rank test, as appropriate. On the binary classification level, the sensitivity and specificity between assisted and unassisted modes were also analysed using a two-sided paired t test. And P < 0.05 was regarded as statistically significant. Statistical analyses were performed in R (version 4.1) software.",
        "results": "Results: The CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without.",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Convolutional neural network assistance significantly improves dermatologists' diagnosis of cutaneous tumours using clinical images.",
        "abstract": "## BACKGROUND\nConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation.\n## METHODS\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18\u00a0\u2264\u00a010 years; 5/18\uff1e10 years) in interpreting the 400 clinical images with or without CNN assistance.\n## RESULTS\nThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P\u00a0<\u00a00.001) and kappa (0.74 vs. 0.59, P\u00a0<\u00a00.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P\u00a0<\u00a00.001) and specificity (87.90% vs. 80.92%, P\u00a0<\u00a00.001) of dermatologists with CNN assistance were also significantly improved than those without.\n## CONCLUSIONS\nCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.\n",
        "methods": "## Methods\n\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN\u2019s performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists in interpreting the 400 clinical images with or without CNN assistance.\n\n### 2.1. Dataset\n\nThe study was approved by the institutional review board of the Chinese PLA General Hospital & Medical School (Approval no. 95S2019-123-01). Data on patient demographics and clinical images between 2009 and 2020 were collected via a retrospective chart review, and patient consent was waived by the institutional review board. Basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma (MM) are the most common types of skin cancer in populations. These malignant entities show increasing incidence rates worldwide. And mimicking entities regularly create diagnostic dilemmas in clinical practice settings. Clinical images with the following diagnoses were included: BCC, SCC including keratoacanthoma, MM, Bowen disease (BD), actinic keratosis (AK), melanocytic nevus (MN), seborrhoeic keratosis (SK), haemangioma including pyogenic granuloma, cherry haemangioma, sinusoidal haemangioma and angiokeratoma, dermatofibroma (DF) and wart. All of the diagnoses were based on pathological examinations. The clinical images were taken by cameras (Canon EOS 550D, Nikon D610 and Canon EOS 70D) and smartphones (iPhone 6 s Plus, 7 Plus, 11 Pro, 11 Pro Max and HUAWEI P30, P40, Mate 30Pro).\n\nThe overall data collection was split into training/validation and testing datasets. Of the 29,280 images, 25,773 images were allocated for training and validation. In the remaining 3507 images, 2107 eligible images (excluding 1400 unqualified images) were used for testing. The excluded images contained two components. (1) Blurry and distant images were deleted from the test set but were still used for training. (2) In the test set, each image corresponds to one lesion (no overlap derived from multiple viewpoints) to maintain the diversity of the test set and to ensure that the test set reflected the real performance of the CNN. For a number of the patients, multiple images were from the same lesion, including images taken from different distances and angles. These images were divided into training or test dataset according to the specific lesion to prevent images from the same lesion being used in both training and testing.\n\n### 2.2. Multi-class CNN development\n\nWe trained our CNN based on the EfficientNet-B3 architecture, which was pre-trained on the ImageNet dataset. The model has 27 layers and 12 million parameters, which is suitable for real-time calculation with high accuracy using mobile phones. A total of 25,773 clinical images were used to train the model. An 80/20 training/validation split was conducted to perform cross-validation. The data splitting was conducted according to the lesions to avoid images from the same lesion being used in both training and validation. During training, each image was trimmed to 720 720 pixels to set the lesion at the centre of the image. For each image, we used a 300 300 pixel sliding window to sample ordered, same-sized patches as the CNN input. In the test phase, the CNN produced 10 outputs for a given patch. The image prediction was obtained by averaging all the patches\u2019 probabilities. Colour constancy was performed on all images during training and validation using the Shades of Grey method with Minkowski norm p Z 6.\n\nData augmentation in our study included random rotations and random flips (horizontal and vertical). To address the images unbalanced in different tumours, we trained CNN with weighted loss functions where weights were determined using the inverse frequency of classes in the training data. Specifically, the weight of the class with the largest images (Nmax) was 1.0, and the image number of class k was Nk. Then, the weight of the image from class k was represented as Nmax/Nk. The developed CNN can be accessed at the following URL (http://ai.skinreader.cn/) and has an online prediction function.\n\n### 2.3. Test set enrolment for evaluation\n\nTo make a statistically adequate comparison for each category of cutaneous tumour, with or without CNN assistance, while keeping a feasible total number of clinical images, a total of 400 images (40 images per category) were randomly selected from the test dataset.\n\n### 2.4. Dermatologists\n\nA total of 18 board-certified dermatologists from nine different hospitals participated as readers in this study. Their years of clinical work experience ranged from 1 to 20 years. The majority of them were not experts (experience: 13/18 10 years; 5/18\uff1e10 years). All dermatologists voluntarily participated, understood and agreed with the basic principles and objectives of this study.\n\n### 2.5. Study design\n\nTo evaluate dermatologists\u2019 performance in both assisted and unassisted conditions, the study was designed as a fully crossed, self-control, MRMC study. Dermatologists interpreted all the clinical images in both modalities (with or without CNN assistance), which was separated by a washout period of 2 months. In each modality, dermatologists were only offered one clinical image for their diagnosis and no additional information was supplied. To reduce possible performance differences at the beginning versus the end of interpreting the test set, 400 clinical images were divided into blocks of 40 images with each block containing a roughly similar distribution of tumour categories with a random order. Eighteen dermatologists were randomised into two groups, either of which began with (mode 1) or without (mode 2) CNN assistance first. In either mode, the order and images interpreted were identical; the difference was solely in modality (assisted or not assisted). The dermatologists independently assessed 400 clinical images on a 24\u201d LED monitor (Lenovo Q24i-1L) at a self-controlled pace. When the dermatologists reviewed the images, modalities (with or without CNN assistance) switched every 40 image intervals. If the image had CNN assistance, the top 3 predictions of the CNN appeared in the upper right corner of the corresponding image. If the image was without CNN assistance, only a clinical image was displayed. The participants gave the most likely diagnosis to the corresponding image.\n\n### 2.6. Statistical analysis\n\nThe predicted results of the multi-class CNN on 2107 test images were shown with a confusion matrix. The performance of the CNN model was evaluated by accuracy and kappa coefficient. In the performance evaluation of dermatologists with or without CNN assistance, the primary endpoint was the overall accuracy of dermatologists with pathological diagnosis as the gold standard. The sensitivity and specificity of binary classification was the secondary evaluation endpoint, and kappa consistency analysis was used to evaluate the consistency between dermatologist\u2019s diagnoses (with or without CNN assisted) and pathological results. Dermatologists\u2019 accuracy and kappa with or without assistance were compared by a two-sided paired t test or Wilcoxon signed-rank test, as appropriate. On the binary classification level, the sensitivity and specificity between assisted and unassisted modes were also analysed using a two-sided paired t test. And P < 0.05 was regarded as statistically significant. Statistical analyses were performed in R (version 4.1) software.",
        "results": "Results: The CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without.",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Convolutional neural network assistance significantly improves dermatologists' diagnosis of cutaneous tumours using clinical images.",
        "abstract": "## BACKGROUND\nConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation.\n## METHODS\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18\u00a0\u2264\u00a010 years; 5/18\uff1e10 years) in interpreting the 400 clinical images with or without CNN assistance.\n## RESULTS\nThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P\u00a0<\u00a00.001) and kappa (0.74 vs. 0.59, P\u00a0<\u00a00.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P\u00a0<\u00a00.001) and specificity (87.90% vs. 80.92%, P\u00a0<\u00a00.001) of dermatologists with CNN assistance were also significantly improved than those without.\n## CONCLUSIONS\nCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.\n",
        "methods": "## Methods\n\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN\u2019s performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists in interpreting the 400 clinical images with or without CNN assistance.\n\n### 2.1. Dataset\n\nThe study was approved by the institutional review board of the Chinese PLA General Hospital & Medical School (Approval no. 95S2019-123-01). Data on patient demographics and clinical images between 2009 and 2020 were collected via a retrospective chart review, and patient consent was waived by the institutional review board. Basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma (MM) are the most common types of skin cancer in populations. These malignant entities show increasing incidence rates worldwide. And mimicking entities regularly create diagnostic dilemmas in clinical practice settings. Clinical images with the following diagnoses were included: BCC, SCC including keratoacanthoma, MM, Bowen disease (BD), actinic keratosis (AK), melanocytic nevus (MN), seborrhoeic keratosis (SK), haemangioma including pyogenic granuloma, cherry haemangioma, sinusoidal haemangioma and angiokeratoma, dermatofibroma (DF) and wart. All of the diagnoses were based on pathological examinations. The clinical images were taken by cameras (Canon EOS 550D, Nikon D610 and Canon EOS 70D) and smartphones (iPhone 6 s Plus, 7 Plus, 11 Pro, 11 Pro Max and HUAWEI P30, P40, Mate 30Pro).\n\nThe overall data collection was split into training/validation and testing datasets. Of the 29,280 images, 25,773 images were allocated for training and validation. In the remaining 3507 images, 2107 eligible images (excluding 1400 unqualified images) were used for testing. The excluded images contained two components. (1) Blurry and distant images were deleted from the test set but were still used for training. (2) In the test set, each image corresponds to one lesion (no overlap derived from multiple viewpoints) to maintain the diversity of the test set and to ensure that the test set reflected the real performance of the CNN. For a number of the patients, multiple images were from the same lesion, including images taken from different distances and angles. These images were divided into training or test dataset according to the specific lesion to prevent images from the same lesion being used in both training and testing.\n\n### 2.2. Multi-class CNN development\n\nWe trained our CNN based on the EfficientNet-B3 architecture, which was pre-trained on the ImageNet dataset. The model has 27 layers and 12 million parameters, which is suitable for real-time calculation with high accuracy using mobile phones. A total of 25,773 clinical images were used to train the model. An 80/20 training/validation split was conducted to perform cross-validation. The data splitting was conducted according to the lesions to avoid images from the same lesion being used in both training and validation. During training, each image was trimmed to 720 720 pixels to set the lesion at the centre of the image. For each image, we used a 300 300 pixel sliding window to sample ordered, same-sized patches as the CNN input. In the test phase, the CNN produced 10 outputs for a given patch. The image prediction was obtained by averaging all the patches\u2019 probabilities. Colour constancy was performed on all images during training and validation using the Shades of Grey method with Minkowski norm p Z 6.\n\nData augmentation in our study included random rotations and random flips (horizontal and vertical). To address the images unbalanced in different tumours, we trained CNN with weighted loss functions where weights were determined using the inverse frequency of classes in the training data. Specifically, the weight of the class with the largest images (Nmax) was 1.0, and the image number of class k was Nk. Then, the weight of the image from class k was represented as Nmax/Nk. The developed CNN can be accessed at the following URL (http://ai.skinreader.cn/) and has an online prediction function.\n\n### 2.3. Test set enrolment for evaluation\n\nTo make a statistically adequate comparison for each category of cutaneous tumour, with or without CNN assistance, while keeping a feasible total number of clinical images, a total of 400 images (40 images per category) were randomly selected from the test dataset.\n\n### 2.4. Dermatologists\n\nA total of 18 board-certified dermatologists from nine different hospitals participated as readers in this study. Their years of clinical work experience ranged from 1 to 20 years. The majority of them were not experts (experience: 13/18 10 years; 5/18\uff1e10 years). All dermatologists voluntarily participated, understood and agreed with the basic principles and objectives of this study.\n\n### 2.5. Study design\n\nTo evaluate dermatologists\u2019 performance in both assisted and unassisted conditions, the study was designed as a fully crossed, self-control, MRMC study. Dermatologists interpreted all the clinical images in both modalities (with or without CNN assistance), which was separated by a washout period of 2 months. In each modality, dermatologists were only offered one clinical image for their diagnosis and no additional information was supplied. To reduce possible performance differences at the beginning versus the end of interpreting the test set, 400 clinical images were divided into blocks of 40 images with each block containing a roughly similar distribution of tumour categories with a random order. Eighteen dermatologists were randomised into two groups, either of which began with (mode 1) or without (mode 2) CNN assistance first. In either mode, the order and images interpreted were identical; the difference was solely in modality (assisted or not assisted). The dermatologists independently assessed 400 clinical images on a 24\u201d LED monitor (Lenovo Q24i-1L) at a self-controlled pace. When the dermatologists reviewed the images, modalities (with or without CNN assistance) switched every 40 image intervals. If the image had CNN assistance, the top 3 predictions of the CNN appeared in the upper right corner of the corresponding image. If the image was without CNN assistance, only a clinical image was displayed. The participants gave the most likely diagnosis to the corresponding image.\n\n### 2.6. Statistical analysis\n\nThe predicted results of the multi-class CNN on 2107 test images were shown with a confusion matrix. The performance of the CNN model was evaluated by accuracy and kappa coefficient. In the performance evaluation of dermatologists with or without CNN assistance, the primary endpoint was the overall accuracy of dermatologists with pathological diagnosis as the gold standard. The sensitivity and specificity of binary classification was the secondary evaluation endpoint, and kappa consistency analysis was used to evaluate the consistency between dermatologist\u2019s diagnoses (with or without CNN assisted) and pathological results. Dermatologists\u2019 accuracy and kappa with or without assistance were compared by a two-sided paired t test or Wilcoxon signed-rank test, as appropriate. On the binary classification level, the sensitivity and specificity between assisted and unassisted modes were also analysed using a two-sided paired t test. And P < 0.05 was regarded as statistically significant. Statistical analyses were performed in R (version 4.1) software.",
        "results": "Results: The CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without.",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Convolutional neural network assistance significantly improves dermatologists' diagnosis of cutaneous tumours using clinical images.",
        "abstract": "## BACKGROUND\nConvolutional neural networks (CNNs) have demonstrated expert-level performance in cutaneous tumour classification using clinical images, but most previous studies have focused on dermatologist-versus-CNN comparisons rather than their combination. The objective of our study was to evaluate the potential impact of CNN assistance on dermatologists for clinical image interpretation.\n## METHODS\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN's performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists (experience: 13/18\u00a0\u2264\u00a010 years; 5/18\uff1e10 years) in interpreting the 400 clinical images with or without CNN assistance.\n## RESULTS\nThe CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P\u00a0<\u00a00.001) and kappa (0.74 vs. 0.59, P\u00a0<\u00a00.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P\u00a0<\u00a00.001) and specificity (87.90% vs. 80.92%, P\u00a0<\u00a00.001) of dermatologists with CNN assistance were also significantly improved than those without.\n## CONCLUSIONS\nCNN assistance improved dermatologist accuracy in interpreting cutaneous tumours and could further boost the acceptance of this new technique.\n",
        "methods": "## Methods\n\nA multi-class CNN was trained and validated using a dataset of 25,773 clinical images comprising 10 categories of cutaneous tumours. The CNN\u2019s performance was tested on an independent dataset of 2107 images. A total of 400 images (40 per category) were randomly selected from the test dataset. A fully crossed, self-control, multi-reader multi-case (MRMC) study was conducted to compare the performance of 18 board-certified dermatologists in interpreting the 400 clinical images with or without CNN assistance.\n\n### 2.1. Dataset\n\nThe study was approved by the institutional review board of the Chinese PLA General Hospital & Medical School (Approval no. 95S2019-123-01). Data on patient demographics and clinical images between 2009 and 2020 were collected via a retrospective chart review, and patient consent was waived by the institutional review board. Basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma (MM) are the most common types of skin cancer in populations. These malignant entities show increasing incidence rates worldwide. And mimicking entities regularly create diagnostic dilemmas in clinical practice settings. Clinical images with the following diagnoses were included: BCC, SCC including keratoacanthoma, MM, Bowen disease (BD), actinic keratosis (AK), melanocytic nevus (MN), seborrhoeic keratosis (SK), haemangioma including pyogenic granuloma, cherry haemangioma, sinusoidal haemangioma and angiokeratoma, dermatofibroma (DF) and wart. All of the diagnoses were based on pathological examinations. The clinical images were taken by cameras (Canon EOS 550D, Nikon D610 and Canon EOS 70D) and smartphones (iPhone 6 s Plus, 7 Plus, 11 Pro, 11 Pro Max and HUAWEI P30, P40, Mate 30Pro).\n\nThe overall data collection was split into training/validation and testing datasets. Of the 29,280 images, 25,773 images were allocated for training and validation. In the remaining 3507 images, 2107 eligible images (excluding 1400 unqualified images) were used for testing. The excluded images contained two components. (1) Blurry and distant images were deleted from the test set but were still used for training. (2) In the test set, each image corresponds to one lesion (no overlap derived from multiple viewpoints) to maintain the diversity of the test set and to ensure that the test set reflected the real performance of the CNN. For a number of the patients, multiple images were from the same lesion, including images taken from different distances and angles. These images were divided into training or test dataset according to the specific lesion to prevent images from the same lesion being used in both training and testing.\n\n### 2.2. Multi-class CNN development\n\nWe trained our CNN based on the EfficientNet-B3 architecture, which was pre-trained on the ImageNet dataset. The model has 27 layers and 12 million parameters, which is suitable for real-time calculation with high accuracy using mobile phones. A total of 25,773 clinical images were used to train the model. An 80/20 training/validation split was conducted to perform cross-validation. The data splitting was conducted according to the lesions to avoid images from the same lesion being used in both training and validation. During training, each image was trimmed to 720 720 pixels to set the lesion at the centre of the image. For each image, we used a 300 300 pixel sliding window to sample ordered, same-sized patches as the CNN input. In the test phase, the CNN produced 10 outputs for a given patch. The image prediction was obtained by averaging all the patches\u2019 probabilities. Colour constancy was performed on all images during training and validation using the Shades of Grey method with Minkowski norm p Z 6.\n\nData augmentation in our study included random rotations and random flips (horizontal and vertical). To address the images unbalanced in different tumours, we trained CNN with weighted loss functions where weights were determined using the inverse frequency of classes in the training data. Specifically, the weight of the class with the largest images (Nmax) was 1.0, and the image number of class k was Nk. Then, the weight of the image from class k was represented as Nmax/Nk. The developed CNN can be accessed at the following URL (http://ai.skinreader.cn/) and has an online prediction function.\n\n### 2.3. Test set enrolment for evaluation\n\nTo make a statistically adequate comparison for each category of cutaneous tumour, with or without CNN assistance, while keeping a feasible total number of clinical images, a total of 400 images (40 images per category) were randomly selected from the test dataset.\n\n### 2.4. Dermatologists\n\nA total of 18 board-certified dermatologists from nine different hospitals participated as readers in this study. Their years of clinical work experience ranged from 1 to 20 years. The majority of them were not experts (experience: 13/18 10 years; 5/18\uff1e10 years). All dermatologists voluntarily participated, understood and agreed with the basic principles and objectives of this study.\n\n### 2.5. Study design\n\nTo evaluate dermatologists\u2019 performance in both assisted and unassisted conditions, the study was designed as a fully crossed, self-control, MRMC study. Dermatologists interpreted all the clinical images in both modalities (with or without CNN assistance), which was separated by a washout period of 2 months. In each modality, dermatologists were only offered one clinical image for their diagnosis and no additional information was supplied. To reduce possible performance differences at the beginning versus the end of interpreting the test set, 400 clinical images were divided into blocks of 40 images with each block containing a roughly similar distribution of tumour categories with a random order. Eighteen dermatologists were randomised into two groups, either of which began with (mode 1) or without (mode 2) CNN assistance first. In either mode, the order and images interpreted were identical; the difference was solely in modality (assisted or not assisted). The dermatologists independently assessed 400 clinical images on a 24\u201d LED monitor (Lenovo Q24i-1L) at a self-controlled pace. When the dermatologists reviewed the images, modalities (with or without CNN assistance) switched every 40 image intervals. If the image had CNN assistance, the top 3 predictions of the CNN appeared in the upper right corner of the corresponding image. If the image was without CNN assistance, only a clinical image was displayed. The participants gave the most likely diagnosis to the corresponding image.\n\n### 2.6. Statistical analysis\n\nThe predicted results of the multi-class CNN on 2107 test images were shown with a confusion matrix. The performance of the CNN model was evaluated by accuracy and kappa coefficient. In the performance evaluation of dermatologists with or without CNN assistance, the primary endpoint was the overall accuracy of dermatologists with pathological diagnosis as the gold standard. The sensitivity and specificity of binary classification was the secondary evaluation endpoint, and kappa consistency analysis was used to evaluate the consistency between dermatologist\u2019s diagnoses (with or without CNN assisted) and pathological results. Dermatologists\u2019 accuracy and kappa with or without assistance were compared by a two-sided paired t test or Wilcoxon signed-rank test, as appropriate. On the binary classification level, the sensitivity and specificity between assisted and unassisted modes were also analysed using a two-sided paired t test. And P < 0.05 was regarded as statistically significant. Statistical analyses were performed in R (version 4.1) software.",
        "results": "Results: The CNN achieved an overall accuracy of 78.45% and kappa of 0.73 in the classification of 10 types of cutaneous tumours on 2107 images. CNN-assisted dermatologists achieved a higher accuracy (76.60% vs. 62.78%, P < 0.001) and kappa (0.74 vs. 0.59, P < 0.001) than unassisted dermatologists in interpreting the 400 clinical images. Dermatologists with less experience benefited more from CNN assistance. At the binary classification level (malignant or benign), the sensitivity (89.56% vs. 83.21%, P < 0.001) and specificity (87.90% vs. 80.92%, P < 0.001) of dermatologists with CNN assistance were also significantly improved than those without.",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices.",
        "abstract": "## IMPORTANCE\nMost dermatologic cases are initially evaluated by nondermatologists such as primary care physicians (PCPs) or nurse practitioners (NPs).\n## OBJECTIVE\nTo evaluate an artificial intelligence (AI)-based tool that assists with diagnoses of dermatologic conditions.\n## DESIGN, SETTING, AND PARTICIPANTS\nThis multiple-reader, multiple-case diagnostic study developed an AI-based tool and evaluated its utility. Primary care physicians and NPs retrospectively reviewed an enriched set of cases representing 120 different skin conditions. Randomization was used to ensure each clinician reviewed each case either with or without AI assistance; each clinician alternated between batches of 50 cases in each modality. The reviews occurred from February 21 to April 28, 2020. Data were analyzed from May 26, 2020, to January 27, 2021.\n## EXPOSURES\nAn AI-based assistive tool for interpreting clinical images and associated medical history.\n## MAIN OUTCOMES AND MEASURES\nThe primary analysis evaluated agreement with reference diagnoses provided by a panel of 3 dermatologists for PCPs and NPs. Secondary analyses included diagnostic accuracy for biopsy-confirmed cases, biopsy and referral rates, review time, and diagnostic confidence.\n## RESULTS\nForty board-certified clinicians, including 20 PCPs (14 women [70.0%]; mean experience, 11.3 [range, 2-32] years) and 20 NPs (18 women [90.0%]; mean experience, 13.1 [range, 2-34] years) reviewed 1048 retrospective cases (672 female [64.2%]; median age, 43 [interquartile range, 30-56] years; 41\u202f920 total reviews) from a teledermatology practice serving 11 sites and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]). The PCPs were located across 12 states, and the NPs practiced in primary care without physician supervision across 9 states. The NPs had a mean of 13.1 (range, 2-34) years of experience and practiced in primary care without physician supervision across 9 states. Artificial intelligence assistance was significantly associated with higher agreement with reference diagnoses. For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P\u2009<\u2009.001), from 48% to 58%; for NPs, the increase was 12% (95% CI, 10%-14%; P\u2009<\u2009.001), from 46% to 58%. In secondary analyses, agreement with biopsy-obtained diagnosis categories of maglignant, precancerous, or benign increased by 3% (95% CI, -1% to 7%) for PCPs and by 8% (95% CI, 3%-13%) for NPs. Rates of desire for biopsies decreased by 1% (95% CI, 0-3%) for PCPs and 2% (95% CI, 1%-3%) for NPs; the rate of desire for referrals decreased by 3% (95% CI, 1%-4%) for PCPs and NPs. Diagnostic agreement on cases not indicated for a dermatologist referral increased by 10% (95% CI, 8%-12%) for PCPs and 12% (95% CI, 10%-14%) for NPs, and median review time increased slightly by 5 (95% CI, 0-8) seconds for PCPs and 7 (95% CI, 5-10) seconds for NPs per case.\n## CONCLUSIONS AND RELEVANCE\nArtificial intelligence assistance was associated with improved diagnoses by PCPs and NPs for 1 in every 8 to 10 cases, indicating potential for improving the quality of dermatologic care.\n",
        "methods": "### Methods\n\nThis study was approved by the Quorum Institutional Review Board, Seattle, Washington, and deemed exempt from informed consent because all data and images were deidentified. The [Standards for Reporting of Diagnostic Accuracy (STARD)[25]](http://www.equator-network.org/reporting-guidelines/stard/) reporting guideline was followed for this study.\n\n#### The AI Tool\nLiu et al[26] previously described an AI algorithm that provides a differential diagnosis given clinical photographs of skin conditions and the medical history (eTable 1 in the Supplement). Their AI model was developed using 16 114 cases and used a convolutional neural network to output prediction scores across 419 skin conditions. In the present study, we created a web-based tool using the AI model described by Liu et al by incorporating user experience insights (Figure 1).\n\nThe tool provides information about the case, including demographic information, history of present illness, and other elements of the patient\u2019s medical history. For each case, 1 to 6 images were available for review (median, 4), and readers could toggle between or zoom in on images. Primary care physicians and NPs reviewed these cases using a laptop and could consult additional resources as they would in clinical practice.\n\nThe AI assistance component of the web-based tool was only available during the assisted mode of the study (described below). At the top of the panel, the interface displayed the skin conditions that were output by the AI, sorted in order of the AI\u2019s predicted likelihood scores. Artificial intelligence predictions with low scores (<0.05) were removed, and the list was limited to 5 skin conditions to avoid presenting extraneous information. Each condition could be clicked on to display additional information (Figure 1 and eFigure 1 and the AI Tool Interface section in the eMethods in the Supplement).\n\n#### Study Design\nTo evaluate whether this tool could assist primary care clinicians in diagnosing skin conditions, we conducted a multiple-reader, multiple-case diagnostic study with 20 PCPs and 20 NPs (Figure 1). The characteristics of the clinicians are described in the Reader Characteristics section of the eMethods and eFigures 2 and 3 in the Supplement. Before reviewing the study cases, each reader was presented with materials describing how to use the AI assistant and given the opportunity to practice using the AI assistant with 2 sample cases (independent of the study cases). Additional details of this training[27] can be found in the Onboarding Process section in the eMethods in the Supplement.\n\nThe study used cases from 2 retrospective data sets from California and Hawaii previously used to validate the AI algorithm.[26] Specifically, the prior study used a validation set A and a subset (validation set B) enriched for rarer conditions via random sampling stratified by condition. Validation set B (963 cases) was included in its entirety. From validation set A, all 85 cases for which biopsy results were available were also included to yield a total of 1048 cases (Table). None of the PCPs or NPs in this study previously reviewed these cases, and the AI algorithm used was identical to the one used in the previous study.[26]\n\nEach reader was randomly assigned to 1 of 2 reader cohorts. The 2 reader cohorts read the same cases but with the opposite assistance modalities (ie, unassisted vs AI assisted) for each case. To reduce effects associated with switching modalities, the 1048 cases were divided into batches of 50 cases (except the last 48 cases, which were divided into 2 batches of 24 cases), and the assistance modality switched after each batch of cases. For the first batch of 50 cases, reader cohort 1 reviewed these cases with AI assistance, whereas reader cohort 2 reviewed the same cases unassisted. The next batch of cases were reviewed in the opposite modality (Figure 1). By ensuring each reader reviewed each case only once in either the assisted or unassisted modality, this design eliminated any memory effect associated with a crossover study (where memorable cases may inflate the diagnostic performance when reviewed a second time by the same readers).\n\nDuring the case reviews, the readers either provided their top differential diagnoses or indicated that they were unable to diagnose a case. They also answered a few questions on their intended clinical next steps for each case (see the Study End Points section below). Reviews were performed without time constraint. These reviews occurred from February 21 to April 28, 2020.\n\n#### Reference Diagnoses\nReference diagnoses were provided by a panel of dermatologists.[26] Briefly, 3 US board-certified dermatologists (from a pool of 12) independently reviewed each case. The dermatologists participated in the study via Advanced Clinical, Deerfield, Illinois; had 5 to 13 years of experience (mean [SD], 7.2 [2.7] years); and practiced in multiple states, including Colorado, Hawaii, Iowa, Maryland, New York, South Carolina, Tennessee, and Texas. Reference diagnoses were obtained using a previously described collective intelligence approach, which results in more reproducible diagnoses than diagnoses obtained by individual dermatologist review (eTable 2 in the Supplement). This approach assigns a vote to each diagnosis based on its ranking: the first diagnosis in a dermatologist\u2019s differential was given a weight of 1/1 = 1; the secondary diagnosis was given a weight of 1/2 = 0.5. The votes for each diagnosis were summed across the 3 dermatologists, and the top-voted diagnosis was considered the primary diagnosis of the panel.\n\nAgreement was also assessed against biopsy-confirmed diagnoses when available. Diagnoses were extracted from pathology reports by the teledermatology service before transfer to study investigators. These diagnoses were then mapped to skin conditions by US board-certified dermatologists (including K.K. and S.J.H.). The case distribution across these diagnoses (both clinical and histologic) are presented in the Table; of 152 cases with available biopsy results, the diagnosis of 141 cases was growths.\n\n#### Study End Points\nOur study was designed to evaluate 2 prespecified primary end points: (1) the agreement rate of the primary differential diagnosis of the PCPs with the reference diagnosis and (2) the agreement rate of the primary differential diagnosis of the NPs with the reference diagnosis. Based on the relative frequencies of conditions in this data set, the chance agreement is 3.77%.\n\nSeveral secondary analyses were planned. First, for cases with biopsy results, diagnoses were classified as malignant, precancerous, or benign and were evaluated against biopsy-determined diagnoses. Clinicians were also asked to report whether they would have recommended a biopsy or referred the case to a dermatologist. For the subset of reads in which clinicians reported they would not opt for a referral, we assessed the diagnostic agreement rate. We also analyzed the time taken to review cases and self-reported diagnostic confidence.\n\nFinally, 2 additional metrics (top-3 agreement and average overlap) were used for more comprehensive evaluation of cases in which additional follow-up may be needed to arrive at a definitive diagnosis (Additional Evaluation Metrics section in the eMethods in the Supplement). An exploratory analysis also measured the effect of AI assistance on dermatologist agreement with reference diagnoses.\n\n#### Statistical Analysis\nData were analyzed from May 26, 2020, to January 27, 2021. To compare clinicians reviewing cases with AI assistance and reviewing cases without, we used a permutation test with 1000 iterations. In each iteration, we permuted the assignment of whether reads were assisted or unassisted (ie, one-half of the full set of assisted and unassisted reads per case were selected to be assisted and the other half unassisted). Sensitivity analysis using a permutation test that preserved the reader cohorts\u2019 structure and another statistical analysis via a generalized linear mixed model produced similar results (see the Alternative Statistical Analyses section in the eMethods in the Supplement).\n\nBecause this study had 2 prespecified primary end points (both 1-tailed superiority tests), we applied the Bonferroni correction, and P < .0125 was considered statistically significant (halved from \u03b1 = 0.05 owing to 1-tailed tests and halved again owing to having 2 primary end points). Confidence intervals were computed by bootstrapping across both cases and readers for each sampled case (1000 iterations; sampling both cases and reader with replacement in each iteration). Hypothesis tests were conducted in Python, version 3.6.7 (Python Software Foundation).",
        "results": "### Results\n\nThis study involved the participation of 40 board-certified clinicians, including 20 PCPs (14 women [70.0%] and 6 men [30.0%]; mean experience, 11.3 [range, 2-32] years) who were located across 12 states and 20 NPs (18 women [90.0%] and 2 men [10.0%]; mean experience, 13.1 [range, 2-34] years) who practiced in primary care without physician supervision across 9 states. These clinicians reviewed 1048 teledermatology cases (672 women [64.2%] and 375 men [35.8%], with 1 missing; median age, 43 [interquartile range, 30-56] years) from 11 sites (Table) and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]), for a total of 41 920 case reviews. Every PCP and NP reviewed each case only once, either with or without AI assistance (Figure 1).\n\nArtificial intelligence assistance was associated with significantly higher top-1 agreement with the reference diagnosis (Figure 2A and eTable 3 in the Supplement). For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P < .001), from 48% to 58%; for NPs, the improvement was 12% (95% CI, 10%-14%; P < .001), from 46% to 58%. Assistance was associated with improvements for all 40 readers, although the magnitude varied by reader (range, 2%-22%; median, 10%) (Figure 2B). Similar improvements were observed beyond the primary diagnosis based on the top-3 agreement, average overlap, per-condition sensitivity, and \u03ba value (eFigures 4 and 5 and eTable 3 in the Supplement). In an exploratory analysis, 2 dermatologists\u2019 agreement with the reference diagnosis remained largely unchanged with AI assistance, increasing by 2% (95% CI, \u22121% to 5%), from 63% to 66% (eFigures 4 and 5 in the Supplement).\n\nFor cases with available biopsy diagnoses (n = 141), the readers\u2019 accuracy at classifying lesions as malignant, precancerous, or benign trended upward by 3% for PCPs (95% CI, \u22121% to 7%) from 64% to 67% and by 8% for NPs (95% CI, 3%-13%) from 60% to 68% (Figure 2C-D). Subgroup analysis further found that sensitivity for malignant lesions, precancerous lesions, infectious skin diseases, and categories of hair loss trended upward or remained similar with assistance for both NPs and PCPs, with improvements ranging from \u22121% to 36% (eTable 4 in the Supplement).\n\nOn the subset of cases in which the top prediction of AI was accurate (63% of cases), the use of assistance was associated with an increased top-1 agreement with reference diagnosis of 18% (95% CI, 16%-20%) for PCPs and 21% (95% CI, 19%-23%) for NPs. On the contrary, when none of the AI tool\u2019s predictions was correct (13% of cases), the agreement was 8% lower (95% CI, 5%-12%) for PCPs and 9% lower (95% CI, 6%-12%) for NPs. The effects were intermediate when the correct diagnosis was in the second or third position instead of the first (see the Impact of AI Accuracy on Assistance section of eMethods and eFigures 6 and 7 in the Supplement). An exploratory analysis also suggested that assistance was particularly beneficial for less ambiguous cases. For example, in the subset of cases in which the dermatologist panel had unanimous agreement, the use of AI assistance was associated with a top-1 agreement increase of 13% (95% CI, 10%-15%) for PCPs and of 16% (95% CI, 14%-19%) for NPs (eFigure 8 in the Supplement). Subanalyses also indicated that assistance-associated benefits were consistent during the study and across several skin types (eFigures 9 and 10 in the Supplement).\n\nArtificial intelligence assistance was also associated with changes in several simulated clinical decisions (Figure 3A-B). The rates of indicating a need for biopsy were 1% lower (95% CI, 0%-3%) for PCPs and 2% lower (95% CI, 1%-3%) for NPs; the rate of desire was 3% lower (95% CI, 1%-4%) for both PCPs and NPs (eTable 5 in the Supplement). For cases in which readers indicated referrals were unnecessary, their top-1 agreement rate with dermatologists was higher by 10% for PCPs (95% CI, 8%-12%), from 51% to 61%, and by 12% for NPs (95% CI, 10%-14%), from 47% to 59%, with a similar effect on referred cases (Figure 3C-D and eFigure 11 in the Supplement). In related findings, self-reported diagnostic confidence was substantially higher with AI assistance for both reader cohorts (Figure 4A). The top-1 agreement of cases that were rated with more than 90% confidence was substantially higher (73% vs 64% for PCPs and 68% vs 58% for NPs) (eFigure 12 in the Supplement).\n\nIn terms of review time per case, AI assistance was associated with a slightly increased median review time. A difference of 5 (95% CI, 0-8) seconds, from 89 to 94 seconds, was observed for PCPs and a difference of 7 (95% CI, 5-10) seconds, from 77 to 84 seconds, was observed for NPs (Figure 4B and eFigure 9D in the Supplement). We also present representative examples of cases in which AI assistance was associated with the largest increases or decreases in agreement with reference diagnoses (eFigures 13 and 14 in the Supplement) and results of follow-up surveys investigating the usefulness of various AI assistant features (eFigures 15-17 in the Supplement).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices.",
        "abstract": "## IMPORTANCE\nMost dermatologic cases are initially evaluated by nondermatologists such as primary care physicians (PCPs) or nurse practitioners (NPs).\n## OBJECTIVE\nTo evaluate an artificial intelligence (AI)-based tool that assists with diagnoses of dermatologic conditions.\n## DESIGN, SETTING, AND PARTICIPANTS\nThis multiple-reader, multiple-case diagnostic study developed an AI-based tool and evaluated its utility. Primary care physicians and NPs retrospectively reviewed an enriched set of cases representing 120 different skin conditions. Randomization was used to ensure each clinician reviewed each case either with or without AI assistance; each clinician alternated between batches of 50 cases in each modality. The reviews occurred from February 21 to April 28, 2020. Data were analyzed from May 26, 2020, to January 27, 2021.\n## EXPOSURES\nAn AI-based assistive tool for interpreting clinical images and associated medical history.\n## MAIN OUTCOMES AND MEASURES\nThe primary analysis evaluated agreement with reference diagnoses provided by a panel of 3 dermatologists for PCPs and NPs. Secondary analyses included diagnostic accuracy for biopsy-confirmed cases, biopsy and referral rates, review time, and diagnostic confidence.\n## RESULTS\nForty board-certified clinicians, including 20 PCPs (14 women [70.0%]; mean experience, 11.3 [range, 2-32] years) and 20 NPs (18 women [90.0%]; mean experience, 13.1 [range, 2-34] years) reviewed 1048 retrospective cases (672 female [64.2%]; median age, 43 [interquartile range, 30-56] years; 41\u202f920 total reviews) from a teledermatology practice serving 11 sites and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]). The PCPs were located across 12 states, and the NPs practiced in primary care without physician supervision across 9 states. The NPs had a mean of 13.1 (range, 2-34) years of experience and practiced in primary care without physician supervision across 9 states. Artificial intelligence assistance was significantly associated with higher agreement with reference diagnoses. For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P\u2009<\u2009.001), from 48% to 58%; for NPs, the increase was 12% (95% CI, 10%-14%; P\u2009<\u2009.001), from 46% to 58%. In secondary analyses, agreement with biopsy-obtained diagnosis categories of maglignant, precancerous, or benign increased by 3% (95% CI, -1% to 7%) for PCPs and by 8% (95% CI, 3%-13%) for NPs. Rates of desire for biopsies decreased by 1% (95% CI, 0-3%) for PCPs and 2% (95% CI, 1%-3%) for NPs; the rate of desire for referrals decreased by 3% (95% CI, 1%-4%) for PCPs and NPs. Diagnostic agreement on cases not indicated for a dermatologist referral increased by 10% (95% CI, 8%-12%) for PCPs and 12% (95% CI, 10%-14%) for NPs, and median review time increased slightly by 5 (95% CI, 0-8) seconds for PCPs and 7 (95% CI, 5-10) seconds for NPs per case.\n## CONCLUSIONS AND RELEVANCE\nArtificial intelligence assistance was associated with improved diagnoses by PCPs and NPs for 1 in every 8 to 10 cases, indicating potential for improving the quality of dermatologic care.\n",
        "methods": "### Methods\n\nThis study was approved by the Quorum Institutional Review Board, Seattle, Washington, and deemed exempt from informed consent because all data and images were deidentified. The [Standards for Reporting of Diagnostic Accuracy (STARD)[25]](http://www.equator-network.org/reporting-guidelines/stard/) reporting guideline was followed for this study.\n\n#### The AI Tool\nLiu et al[26] previously described an AI algorithm that provides a differential diagnosis given clinical photographs of skin conditions and the medical history (eTable 1 in the Supplement). Their AI model was developed using 16 114 cases and used a convolutional neural network to output prediction scores across 419 skin conditions. In the present study, we created a web-based tool using the AI model described by Liu et al by incorporating user experience insights (Figure 1).\n\nThe tool provides information about the case, including demographic information, history of present illness, and other elements of the patient\u2019s medical history. For each case, 1 to 6 images were available for review (median, 4), and readers could toggle between or zoom in on images. Primary care physicians and NPs reviewed these cases using a laptop and could consult additional resources as they would in clinical practice.\n\nThe AI assistance component of the web-based tool was only available during the assisted mode of the study (described below). At the top of the panel, the interface displayed the skin conditions that were output by the AI, sorted in order of the AI\u2019s predicted likelihood scores. Artificial intelligence predictions with low scores (<0.05) were removed, and the list was limited to 5 skin conditions to avoid presenting extraneous information. Each condition could be clicked on to display additional information (Figure 1 and eFigure 1 and the AI Tool Interface section in the eMethods in the Supplement).\n\n#### Study Design\nTo evaluate whether this tool could assist primary care clinicians in diagnosing skin conditions, we conducted a multiple-reader, multiple-case diagnostic study with 20 PCPs and 20 NPs (Figure 1). The characteristics of the clinicians are described in the Reader Characteristics section of the eMethods and eFigures 2 and 3 in the Supplement. Before reviewing the study cases, each reader was presented with materials describing how to use the AI assistant and given the opportunity to practice using the AI assistant with 2 sample cases (independent of the study cases). Additional details of this training[27] can be found in the Onboarding Process section in the eMethods in the Supplement.\n\nThe study used cases from 2 retrospective data sets from California and Hawaii previously used to validate the AI algorithm.[26] Specifically, the prior study used a validation set A and a subset (validation set B) enriched for rarer conditions via random sampling stratified by condition. Validation set B (963 cases) was included in its entirety. From validation set A, all 85 cases for which biopsy results were available were also included to yield a total of 1048 cases (Table). None of the PCPs or NPs in this study previously reviewed these cases, and the AI algorithm used was identical to the one used in the previous study.[26]\n\nEach reader was randomly assigned to 1 of 2 reader cohorts. The 2 reader cohorts read the same cases but with the opposite assistance modalities (ie, unassisted vs AI assisted) for each case. To reduce effects associated with switching modalities, the 1048 cases were divided into batches of 50 cases (except the last 48 cases, which were divided into 2 batches of 24 cases), and the assistance modality switched after each batch of cases. For the first batch of 50 cases, reader cohort 1 reviewed these cases with AI assistance, whereas reader cohort 2 reviewed the same cases unassisted. The next batch of cases were reviewed in the opposite modality (Figure 1). By ensuring each reader reviewed each case only once in either the assisted or unassisted modality, this design eliminated any memory effect associated with a crossover study (where memorable cases may inflate the diagnostic performance when reviewed a second time by the same readers).\n\nDuring the case reviews, the readers either provided their top differential diagnoses or indicated that they were unable to diagnose a case. They also answered a few questions on their intended clinical next steps for each case (see the Study End Points section below). Reviews were performed without time constraint. These reviews occurred from February 21 to April 28, 2020.\n\n#### Reference Diagnoses\nReference diagnoses were provided by a panel of dermatologists.[26] Briefly, 3 US board-certified dermatologists (from a pool of 12) independently reviewed each case. The dermatologists participated in the study via Advanced Clinical, Deerfield, Illinois; had 5 to 13 years of experience (mean [SD], 7.2 [2.7] years); and practiced in multiple states, including Colorado, Hawaii, Iowa, Maryland, New York, South Carolina, Tennessee, and Texas. Reference diagnoses were obtained using a previously described collective intelligence approach, which results in more reproducible diagnoses than diagnoses obtained by individual dermatologist review (eTable 2 in the Supplement). This approach assigns a vote to each diagnosis based on its ranking: the first diagnosis in a dermatologist\u2019s differential was given a weight of 1/1 = 1; the secondary diagnosis was given a weight of 1/2 = 0.5. The votes for each diagnosis were summed across the 3 dermatologists, and the top-voted diagnosis was considered the primary diagnosis of the panel.\n\nAgreement was also assessed against biopsy-confirmed diagnoses when available. Diagnoses were extracted from pathology reports by the teledermatology service before transfer to study investigators. These diagnoses were then mapped to skin conditions by US board-certified dermatologists (including K.K. and S.J.H.). The case distribution across these diagnoses (both clinical and histologic) are presented in the Table; of 152 cases with available biopsy results, the diagnosis of 141 cases was growths.\n\n#### Study End Points\nOur study was designed to evaluate 2 prespecified primary end points: (1) the agreement rate of the primary differential diagnosis of the PCPs with the reference diagnosis and (2) the agreement rate of the primary differential diagnosis of the NPs with the reference diagnosis. Based on the relative frequencies of conditions in this data set, the chance agreement is 3.77%.\n\nSeveral secondary analyses were planned. First, for cases with biopsy results, diagnoses were classified as malignant, precancerous, or benign and were evaluated against biopsy-determined diagnoses. Clinicians were also asked to report whether they would have recommended a biopsy or referred the case to a dermatologist. For the subset of reads in which clinicians reported they would not opt for a referral, we assessed the diagnostic agreement rate. We also analyzed the time taken to review cases and self-reported diagnostic confidence.\n\nFinally, 2 additional metrics (top-3 agreement and average overlap) were used for more comprehensive evaluation of cases in which additional follow-up may be needed to arrive at a definitive diagnosis (Additional Evaluation Metrics section in the eMethods in the Supplement). An exploratory analysis also measured the effect of AI assistance on dermatologist agreement with reference diagnoses.\n\n#### Statistical Analysis\nData were analyzed from May 26, 2020, to January 27, 2021. To compare clinicians reviewing cases with AI assistance and reviewing cases without, we used a permutation test with 1000 iterations. In each iteration, we permuted the assignment of whether reads were assisted or unassisted (ie, one-half of the full set of assisted and unassisted reads per case were selected to be assisted and the other half unassisted). Sensitivity analysis using a permutation test that preserved the reader cohorts\u2019 structure and another statistical analysis via a generalized linear mixed model produced similar results (see the Alternative Statistical Analyses section in the eMethods in the Supplement).\n\nBecause this study had 2 prespecified primary end points (both 1-tailed superiority tests), we applied the Bonferroni correction, and P < .0125 was considered statistically significant (halved from \u03b1 = 0.05 owing to 1-tailed tests and halved again owing to having 2 primary end points). Confidence intervals were computed by bootstrapping across both cases and readers for each sampled case (1000 iterations; sampling both cases and reader with replacement in each iteration). Hypothesis tests were conducted in Python, version 3.6.7 (Python Software Foundation).",
        "results": "### Results\n\nThis study involved the participation of 40 board-certified clinicians, including 20 PCPs (14 women [70.0%] and 6 men [30.0%]; mean experience, 11.3 [range, 2-32] years) who were located across 12 states and 20 NPs (18 women [90.0%] and 2 men [10.0%]; mean experience, 13.1 [range, 2-34] years) who practiced in primary care without physician supervision across 9 states. These clinicians reviewed 1048 teledermatology cases (672 women [64.2%] and 375 men [35.8%], with 1 missing; median age, 43 [interquartile range, 30-56] years) from 11 sites (Table) and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]), for a total of 41 920 case reviews. Every PCP and NP reviewed each case only once, either with or without AI assistance (Figure 1).\n\nArtificial intelligence assistance was associated with significantly higher top-1 agreement with the reference diagnosis (Figure 2A and eTable 3 in the Supplement). For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P < .001), from 48% to 58%; for NPs, the improvement was 12% (95% CI, 10%-14%; P < .001), from 46% to 58%. Assistance was associated with improvements for all 40 readers, although the magnitude varied by reader (range, 2%-22%; median, 10%) (Figure 2B). Similar improvements were observed beyond the primary diagnosis based on the top-3 agreement, average overlap, per-condition sensitivity, and \u03ba value (eFigures 4 and 5 and eTable 3 in the Supplement). In an exploratory analysis, 2 dermatologists\u2019 agreement with the reference diagnosis remained largely unchanged with AI assistance, increasing by 2% (95% CI, \u22121% to 5%), from 63% to 66% (eFigures 4 and 5 in the Supplement).\n\nFor cases with available biopsy diagnoses (n = 141), the readers\u2019 accuracy at classifying lesions as malignant, precancerous, or benign trended upward by 3% for PCPs (95% CI, \u22121% to 7%) from 64% to 67% and by 8% for NPs (95% CI, 3%-13%) from 60% to 68% (Figure 2C-D). Subgroup analysis further found that sensitivity for malignant lesions, precancerous lesions, infectious skin diseases, and categories of hair loss trended upward or remained similar with assistance for both NPs and PCPs, with improvements ranging from \u22121% to 36% (eTable 4 in the Supplement).\n\nOn the subset of cases in which the top prediction of AI was accurate (63% of cases), the use of assistance was associated with an increased top-1 agreement with reference diagnosis of 18% (95% CI, 16%-20%) for PCPs and 21% (95% CI, 19%-23%) for NPs. On the contrary, when none of the AI tool\u2019s predictions was correct (13% of cases), the agreement was 8% lower (95% CI, 5%-12%) for PCPs and 9% lower (95% CI, 6%-12%) for NPs. The effects were intermediate when the correct diagnosis was in the second or third position instead of the first (see the Impact of AI Accuracy on Assistance section of eMethods and eFigures 6 and 7 in the Supplement). An exploratory analysis also suggested that assistance was particularly beneficial for less ambiguous cases. For example, in the subset of cases in which the dermatologist panel had unanimous agreement, the use of AI assistance was associated with a top-1 agreement increase of 13% (95% CI, 10%-15%) for PCPs and of 16% (95% CI, 14%-19%) for NPs (eFigure 8 in the Supplement). Subanalyses also indicated that assistance-associated benefits were consistent during the study and across several skin types (eFigures 9 and 10 in the Supplement).\n\nArtificial intelligence assistance was also associated with changes in several simulated clinical decisions (Figure 3A-B). The rates of indicating a need for biopsy were 1% lower (95% CI, 0%-3%) for PCPs and 2% lower (95% CI, 1%-3%) for NPs; the rate of desire was 3% lower (95% CI, 1%-4%) for both PCPs and NPs (eTable 5 in the Supplement). For cases in which readers indicated referrals were unnecessary, their top-1 agreement rate with dermatologists was higher by 10% for PCPs (95% CI, 8%-12%), from 51% to 61%, and by 12% for NPs (95% CI, 10%-14%), from 47% to 59%, with a similar effect on referred cases (Figure 3C-D and eFigure 11 in the Supplement). In related findings, self-reported diagnostic confidence was substantially higher with AI assistance for both reader cohorts (Figure 4A). The top-1 agreement of cases that were rated with more than 90% confidence was substantially higher (73% vs 64% for PCPs and 68% vs 58% for NPs) (eFigure 12 in the Supplement).\n\nIn terms of review time per case, AI assistance was associated with a slightly increased median review time. A difference of 5 (95% CI, 0-8) seconds, from 89 to 94 seconds, was observed for PCPs and a difference of 7 (95% CI, 5-10) seconds, from 77 to 84 seconds, was observed for NPs (Figure 4B and eFigure 9D in the Supplement). We also present representative examples of cases in which AI assistance was associated with the largest increases or decreases in agreement with reference diagnoses (eFigures 13 and 14 in the Supplement) and results of follow-up surveys investigating the usefulness of various AI assistant features (eFigures 15-17 in the Supplement).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices.",
        "abstract": "## IMPORTANCE\nMost dermatologic cases are initially evaluated by nondermatologists such as primary care physicians (PCPs) or nurse practitioners (NPs).\n## OBJECTIVE\nTo evaluate an artificial intelligence (AI)-based tool that assists with diagnoses of dermatologic conditions.\n## DESIGN, SETTING, AND PARTICIPANTS\nThis multiple-reader, multiple-case diagnostic study developed an AI-based tool and evaluated its utility. Primary care physicians and NPs retrospectively reviewed an enriched set of cases representing 120 different skin conditions. Randomization was used to ensure each clinician reviewed each case either with or without AI assistance; each clinician alternated between batches of 50 cases in each modality. The reviews occurred from February 21 to April 28, 2020. Data were analyzed from May 26, 2020, to January 27, 2021.\n## EXPOSURES\nAn AI-based assistive tool for interpreting clinical images and associated medical history.\n## MAIN OUTCOMES AND MEASURES\nThe primary analysis evaluated agreement with reference diagnoses provided by a panel of 3 dermatologists for PCPs and NPs. Secondary analyses included diagnostic accuracy for biopsy-confirmed cases, biopsy and referral rates, review time, and diagnostic confidence.\n## RESULTS\nForty board-certified clinicians, including 20 PCPs (14 women [70.0%]; mean experience, 11.3 [range, 2-32] years) and 20 NPs (18 women [90.0%]; mean experience, 13.1 [range, 2-34] years) reviewed 1048 retrospective cases (672 female [64.2%]; median age, 43 [interquartile range, 30-56] years; 41\u202f920 total reviews) from a teledermatology practice serving 11 sites and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]). The PCPs were located across 12 states, and the NPs practiced in primary care without physician supervision across 9 states. The NPs had a mean of 13.1 (range, 2-34) years of experience and practiced in primary care without physician supervision across 9 states. Artificial intelligence assistance was significantly associated with higher agreement with reference diagnoses. For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P\u2009<\u2009.001), from 48% to 58%; for NPs, the increase was 12% (95% CI, 10%-14%; P\u2009<\u2009.001), from 46% to 58%. In secondary analyses, agreement with biopsy-obtained diagnosis categories of maglignant, precancerous, or benign increased by 3% (95% CI, -1% to 7%) for PCPs and by 8% (95% CI, 3%-13%) for NPs. Rates of desire for biopsies decreased by 1% (95% CI, 0-3%) for PCPs and 2% (95% CI, 1%-3%) for NPs; the rate of desire for referrals decreased by 3% (95% CI, 1%-4%) for PCPs and NPs. Diagnostic agreement on cases not indicated for a dermatologist referral increased by 10% (95% CI, 8%-12%) for PCPs and 12% (95% CI, 10%-14%) for NPs, and median review time increased slightly by 5 (95% CI, 0-8) seconds for PCPs and 7 (95% CI, 5-10) seconds for NPs per case.\n## CONCLUSIONS AND RELEVANCE\nArtificial intelligence assistance was associated with improved diagnoses by PCPs and NPs for 1 in every 8 to 10 cases, indicating potential for improving the quality of dermatologic care.\n",
        "methods": "### Methods\n\nThis study was approved by the Quorum Institutional Review Board, Seattle, Washington, and deemed exempt from informed consent because all data and images were deidentified. The [Standards for Reporting of Diagnostic Accuracy (STARD)[25]](http://www.equator-network.org/reporting-guidelines/stard/) reporting guideline was followed for this study.\n\n#### The AI Tool\nLiu et al[26] previously described an AI algorithm that provides a differential diagnosis given clinical photographs of skin conditions and the medical history (eTable 1 in the Supplement). Their AI model was developed using 16 114 cases and used a convolutional neural network to output prediction scores across 419 skin conditions. In the present study, we created a web-based tool using the AI model described by Liu et al by incorporating user experience insights (Figure 1).\n\nThe tool provides information about the case, including demographic information, history of present illness, and other elements of the patient\u2019s medical history. For each case, 1 to 6 images were available for review (median, 4), and readers could toggle between or zoom in on images. Primary care physicians and NPs reviewed these cases using a laptop and could consult additional resources as they would in clinical practice.\n\nThe AI assistance component of the web-based tool was only available during the assisted mode of the study (described below). At the top of the panel, the interface displayed the skin conditions that were output by the AI, sorted in order of the AI\u2019s predicted likelihood scores. Artificial intelligence predictions with low scores (<0.05) were removed, and the list was limited to 5 skin conditions to avoid presenting extraneous information. Each condition could be clicked on to display additional information (Figure 1 and eFigure 1 and the AI Tool Interface section in the eMethods in the Supplement).\n\n#### Study Design\nTo evaluate whether this tool could assist primary care clinicians in diagnosing skin conditions, we conducted a multiple-reader, multiple-case diagnostic study with 20 PCPs and 20 NPs (Figure 1). The characteristics of the clinicians are described in the Reader Characteristics section of the eMethods and eFigures 2 and 3 in the Supplement. Before reviewing the study cases, each reader was presented with materials describing how to use the AI assistant and given the opportunity to practice using the AI assistant with 2 sample cases (independent of the study cases). Additional details of this training[27] can be found in the Onboarding Process section in the eMethods in the Supplement.\n\nThe study used cases from 2 retrospective data sets from California and Hawaii previously used to validate the AI algorithm.[26] Specifically, the prior study used a validation set A and a subset (validation set B) enriched for rarer conditions via random sampling stratified by condition. Validation set B (963 cases) was included in its entirety. From validation set A, all 85 cases for which biopsy results were available were also included to yield a total of 1048 cases (Table). None of the PCPs or NPs in this study previously reviewed these cases, and the AI algorithm used was identical to the one used in the previous study.[26]\n\nEach reader was randomly assigned to 1 of 2 reader cohorts. The 2 reader cohorts read the same cases but with the opposite assistance modalities (ie, unassisted vs AI assisted) for each case. To reduce effects associated with switching modalities, the 1048 cases were divided into batches of 50 cases (except the last 48 cases, which were divided into 2 batches of 24 cases), and the assistance modality switched after each batch of cases. For the first batch of 50 cases, reader cohort 1 reviewed these cases with AI assistance, whereas reader cohort 2 reviewed the same cases unassisted. The next batch of cases were reviewed in the opposite modality (Figure 1). By ensuring each reader reviewed each case only once in either the assisted or unassisted modality, this design eliminated any memory effect associated with a crossover study (where memorable cases may inflate the diagnostic performance when reviewed a second time by the same readers).\n\nDuring the case reviews, the readers either provided their top differential diagnoses or indicated that they were unable to diagnose a case. They also answered a few questions on their intended clinical next steps for each case (see the Study End Points section below). Reviews were performed without time constraint. These reviews occurred from February 21 to April 28, 2020.\n\n#### Reference Diagnoses\nReference diagnoses were provided by a panel of dermatologists.[26] Briefly, 3 US board-certified dermatologists (from a pool of 12) independently reviewed each case. The dermatologists participated in the study via Advanced Clinical, Deerfield, Illinois; had 5 to 13 years of experience (mean [SD], 7.2 [2.7] years); and practiced in multiple states, including Colorado, Hawaii, Iowa, Maryland, New York, South Carolina, Tennessee, and Texas. Reference diagnoses were obtained using a previously described collective intelligence approach, which results in more reproducible diagnoses than diagnoses obtained by individual dermatologist review (eTable 2 in the Supplement). This approach assigns a vote to each diagnosis based on its ranking: the first diagnosis in a dermatologist\u2019s differential was given a weight of 1/1 = 1; the secondary diagnosis was given a weight of 1/2 = 0.5. The votes for each diagnosis were summed across the 3 dermatologists, and the top-voted diagnosis was considered the primary diagnosis of the panel.\n\nAgreement was also assessed against biopsy-confirmed diagnoses when available. Diagnoses were extracted from pathology reports by the teledermatology service before transfer to study investigators. These diagnoses were then mapped to skin conditions by US board-certified dermatologists (including K.K. and S.J.H.). The case distribution across these diagnoses (both clinical and histologic) are presented in the Table; of 152 cases with available biopsy results, the diagnosis of 141 cases was growths.\n\n#### Study End Points\nOur study was designed to evaluate 2 prespecified primary end points: (1) the agreement rate of the primary differential diagnosis of the PCPs with the reference diagnosis and (2) the agreement rate of the primary differential diagnosis of the NPs with the reference diagnosis. Based on the relative frequencies of conditions in this data set, the chance agreement is 3.77%.\n\nSeveral secondary analyses were planned. First, for cases with biopsy results, diagnoses were classified as malignant, precancerous, or benign and were evaluated against biopsy-determined diagnoses. Clinicians were also asked to report whether they would have recommended a biopsy or referred the case to a dermatologist. For the subset of reads in which clinicians reported they would not opt for a referral, we assessed the diagnostic agreement rate. We also analyzed the time taken to review cases and self-reported diagnostic confidence.\n\nFinally, 2 additional metrics (top-3 agreement and average overlap) were used for more comprehensive evaluation of cases in which additional follow-up may be needed to arrive at a definitive diagnosis (Additional Evaluation Metrics section in the eMethods in the Supplement). An exploratory analysis also measured the effect of AI assistance on dermatologist agreement with reference diagnoses.\n\n#### Statistical Analysis\nData were analyzed from May 26, 2020, to January 27, 2021. To compare clinicians reviewing cases with AI assistance and reviewing cases without, we used a permutation test with 1000 iterations. In each iteration, we permuted the assignment of whether reads were assisted or unassisted (ie, one-half of the full set of assisted and unassisted reads per case were selected to be assisted and the other half unassisted). Sensitivity analysis using a permutation test that preserved the reader cohorts\u2019 structure and another statistical analysis via a generalized linear mixed model produced similar results (see the Alternative Statistical Analyses section in the eMethods in the Supplement).\n\nBecause this study had 2 prespecified primary end points (both 1-tailed superiority tests), we applied the Bonferroni correction, and P < .0125 was considered statistically significant (halved from \u03b1 = 0.05 owing to 1-tailed tests and halved again owing to having 2 primary end points). Confidence intervals were computed by bootstrapping across both cases and readers for each sampled case (1000 iterations; sampling both cases and reader with replacement in each iteration). Hypothesis tests were conducted in Python, version 3.6.7 (Python Software Foundation).",
        "results": "### Results\n\nThis study involved the participation of 40 board-certified clinicians, including 20 PCPs (14 women [70.0%] and 6 men [30.0%]; mean experience, 11.3 [range, 2-32] years) who were located across 12 states and 20 NPs (18 women [90.0%] and 2 men [10.0%]; mean experience, 13.1 [range, 2-34] years) who practiced in primary care without physician supervision across 9 states. These clinicians reviewed 1048 teledermatology cases (672 women [64.2%] and 375 men [35.8%], with 1 missing; median age, 43 [interquartile range, 30-56] years) from 11 sites (Table) and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]), for a total of 41 920 case reviews. Every PCP and NP reviewed each case only once, either with or without AI assistance (Figure 1).\n\nArtificial intelligence assistance was associated with significantly higher top-1 agreement with the reference diagnosis (Figure 2A and eTable 3 in the Supplement). For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P < .001), from 48% to 58%; for NPs, the improvement was 12% (95% CI, 10%-14%; P < .001), from 46% to 58%. Assistance was associated with improvements for all 40 readers, although the magnitude varied by reader (range, 2%-22%; median, 10%) (Figure 2B). Similar improvements were observed beyond the primary diagnosis based on the top-3 agreement, average overlap, per-condition sensitivity, and \u03ba value (eFigures 4 and 5 and eTable 3 in the Supplement). In an exploratory analysis, 2 dermatologists\u2019 agreement with the reference diagnosis remained largely unchanged with AI assistance, increasing by 2% (95% CI, \u22121% to 5%), from 63% to 66% (eFigures 4 and 5 in the Supplement).\n\nFor cases with available biopsy diagnoses (n = 141), the readers\u2019 accuracy at classifying lesions as malignant, precancerous, or benign trended upward by 3% for PCPs (95% CI, \u22121% to 7%) from 64% to 67% and by 8% for NPs (95% CI, 3%-13%) from 60% to 68% (Figure 2C-D). Subgroup analysis further found that sensitivity for malignant lesions, precancerous lesions, infectious skin diseases, and categories of hair loss trended upward or remained similar with assistance for both NPs and PCPs, with improvements ranging from \u22121% to 36% (eTable 4 in the Supplement).\n\nOn the subset of cases in which the top prediction of AI was accurate (63% of cases), the use of assistance was associated with an increased top-1 agreement with reference diagnosis of 18% (95% CI, 16%-20%) for PCPs and 21% (95% CI, 19%-23%) for NPs. On the contrary, when none of the AI tool\u2019s predictions was correct (13% of cases), the agreement was 8% lower (95% CI, 5%-12%) for PCPs and 9% lower (95% CI, 6%-12%) for NPs. The effects were intermediate when the correct diagnosis was in the second or third position instead of the first (see the Impact of AI Accuracy on Assistance section of eMethods and eFigures 6 and 7 in the Supplement). An exploratory analysis also suggested that assistance was particularly beneficial for less ambiguous cases. For example, in the subset of cases in which the dermatologist panel had unanimous agreement, the use of AI assistance was associated with a top-1 agreement increase of 13% (95% CI, 10%-15%) for PCPs and of 16% (95% CI, 14%-19%) for NPs (eFigure 8 in the Supplement). Subanalyses also indicated that assistance-associated benefits were consistent during the study and across several skin types (eFigures 9 and 10 in the Supplement).\n\nArtificial intelligence assistance was also associated with changes in several simulated clinical decisions (Figure 3A-B). The rates of indicating a need for biopsy were 1% lower (95% CI, 0%-3%) for PCPs and 2% lower (95% CI, 1%-3%) for NPs; the rate of desire was 3% lower (95% CI, 1%-4%) for both PCPs and NPs (eTable 5 in the Supplement). For cases in which readers indicated referrals were unnecessary, their top-1 agreement rate with dermatologists was higher by 10% for PCPs (95% CI, 8%-12%), from 51% to 61%, and by 12% for NPs (95% CI, 10%-14%), from 47% to 59%, with a similar effect on referred cases (Figure 3C-D and eFigure 11 in the Supplement). In related findings, self-reported diagnostic confidence was substantially higher with AI assistance for both reader cohorts (Figure 4A). The top-1 agreement of cases that were rated with more than 90% confidence was substantially higher (73% vs 64% for PCPs and 68% vs 58% for NPs) (eFigure 12 in the Supplement).\n\nIn terms of review time per case, AI assistance was associated with a slightly increased median review time. A difference of 5 (95% CI, 0-8) seconds, from 89 to 94 seconds, was observed for PCPs and a difference of 7 (95% CI, 5-10) seconds, from 77 to 84 seconds, was observed for NPs (Figure 4B and eFigure 9D in the Supplement). We also present representative examples of cases in which AI assistance was associated with the largest increases or decreases in agreement with reference diagnoses (eFigures 13 and 14 in the Supplement) and results of follow-up surveys investigating the usefulness of various AI assistant features (eFigures 15-17 in the Supplement).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices.",
        "abstract": "## IMPORTANCE\nMost dermatologic cases are initially evaluated by nondermatologists such as primary care physicians (PCPs) or nurse practitioners (NPs).\n## OBJECTIVE\nTo evaluate an artificial intelligence (AI)-based tool that assists with diagnoses of dermatologic conditions.\n## DESIGN, SETTING, AND PARTICIPANTS\nThis multiple-reader, multiple-case diagnostic study developed an AI-based tool and evaluated its utility. Primary care physicians and NPs retrospectively reviewed an enriched set of cases representing 120 different skin conditions. Randomization was used to ensure each clinician reviewed each case either with or without AI assistance; each clinician alternated between batches of 50 cases in each modality. The reviews occurred from February 21 to April 28, 2020. Data were analyzed from May 26, 2020, to January 27, 2021.\n## EXPOSURES\nAn AI-based assistive tool for interpreting clinical images and associated medical history.\n## MAIN OUTCOMES AND MEASURES\nThe primary analysis evaluated agreement with reference diagnoses provided by a panel of 3 dermatologists for PCPs and NPs. Secondary analyses included diagnostic accuracy for biopsy-confirmed cases, biopsy and referral rates, review time, and diagnostic confidence.\n## RESULTS\nForty board-certified clinicians, including 20 PCPs (14 women [70.0%]; mean experience, 11.3 [range, 2-32] years) and 20 NPs (18 women [90.0%]; mean experience, 13.1 [range, 2-34] years) reviewed 1048 retrospective cases (672 female [64.2%]; median age, 43 [interquartile range, 30-56] years; 41\u202f920 total reviews) from a teledermatology practice serving 11 sites and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]). The PCPs were located across 12 states, and the NPs practiced in primary care without physician supervision across 9 states. The NPs had a mean of 13.1 (range, 2-34) years of experience and practiced in primary care without physician supervision across 9 states. Artificial intelligence assistance was significantly associated with higher agreement with reference diagnoses. For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P\u2009<\u2009.001), from 48% to 58%; for NPs, the increase was 12% (95% CI, 10%-14%; P\u2009<\u2009.001), from 46% to 58%. In secondary analyses, agreement with biopsy-obtained diagnosis categories of maglignant, precancerous, or benign increased by 3% (95% CI, -1% to 7%) for PCPs and by 8% (95% CI, 3%-13%) for NPs. Rates of desire for biopsies decreased by 1% (95% CI, 0-3%) for PCPs and 2% (95% CI, 1%-3%) for NPs; the rate of desire for referrals decreased by 3% (95% CI, 1%-4%) for PCPs and NPs. Diagnostic agreement on cases not indicated for a dermatologist referral increased by 10% (95% CI, 8%-12%) for PCPs and 12% (95% CI, 10%-14%) for NPs, and median review time increased slightly by 5 (95% CI, 0-8) seconds for PCPs and 7 (95% CI, 5-10) seconds for NPs per case.\n## CONCLUSIONS AND RELEVANCE\nArtificial intelligence assistance was associated with improved diagnoses by PCPs and NPs for 1 in every 8 to 10 cases, indicating potential for improving the quality of dermatologic care.\n",
        "methods": "### Methods\n\nThis study was approved by the Quorum Institutional Review Board, Seattle, Washington, and deemed exempt from informed consent because all data and images were deidentified. The [Standards for Reporting of Diagnostic Accuracy (STARD)[25]](http://www.equator-network.org/reporting-guidelines/stard/) reporting guideline was followed for this study.\n\n#### The AI Tool\nLiu et al[26] previously described an AI algorithm that provides a differential diagnosis given clinical photographs of skin conditions and the medical history (eTable 1 in the Supplement). Their AI model was developed using 16 114 cases and used a convolutional neural network to output prediction scores across 419 skin conditions. In the present study, we created a web-based tool using the AI model described by Liu et al by incorporating user experience insights (Figure 1).\n\nThe tool provides information about the case, including demographic information, history of present illness, and other elements of the patient\u2019s medical history. For each case, 1 to 6 images were available for review (median, 4), and readers could toggle between or zoom in on images. Primary care physicians and NPs reviewed these cases using a laptop and could consult additional resources as they would in clinical practice.\n\nThe AI assistance component of the web-based tool was only available during the assisted mode of the study (described below). At the top of the panel, the interface displayed the skin conditions that were output by the AI, sorted in order of the AI\u2019s predicted likelihood scores. Artificial intelligence predictions with low scores (<0.05) were removed, and the list was limited to 5 skin conditions to avoid presenting extraneous information. Each condition could be clicked on to display additional information (Figure 1 and eFigure 1 and the AI Tool Interface section in the eMethods in the Supplement).\n\n#### Study Design\nTo evaluate whether this tool could assist primary care clinicians in diagnosing skin conditions, we conducted a multiple-reader, multiple-case diagnostic study with 20 PCPs and 20 NPs (Figure 1). The characteristics of the clinicians are described in the Reader Characteristics section of the eMethods and eFigures 2 and 3 in the Supplement. Before reviewing the study cases, each reader was presented with materials describing how to use the AI assistant and given the opportunity to practice using the AI assistant with 2 sample cases (independent of the study cases). Additional details of this training[27] can be found in the Onboarding Process section in the eMethods in the Supplement.\n\nThe study used cases from 2 retrospective data sets from California and Hawaii previously used to validate the AI algorithm.[26] Specifically, the prior study used a validation set A and a subset (validation set B) enriched for rarer conditions via random sampling stratified by condition. Validation set B (963 cases) was included in its entirety. From validation set A, all 85 cases for which biopsy results were available were also included to yield a total of 1048 cases (Table). None of the PCPs or NPs in this study previously reviewed these cases, and the AI algorithm used was identical to the one used in the previous study.[26]\n\nEach reader was randomly assigned to 1 of 2 reader cohorts. The 2 reader cohorts read the same cases but with the opposite assistance modalities (ie, unassisted vs AI assisted) for each case. To reduce effects associated with switching modalities, the 1048 cases were divided into batches of 50 cases (except the last 48 cases, which were divided into 2 batches of 24 cases), and the assistance modality switched after each batch of cases. For the first batch of 50 cases, reader cohort 1 reviewed these cases with AI assistance, whereas reader cohort 2 reviewed the same cases unassisted. The next batch of cases were reviewed in the opposite modality (Figure 1). By ensuring each reader reviewed each case only once in either the assisted or unassisted modality, this design eliminated any memory effect associated with a crossover study (where memorable cases may inflate the diagnostic performance when reviewed a second time by the same readers).\n\nDuring the case reviews, the readers either provided their top differential diagnoses or indicated that they were unable to diagnose a case. They also answered a few questions on their intended clinical next steps for each case (see the Study End Points section below). Reviews were performed without time constraint. These reviews occurred from February 21 to April 28, 2020.\n\n#### Reference Diagnoses\nReference diagnoses were provided by a panel of dermatologists.[26] Briefly, 3 US board-certified dermatologists (from a pool of 12) independently reviewed each case. The dermatologists participated in the study via Advanced Clinical, Deerfield, Illinois; had 5 to 13 years of experience (mean [SD], 7.2 [2.7] years); and practiced in multiple states, including Colorado, Hawaii, Iowa, Maryland, New York, South Carolina, Tennessee, and Texas. Reference diagnoses were obtained using a previously described collective intelligence approach, which results in more reproducible diagnoses than diagnoses obtained by individual dermatologist review (eTable 2 in the Supplement). This approach assigns a vote to each diagnosis based on its ranking: the first diagnosis in a dermatologist\u2019s differential was given a weight of 1/1 = 1; the secondary diagnosis was given a weight of 1/2 = 0.5. The votes for each diagnosis were summed across the 3 dermatologists, and the top-voted diagnosis was considered the primary diagnosis of the panel.\n\nAgreement was also assessed against biopsy-confirmed diagnoses when available. Diagnoses were extracted from pathology reports by the teledermatology service before transfer to study investigators. These diagnoses were then mapped to skin conditions by US board-certified dermatologists (including K.K. and S.J.H.). The case distribution across these diagnoses (both clinical and histologic) are presented in the Table; of 152 cases with available biopsy results, the diagnosis of 141 cases was growths.\n\n#### Study End Points\nOur study was designed to evaluate 2 prespecified primary end points: (1) the agreement rate of the primary differential diagnosis of the PCPs with the reference diagnosis and (2) the agreement rate of the primary differential diagnosis of the NPs with the reference diagnosis. Based on the relative frequencies of conditions in this data set, the chance agreement is 3.77%.\n\nSeveral secondary analyses were planned. First, for cases with biopsy results, diagnoses were classified as malignant, precancerous, or benign and were evaluated against biopsy-determined diagnoses. Clinicians were also asked to report whether they would have recommended a biopsy or referred the case to a dermatologist. For the subset of reads in which clinicians reported they would not opt for a referral, we assessed the diagnostic agreement rate. We also analyzed the time taken to review cases and self-reported diagnostic confidence.\n\nFinally, 2 additional metrics (top-3 agreement and average overlap) were used for more comprehensive evaluation of cases in which additional follow-up may be needed to arrive at a definitive diagnosis (Additional Evaluation Metrics section in the eMethods in the Supplement). An exploratory analysis also measured the effect of AI assistance on dermatologist agreement with reference diagnoses.\n\n#### Statistical Analysis\nData were analyzed from May 26, 2020, to January 27, 2021. To compare clinicians reviewing cases with AI assistance and reviewing cases without, we used a permutation test with 1000 iterations. In each iteration, we permuted the assignment of whether reads were assisted or unassisted (ie, one-half of the full set of assisted and unassisted reads per case were selected to be assisted and the other half unassisted). Sensitivity analysis using a permutation test that preserved the reader cohorts\u2019 structure and another statistical analysis via a generalized linear mixed model produced similar results (see the Alternative Statistical Analyses section in the eMethods in the Supplement).\n\nBecause this study had 2 prespecified primary end points (both 1-tailed superiority tests), we applied the Bonferroni correction, and P < .0125 was considered statistically significant (halved from \u03b1 = 0.05 owing to 1-tailed tests and halved again owing to having 2 primary end points). Confidence intervals were computed by bootstrapping across both cases and readers for each sampled case (1000 iterations; sampling both cases and reader with replacement in each iteration). Hypothesis tests were conducted in Python, version 3.6.7 (Python Software Foundation).",
        "results": "### Results\n\nThis study involved the participation of 40 board-certified clinicians, including 20 PCPs (14 women [70.0%] and 6 men [30.0%]; mean experience, 11.3 [range, 2-32] years) who were located across 12 states and 20 NPs (18 women [90.0%] and 2 men [10.0%]; mean experience, 13.1 [range, 2-34] years) who practiced in primary care without physician supervision across 9 states. These clinicians reviewed 1048 teledermatology cases (672 women [64.2%] and 375 men [35.8%], with 1 missing; median age, 43 [interquartile range, 30-56] years) from 11 sites (Table) and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]), for a total of 41 920 case reviews. Every PCP and NP reviewed each case only once, either with or without AI assistance (Figure 1).\n\nArtificial intelligence assistance was associated with significantly higher top-1 agreement with the reference diagnosis (Figure 2A and eTable 3 in the Supplement). For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P < .001), from 48% to 58%; for NPs, the improvement was 12% (95% CI, 10%-14%; P < .001), from 46% to 58%. Assistance was associated with improvements for all 40 readers, although the magnitude varied by reader (range, 2%-22%; median, 10%) (Figure 2B). Similar improvements were observed beyond the primary diagnosis based on the top-3 agreement, average overlap, per-condition sensitivity, and \u03ba value (eFigures 4 and 5 and eTable 3 in the Supplement). In an exploratory analysis, 2 dermatologists\u2019 agreement with the reference diagnosis remained largely unchanged with AI assistance, increasing by 2% (95% CI, \u22121% to 5%), from 63% to 66% (eFigures 4 and 5 in the Supplement).\n\nFor cases with available biopsy diagnoses (n = 141), the readers\u2019 accuracy at classifying lesions as malignant, precancerous, or benign trended upward by 3% for PCPs (95% CI, \u22121% to 7%) from 64% to 67% and by 8% for NPs (95% CI, 3%-13%) from 60% to 68% (Figure 2C-D). Subgroup analysis further found that sensitivity for malignant lesions, precancerous lesions, infectious skin diseases, and categories of hair loss trended upward or remained similar with assistance for both NPs and PCPs, with improvements ranging from \u22121% to 36% (eTable 4 in the Supplement).\n\nOn the subset of cases in which the top prediction of AI was accurate (63% of cases), the use of assistance was associated with an increased top-1 agreement with reference diagnosis of 18% (95% CI, 16%-20%) for PCPs and 21% (95% CI, 19%-23%) for NPs. On the contrary, when none of the AI tool\u2019s predictions was correct (13% of cases), the agreement was 8% lower (95% CI, 5%-12%) for PCPs and 9% lower (95% CI, 6%-12%) for NPs. The effects were intermediate when the correct diagnosis was in the second or third position instead of the first (see the Impact of AI Accuracy on Assistance section of eMethods and eFigures 6 and 7 in the Supplement). An exploratory analysis also suggested that assistance was particularly beneficial for less ambiguous cases. For example, in the subset of cases in which the dermatologist panel had unanimous agreement, the use of AI assistance was associated with a top-1 agreement increase of 13% (95% CI, 10%-15%) for PCPs and of 16% (95% CI, 14%-19%) for NPs (eFigure 8 in the Supplement). Subanalyses also indicated that assistance-associated benefits were consistent during the study and across several skin types (eFigures 9 and 10 in the Supplement).\n\nArtificial intelligence assistance was also associated with changes in several simulated clinical decisions (Figure 3A-B). The rates of indicating a need for biopsy were 1% lower (95% CI, 0%-3%) for PCPs and 2% lower (95% CI, 1%-3%) for NPs; the rate of desire was 3% lower (95% CI, 1%-4%) for both PCPs and NPs (eTable 5 in the Supplement). For cases in which readers indicated referrals were unnecessary, their top-1 agreement rate with dermatologists was higher by 10% for PCPs (95% CI, 8%-12%), from 51% to 61%, and by 12% for NPs (95% CI, 10%-14%), from 47% to 59%, with a similar effect on referred cases (Figure 3C-D and eFigure 11 in the Supplement). In related findings, self-reported diagnostic confidence was substantially higher with AI assistance for both reader cohorts (Figure 4A). The top-1 agreement of cases that were rated with more than 90% confidence was substantially higher (73% vs 64% for PCPs and 68% vs 58% for NPs) (eFigure 12 in the Supplement).\n\nIn terms of review time per case, AI assistance was associated with a slightly increased median review time. A difference of 5 (95% CI, 0-8) seconds, from 89 to 94 seconds, was observed for PCPs and a difference of 7 (95% CI, 5-10) seconds, from 77 to 84 seconds, was observed for NPs (Figure 4B and eFigure 9D in the Supplement). We also present representative examples of cases in which AI assistance was associated with the largest increases or decreases in agreement with reference diagnoses (eFigures 13 and 14 in the Supplement) and results of follow-up surveys investigating the usefulness of various AI assistant features (eFigures 15-17 in the Supplement).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices.",
        "abstract": "## IMPORTANCE\nMost dermatologic cases are initially evaluated by nondermatologists such as primary care physicians (PCPs) or nurse practitioners (NPs).\n## OBJECTIVE\nTo evaluate an artificial intelligence (AI)-based tool that assists with diagnoses of dermatologic conditions.\n## DESIGN, SETTING, AND PARTICIPANTS\nThis multiple-reader, multiple-case diagnostic study developed an AI-based tool and evaluated its utility. Primary care physicians and NPs retrospectively reviewed an enriched set of cases representing 120 different skin conditions. Randomization was used to ensure each clinician reviewed each case either with or without AI assistance; each clinician alternated between batches of 50 cases in each modality. The reviews occurred from February 21 to April 28, 2020. Data were analyzed from May 26, 2020, to January 27, 2021.\n## EXPOSURES\nAn AI-based assistive tool for interpreting clinical images and associated medical history.\n## MAIN OUTCOMES AND MEASURES\nThe primary analysis evaluated agreement with reference diagnoses provided by a panel of 3 dermatologists for PCPs and NPs. Secondary analyses included diagnostic accuracy for biopsy-confirmed cases, biopsy and referral rates, review time, and diagnostic confidence.\n## RESULTS\nForty board-certified clinicians, including 20 PCPs (14 women [70.0%]; mean experience, 11.3 [range, 2-32] years) and 20 NPs (18 women [90.0%]; mean experience, 13.1 [range, 2-34] years) reviewed 1048 retrospective cases (672 female [64.2%]; median age, 43 [interquartile range, 30-56] years; 41\u202f920 total reviews) from a teledermatology practice serving 11 sites and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]). The PCPs were located across 12 states, and the NPs practiced in primary care without physician supervision across 9 states. The NPs had a mean of 13.1 (range, 2-34) years of experience and practiced in primary care without physician supervision across 9 states. Artificial intelligence assistance was significantly associated with higher agreement with reference diagnoses. For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P\u2009<\u2009.001), from 48% to 58%; for NPs, the increase was 12% (95% CI, 10%-14%; P\u2009<\u2009.001), from 46% to 58%. In secondary analyses, agreement with biopsy-obtained diagnosis categories of maglignant, precancerous, or benign increased by 3% (95% CI, -1% to 7%) for PCPs and by 8% (95% CI, 3%-13%) for NPs. Rates of desire for biopsies decreased by 1% (95% CI, 0-3%) for PCPs and 2% (95% CI, 1%-3%) for NPs; the rate of desire for referrals decreased by 3% (95% CI, 1%-4%) for PCPs and NPs. Diagnostic agreement on cases not indicated for a dermatologist referral increased by 10% (95% CI, 8%-12%) for PCPs and 12% (95% CI, 10%-14%) for NPs, and median review time increased slightly by 5 (95% CI, 0-8) seconds for PCPs and 7 (95% CI, 5-10) seconds for NPs per case.\n## CONCLUSIONS AND RELEVANCE\nArtificial intelligence assistance was associated with improved diagnoses by PCPs and NPs for 1 in every 8 to 10 cases, indicating potential for improving the quality of dermatologic care.\n",
        "methods": "### Methods\n\nThis study was approved by the Quorum Institutional Review Board, Seattle, Washington, and deemed exempt from informed consent because all data and images were deidentified. The [Standards for Reporting of Diagnostic Accuracy (STARD)[25]](http://www.equator-network.org/reporting-guidelines/stard/) reporting guideline was followed for this study.\n\n#### The AI Tool\nLiu et al[26] previously described an AI algorithm that provides a differential diagnosis given clinical photographs of skin conditions and the medical history (eTable 1 in the Supplement). Their AI model was developed using 16 114 cases and used a convolutional neural network to output prediction scores across 419 skin conditions. In the present study, we created a web-based tool using the AI model described by Liu et al by incorporating user experience insights (Figure 1).\n\nThe tool provides information about the case, including demographic information, history of present illness, and other elements of the patient\u2019s medical history. For each case, 1 to 6 images were available for review (median, 4), and readers could toggle between or zoom in on images. Primary care physicians and NPs reviewed these cases using a laptop and could consult additional resources as they would in clinical practice.\n\nThe AI assistance component of the web-based tool was only available during the assisted mode of the study (described below). At the top of the panel, the interface displayed the skin conditions that were output by the AI, sorted in order of the AI\u2019s predicted likelihood scores. Artificial intelligence predictions with low scores (<0.05) were removed, and the list was limited to 5 skin conditions to avoid presenting extraneous information. Each condition could be clicked on to display additional information (Figure 1 and eFigure 1 and the AI Tool Interface section in the eMethods in the Supplement).\n\n#### Study Design\nTo evaluate whether this tool could assist primary care clinicians in diagnosing skin conditions, we conducted a multiple-reader, multiple-case diagnostic study with 20 PCPs and 20 NPs (Figure 1). The characteristics of the clinicians are described in the Reader Characteristics section of the eMethods and eFigures 2 and 3 in the Supplement. Before reviewing the study cases, each reader was presented with materials describing how to use the AI assistant and given the opportunity to practice using the AI assistant with 2 sample cases (independent of the study cases). Additional details of this training[27] can be found in the Onboarding Process section in the eMethods in the Supplement.\n\nThe study used cases from 2 retrospective data sets from California and Hawaii previously used to validate the AI algorithm.[26] Specifically, the prior study used a validation set A and a subset (validation set B) enriched for rarer conditions via random sampling stratified by condition. Validation set B (963 cases) was included in its entirety. From validation set A, all 85 cases for which biopsy results were available were also included to yield a total of 1048 cases (Table). None of the PCPs or NPs in this study previously reviewed these cases, and the AI algorithm used was identical to the one used in the previous study.[26]\n\nEach reader was randomly assigned to 1 of 2 reader cohorts. The 2 reader cohorts read the same cases but with the opposite assistance modalities (ie, unassisted vs AI assisted) for each case. To reduce effects associated with switching modalities, the 1048 cases were divided into batches of 50 cases (except the last 48 cases, which were divided into 2 batches of 24 cases), and the assistance modality switched after each batch of cases. For the first batch of 50 cases, reader cohort 1 reviewed these cases with AI assistance, whereas reader cohort 2 reviewed the same cases unassisted. The next batch of cases were reviewed in the opposite modality (Figure 1). By ensuring each reader reviewed each case only once in either the assisted or unassisted modality, this design eliminated any memory effect associated with a crossover study (where memorable cases may inflate the diagnostic performance when reviewed a second time by the same readers).\n\nDuring the case reviews, the readers either provided their top differential diagnoses or indicated that they were unable to diagnose a case. They also answered a few questions on their intended clinical next steps for each case (see the Study End Points section below). Reviews were performed without time constraint. These reviews occurred from February 21 to April 28, 2020.\n\n#### Reference Diagnoses\nReference diagnoses were provided by a panel of dermatologists.[26] Briefly, 3 US board-certified dermatologists (from a pool of 12) independently reviewed each case. The dermatologists participated in the study via Advanced Clinical, Deerfield, Illinois; had 5 to 13 years of experience (mean [SD], 7.2 [2.7] years); and practiced in multiple states, including Colorado, Hawaii, Iowa, Maryland, New York, South Carolina, Tennessee, and Texas. Reference diagnoses were obtained using a previously described collective intelligence approach, which results in more reproducible diagnoses than diagnoses obtained by individual dermatologist review (eTable 2 in the Supplement). This approach assigns a vote to each diagnosis based on its ranking: the first diagnosis in a dermatologist\u2019s differential was given a weight of 1/1 = 1; the secondary diagnosis was given a weight of 1/2 = 0.5. The votes for each diagnosis were summed across the 3 dermatologists, and the top-voted diagnosis was considered the primary diagnosis of the panel.\n\nAgreement was also assessed against biopsy-confirmed diagnoses when available. Diagnoses were extracted from pathology reports by the teledermatology service before transfer to study investigators. These diagnoses were then mapped to skin conditions by US board-certified dermatologists (including K.K. and S.J.H.). The case distribution across these diagnoses (both clinical and histologic) are presented in the Table; of 152 cases with available biopsy results, the diagnosis of 141 cases was growths.\n\n#### Study End Points\nOur study was designed to evaluate 2 prespecified primary end points: (1) the agreement rate of the primary differential diagnosis of the PCPs with the reference diagnosis and (2) the agreement rate of the primary differential diagnosis of the NPs with the reference diagnosis. Based on the relative frequencies of conditions in this data set, the chance agreement is 3.77%.\n\nSeveral secondary analyses were planned. First, for cases with biopsy results, diagnoses were classified as malignant, precancerous, or benign and were evaluated against biopsy-determined diagnoses. Clinicians were also asked to report whether they would have recommended a biopsy or referred the case to a dermatologist. For the subset of reads in which clinicians reported they would not opt for a referral, we assessed the diagnostic agreement rate. We also analyzed the time taken to review cases and self-reported diagnostic confidence.\n\nFinally, 2 additional metrics (top-3 agreement and average overlap) were used for more comprehensive evaluation of cases in which additional follow-up may be needed to arrive at a definitive diagnosis (Additional Evaluation Metrics section in the eMethods in the Supplement). An exploratory analysis also measured the effect of AI assistance on dermatologist agreement with reference diagnoses.\n\n#### Statistical Analysis\nData were analyzed from May 26, 2020, to January 27, 2021. To compare clinicians reviewing cases with AI assistance and reviewing cases without, we used a permutation test with 1000 iterations. In each iteration, we permuted the assignment of whether reads were assisted or unassisted (ie, one-half of the full set of assisted and unassisted reads per case were selected to be assisted and the other half unassisted). Sensitivity analysis using a permutation test that preserved the reader cohorts\u2019 structure and another statistical analysis via a generalized linear mixed model produced similar results (see the Alternative Statistical Analyses section in the eMethods in the Supplement).\n\nBecause this study had 2 prespecified primary end points (both 1-tailed superiority tests), we applied the Bonferroni correction, and P < .0125 was considered statistically significant (halved from \u03b1 = 0.05 owing to 1-tailed tests and halved again owing to having 2 primary end points). Confidence intervals were computed by bootstrapping across both cases and readers for each sampled case (1000 iterations; sampling both cases and reader with replacement in each iteration). Hypothesis tests were conducted in Python, version 3.6.7 (Python Software Foundation).",
        "results": "### Results\n\nThis study involved the participation of 40 board-certified clinicians, including 20 PCPs (14 women [70.0%] and 6 men [30.0%]; mean experience, 11.3 [range, 2-32] years) who were located across 12 states and 20 NPs (18 women [90.0%] and 2 men [10.0%]; mean experience, 13.1 [range, 2-34] years) who practiced in primary care without physician supervision across 9 states. These clinicians reviewed 1048 teledermatology cases (672 women [64.2%] and 375 men [35.8%], with 1 missing; median age, 43 [interquartile range, 30-56] years) from 11 sites (Table) and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]), for a total of 41 920 case reviews. Every PCP and NP reviewed each case only once, either with or without AI assistance (Figure 1).\n\nArtificial intelligence assistance was associated with significantly higher top-1 agreement with the reference diagnosis (Figure 2A and eTable 3 in the Supplement). For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P < .001), from 48% to 58%; for NPs, the improvement was 12% (95% CI, 10%-14%; P < .001), from 46% to 58%. Assistance was associated with improvements for all 40 readers, although the magnitude varied by reader (range, 2%-22%; median, 10%) (Figure 2B). Similar improvements were observed beyond the primary diagnosis based on the top-3 agreement, average overlap, per-condition sensitivity, and \u03ba value (eFigures 4 and 5 and eTable 3 in the Supplement). In an exploratory analysis, 2 dermatologists\u2019 agreement with the reference diagnosis remained largely unchanged with AI assistance, increasing by 2% (95% CI, \u22121% to 5%), from 63% to 66% (eFigures 4 and 5 in the Supplement).\n\nFor cases with available biopsy diagnoses (n = 141), the readers\u2019 accuracy at classifying lesions as malignant, precancerous, or benign trended upward by 3% for PCPs (95% CI, \u22121% to 7%) from 64% to 67% and by 8% for NPs (95% CI, 3%-13%) from 60% to 68% (Figure 2C-D). Subgroup analysis further found that sensitivity for malignant lesions, precancerous lesions, infectious skin diseases, and categories of hair loss trended upward or remained similar with assistance for both NPs and PCPs, with improvements ranging from \u22121% to 36% (eTable 4 in the Supplement).\n\nOn the subset of cases in which the top prediction of AI was accurate (63% of cases), the use of assistance was associated with an increased top-1 agreement with reference diagnosis of 18% (95% CI, 16%-20%) for PCPs and 21% (95% CI, 19%-23%) for NPs. On the contrary, when none of the AI tool\u2019s predictions was correct (13% of cases), the agreement was 8% lower (95% CI, 5%-12%) for PCPs and 9% lower (95% CI, 6%-12%) for NPs. The effects were intermediate when the correct diagnosis was in the second or third position instead of the first (see the Impact of AI Accuracy on Assistance section of eMethods and eFigures 6 and 7 in the Supplement). An exploratory analysis also suggested that assistance was particularly beneficial for less ambiguous cases. For example, in the subset of cases in which the dermatologist panel had unanimous agreement, the use of AI assistance was associated with a top-1 agreement increase of 13% (95% CI, 10%-15%) for PCPs and of 16% (95% CI, 14%-19%) for NPs (eFigure 8 in the Supplement). Subanalyses also indicated that assistance-associated benefits were consistent during the study and across several skin types (eFigures 9 and 10 in the Supplement).\n\nArtificial intelligence assistance was also associated with changes in several simulated clinical decisions (Figure 3A-B). The rates of indicating a need for biopsy were 1% lower (95% CI, 0%-3%) for PCPs and 2% lower (95% CI, 1%-3%) for NPs; the rate of desire was 3% lower (95% CI, 1%-4%) for both PCPs and NPs (eTable 5 in the Supplement). For cases in which readers indicated referrals were unnecessary, their top-1 agreement rate with dermatologists was higher by 10% for PCPs (95% CI, 8%-12%), from 51% to 61%, and by 12% for NPs (95% CI, 10%-14%), from 47% to 59%, with a similar effect on referred cases (Figure 3C-D and eFigure 11 in the Supplement). In related findings, self-reported diagnostic confidence was substantially higher with AI assistance for both reader cohorts (Figure 4A). The top-1 agreement of cases that were rated with more than 90% confidence was substantially higher (73% vs 64% for PCPs and 68% vs 58% for NPs) (eFigure 12 in the Supplement).\n\nIn terms of review time per case, AI assistance was associated with a slightly increased median review time. A difference of 5 (95% CI, 0-8) seconds, from 89 to 94 seconds, was observed for PCPs and a difference of 7 (95% CI, 5-10) seconds, from 77 to 84 seconds, was observed for NPs (Figure 4B and eFigure 9D in the Supplement). We also present representative examples of cases in which AI assistance was associated with the largest increases or decreases in agreement with reference diagnoses (eFigures 13 and 14 in the Supplement) and results of follow-up surveys investigating the usefulness of various AI assistant features (eFigures 15-17 in the Supplement).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices.",
        "abstract": "## IMPORTANCE\nMost dermatologic cases are initially evaluated by nondermatologists such as primary care physicians (PCPs) or nurse practitioners (NPs).\n## OBJECTIVE\nTo evaluate an artificial intelligence (AI)-based tool that assists with diagnoses of dermatologic conditions.\n## DESIGN, SETTING, AND PARTICIPANTS\nThis multiple-reader, multiple-case diagnostic study developed an AI-based tool and evaluated its utility. Primary care physicians and NPs retrospectively reviewed an enriched set of cases representing 120 different skin conditions. Randomization was used to ensure each clinician reviewed each case either with or without AI assistance; each clinician alternated between batches of 50 cases in each modality. The reviews occurred from February 21 to April 28, 2020. Data were analyzed from May 26, 2020, to January 27, 2021.\n## EXPOSURES\nAn AI-based assistive tool for interpreting clinical images and associated medical history.\n## MAIN OUTCOMES AND MEASURES\nThe primary analysis evaluated agreement with reference diagnoses provided by a panel of 3 dermatologists for PCPs and NPs. Secondary analyses included diagnostic accuracy for biopsy-confirmed cases, biopsy and referral rates, review time, and diagnostic confidence.\n## RESULTS\nForty board-certified clinicians, including 20 PCPs (14 women [70.0%]; mean experience, 11.3 [range, 2-32] years) and 20 NPs (18 women [90.0%]; mean experience, 13.1 [range, 2-34] years) reviewed 1048 retrospective cases (672 female [64.2%]; median age, 43 [interquartile range, 30-56] years; 41\u202f920 total reviews) from a teledermatology practice serving 11 sites and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]). The PCPs were located across 12 states, and the NPs practiced in primary care without physician supervision across 9 states. The NPs had a mean of 13.1 (range, 2-34) years of experience and practiced in primary care without physician supervision across 9 states. Artificial intelligence assistance was significantly associated with higher agreement with reference diagnoses. For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P\u2009<\u2009.001), from 48% to 58%; for NPs, the increase was 12% (95% CI, 10%-14%; P\u2009<\u2009.001), from 46% to 58%. In secondary analyses, agreement with biopsy-obtained diagnosis categories of maglignant, precancerous, or benign increased by 3% (95% CI, -1% to 7%) for PCPs and by 8% (95% CI, 3%-13%) for NPs. Rates of desire for biopsies decreased by 1% (95% CI, 0-3%) for PCPs and 2% (95% CI, 1%-3%) for NPs; the rate of desire for referrals decreased by 3% (95% CI, 1%-4%) for PCPs and NPs. Diagnostic agreement on cases not indicated for a dermatologist referral increased by 10% (95% CI, 8%-12%) for PCPs and 12% (95% CI, 10%-14%) for NPs, and median review time increased slightly by 5 (95% CI, 0-8) seconds for PCPs and 7 (95% CI, 5-10) seconds for NPs per case.\n## CONCLUSIONS AND RELEVANCE\nArtificial intelligence assistance was associated with improved diagnoses by PCPs and NPs for 1 in every 8 to 10 cases, indicating potential for improving the quality of dermatologic care.\n",
        "methods": "### Methods\n\nThis study was approved by the Quorum Institutional Review Board, Seattle, Washington, and deemed exempt from informed consent because all data and images were deidentified. The [Standards for Reporting of Diagnostic Accuracy (STARD)[25]](http://www.equator-network.org/reporting-guidelines/stard/) reporting guideline was followed for this study.\n\n#### The AI Tool\nLiu et al[26] previously described an AI algorithm that provides a differential diagnosis given clinical photographs of skin conditions and the medical history (eTable 1 in the Supplement). Their AI model was developed using 16 114 cases and used a convolutional neural network to output prediction scores across 419 skin conditions. In the present study, we created a web-based tool using the AI model described by Liu et al by incorporating user experience insights (Figure 1).\n\nThe tool provides information about the case, including demographic information, history of present illness, and other elements of the patient\u2019s medical history. For each case, 1 to 6 images were available for review (median, 4), and readers could toggle between or zoom in on images. Primary care physicians and NPs reviewed these cases using a laptop and could consult additional resources as they would in clinical practice.\n\nThe AI assistance component of the web-based tool was only available during the assisted mode of the study (described below). At the top of the panel, the interface displayed the skin conditions that were output by the AI, sorted in order of the AI\u2019s predicted likelihood scores. Artificial intelligence predictions with low scores (<0.05) were removed, and the list was limited to 5 skin conditions to avoid presenting extraneous information. Each condition could be clicked on to display additional information (Figure 1 and eFigure 1 and the AI Tool Interface section in the eMethods in the Supplement).\n\n#### Study Design\nTo evaluate whether this tool could assist primary care clinicians in diagnosing skin conditions, we conducted a multiple-reader, multiple-case diagnostic study with 20 PCPs and 20 NPs (Figure 1). The characteristics of the clinicians are described in the Reader Characteristics section of the eMethods and eFigures 2 and 3 in the Supplement. Before reviewing the study cases, each reader was presented with materials describing how to use the AI assistant and given the opportunity to practice using the AI assistant with 2 sample cases (independent of the study cases). Additional details of this training[27] can be found in the Onboarding Process section in the eMethods in the Supplement.\n\nThe study used cases from 2 retrospective data sets from California and Hawaii previously used to validate the AI algorithm.[26] Specifically, the prior study used a validation set A and a subset (validation set B) enriched for rarer conditions via random sampling stratified by condition. Validation set B (963 cases) was included in its entirety. From validation set A, all 85 cases for which biopsy results were available were also included to yield a total of 1048 cases (Table). None of the PCPs or NPs in this study previously reviewed these cases, and the AI algorithm used was identical to the one used in the previous study.[26]\n\nEach reader was randomly assigned to 1 of 2 reader cohorts. The 2 reader cohorts read the same cases but with the opposite assistance modalities (ie, unassisted vs AI assisted) for each case. To reduce effects associated with switching modalities, the 1048 cases were divided into batches of 50 cases (except the last 48 cases, which were divided into 2 batches of 24 cases), and the assistance modality switched after each batch of cases. For the first batch of 50 cases, reader cohort 1 reviewed these cases with AI assistance, whereas reader cohort 2 reviewed the same cases unassisted. The next batch of cases were reviewed in the opposite modality (Figure 1). By ensuring each reader reviewed each case only once in either the assisted or unassisted modality, this design eliminated any memory effect associated with a crossover study (where memorable cases may inflate the diagnostic performance when reviewed a second time by the same readers).\n\nDuring the case reviews, the readers either provided their top differential diagnoses or indicated that they were unable to diagnose a case. They also answered a few questions on their intended clinical next steps for each case (see the Study End Points section below). Reviews were performed without time constraint. These reviews occurred from February 21 to April 28, 2020.\n\n#### Reference Diagnoses\nReference diagnoses were provided by a panel of dermatologists.[26] Briefly, 3 US board-certified dermatologists (from a pool of 12) independently reviewed each case. The dermatologists participated in the study via Advanced Clinical, Deerfield, Illinois; had 5 to 13 years of experience (mean [SD], 7.2 [2.7] years); and practiced in multiple states, including Colorado, Hawaii, Iowa, Maryland, New York, South Carolina, Tennessee, and Texas. Reference diagnoses were obtained using a previously described collective intelligence approach, which results in more reproducible diagnoses than diagnoses obtained by individual dermatologist review (eTable 2 in the Supplement). This approach assigns a vote to each diagnosis based on its ranking: the first diagnosis in a dermatologist\u2019s differential was given a weight of 1/1 = 1; the secondary diagnosis was given a weight of 1/2 = 0.5. The votes for each diagnosis were summed across the 3 dermatologists, and the top-voted diagnosis was considered the primary diagnosis of the panel.\n\nAgreement was also assessed against biopsy-confirmed diagnoses when available. Diagnoses were extracted from pathology reports by the teledermatology service before transfer to study investigators. These diagnoses were then mapped to skin conditions by US board-certified dermatologists (including K.K. and S.J.H.). The case distribution across these diagnoses (both clinical and histologic) are presented in the Table; of 152 cases with available biopsy results, the diagnosis of 141 cases was growths.\n\n#### Study End Points\nOur study was designed to evaluate 2 prespecified primary end points: (1) the agreement rate of the primary differential diagnosis of the PCPs with the reference diagnosis and (2) the agreement rate of the primary differential diagnosis of the NPs with the reference diagnosis. Based on the relative frequencies of conditions in this data set, the chance agreement is 3.77%.\n\nSeveral secondary analyses were planned. First, for cases with biopsy results, diagnoses were classified as malignant, precancerous, or benign and were evaluated against biopsy-determined diagnoses. Clinicians were also asked to report whether they would have recommended a biopsy or referred the case to a dermatologist. For the subset of reads in which clinicians reported they would not opt for a referral, we assessed the diagnostic agreement rate. We also analyzed the time taken to review cases and self-reported diagnostic confidence.\n\nFinally, 2 additional metrics (top-3 agreement and average overlap) were used for more comprehensive evaluation of cases in which additional follow-up may be needed to arrive at a definitive diagnosis (Additional Evaluation Metrics section in the eMethods in the Supplement). An exploratory analysis also measured the effect of AI assistance on dermatologist agreement with reference diagnoses.\n\n#### Statistical Analysis\nData were analyzed from May 26, 2020, to January 27, 2021. To compare clinicians reviewing cases with AI assistance and reviewing cases without, we used a permutation test with 1000 iterations. In each iteration, we permuted the assignment of whether reads were assisted or unassisted (ie, one-half of the full set of assisted and unassisted reads per case were selected to be assisted and the other half unassisted). Sensitivity analysis using a permutation test that preserved the reader cohorts\u2019 structure and another statistical analysis via a generalized linear mixed model produced similar results (see the Alternative Statistical Analyses section in the eMethods in the Supplement).\n\nBecause this study had 2 prespecified primary end points (both 1-tailed superiority tests), we applied the Bonferroni correction, and P < .0125 was considered statistically significant (halved from \u03b1 = 0.05 owing to 1-tailed tests and halved again owing to having 2 primary end points). Confidence intervals were computed by bootstrapping across both cases and readers for each sampled case (1000 iterations; sampling both cases and reader with replacement in each iteration). Hypothesis tests were conducted in Python, version 3.6.7 (Python Software Foundation).",
        "results": "### Results\n\nThis study involved the participation of 40 board-certified clinicians, including 20 PCPs (14 women [70.0%] and 6 men [30.0%]; mean experience, 11.3 [range, 2-32] years) who were located across 12 states and 20 NPs (18 women [90.0%] and 2 men [10.0%]; mean experience, 13.1 [range, 2-34] years) who practiced in primary care without physician supervision across 9 states. These clinicians reviewed 1048 teledermatology cases (672 women [64.2%] and 375 men [35.8%], with 1 missing; median age, 43 [interquartile range, 30-56] years) from 11 sites (Table) and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]), for a total of 41 920 case reviews. Every PCP and NP reviewed each case only once, either with or without AI assistance (Figure 1).\n\nArtificial intelligence assistance was associated with significantly higher top-1 agreement with the reference diagnosis (Figure 2A and eTable 3 in the Supplement). For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P < .001), from 48% to 58%; for NPs, the improvement was 12% (95% CI, 10%-14%; P < .001), from 46% to 58%. Assistance was associated with improvements for all 40 readers, although the magnitude varied by reader (range, 2%-22%; median, 10%) (Figure 2B). Similar improvements were observed beyond the primary diagnosis based on the top-3 agreement, average overlap, per-condition sensitivity, and \u03ba value (eFigures 4 and 5 and eTable 3 in the Supplement). In an exploratory analysis, 2 dermatologists\u2019 agreement with the reference diagnosis remained largely unchanged with AI assistance, increasing by 2% (95% CI, \u22121% to 5%), from 63% to 66% (eFigures 4 and 5 in the Supplement).\n\nFor cases with available biopsy diagnoses (n = 141), the readers\u2019 accuracy at classifying lesions as malignant, precancerous, or benign trended upward by 3% for PCPs (95% CI, \u22121% to 7%) from 64% to 67% and by 8% for NPs (95% CI, 3%-13%) from 60% to 68% (Figure 2C-D). Subgroup analysis further found that sensitivity for malignant lesions, precancerous lesions, infectious skin diseases, and categories of hair loss trended upward or remained similar with assistance for both NPs and PCPs, with improvements ranging from \u22121% to 36% (eTable 4 in the Supplement).\n\nOn the subset of cases in which the top prediction of AI was accurate (63% of cases), the use of assistance was associated with an increased top-1 agreement with reference diagnosis of 18% (95% CI, 16%-20%) for PCPs and 21% (95% CI, 19%-23%) for NPs. On the contrary, when none of the AI tool\u2019s predictions was correct (13% of cases), the agreement was 8% lower (95% CI, 5%-12%) for PCPs and 9% lower (95% CI, 6%-12%) for NPs. The effects were intermediate when the correct diagnosis was in the second or third position instead of the first (see the Impact of AI Accuracy on Assistance section of eMethods and eFigures 6 and 7 in the Supplement). An exploratory analysis also suggested that assistance was particularly beneficial for less ambiguous cases. For example, in the subset of cases in which the dermatologist panel had unanimous agreement, the use of AI assistance was associated with a top-1 agreement increase of 13% (95% CI, 10%-15%) for PCPs and of 16% (95% CI, 14%-19%) for NPs (eFigure 8 in the Supplement). Subanalyses also indicated that assistance-associated benefits were consistent during the study and across several skin types (eFigures 9 and 10 in the Supplement).\n\nArtificial intelligence assistance was also associated with changes in several simulated clinical decisions (Figure 3A-B). The rates of indicating a need for biopsy were 1% lower (95% CI, 0%-3%) for PCPs and 2% lower (95% CI, 1%-3%) for NPs; the rate of desire was 3% lower (95% CI, 1%-4%) for both PCPs and NPs (eTable 5 in the Supplement). For cases in which readers indicated referrals were unnecessary, their top-1 agreement rate with dermatologists was higher by 10% for PCPs (95% CI, 8%-12%), from 51% to 61%, and by 12% for NPs (95% CI, 10%-14%), from 47% to 59%, with a similar effect on referred cases (Figure 3C-D and eFigure 11 in the Supplement). In related findings, self-reported diagnostic confidence was substantially higher with AI assistance for both reader cohorts (Figure 4A). The top-1 agreement of cases that were rated with more than 90% confidence was substantially higher (73% vs 64% for PCPs and 68% vs 58% for NPs) (eFigure 12 in the Supplement).\n\nIn terms of review time per case, AI assistance was associated with a slightly increased median review time. A difference of 5 (95% CI, 0-8) seconds, from 89 to 94 seconds, was observed for PCPs and a difference of 7 (95% CI, 5-10) seconds, from 77 to 84 seconds, was observed for NPs (Figure 4B and eFigure 9D in the Supplement). We also present representative examples of cases in which AI assistance was associated with the largest increases or decreases in agreement with reference diagnoses (eFigures 13 and 14 in the Supplement) and results of follow-up surveys investigating the usefulness of various AI assistant features (eFigures 15-17 in the Supplement).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Development and Assessment of an Artificial Intelligence-Based Tool for Skin Condition Diagnosis by Primary Care Physicians and Nurse Practitioners in Teledermatology Practices.",
        "abstract": "## IMPORTANCE\nMost dermatologic cases are initially evaluated by nondermatologists such as primary care physicians (PCPs) or nurse practitioners (NPs).\n## OBJECTIVE\nTo evaluate an artificial intelligence (AI)-based tool that assists with diagnoses of dermatologic conditions.\n## DESIGN, SETTING, AND PARTICIPANTS\nThis multiple-reader, multiple-case diagnostic study developed an AI-based tool and evaluated its utility. Primary care physicians and NPs retrospectively reviewed an enriched set of cases representing 120 different skin conditions. Randomization was used to ensure each clinician reviewed each case either with or without AI assistance; each clinician alternated between batches of 50 cases in each modality. The reviews occurred from February 21 to April 28, 2020. Data were analyzed from May 26, 2020, to January 27, 2021.\n## EXPOSURES\nAn AI-based assistive tool for interpreting clinical images and associated medical history.\n## MAIN OUTCOMES AND MEASURES\nThe primary analysis evaluated agreement with reference diagnoses provided by a panel of 3 dermatologists for PCPs and NPs. Secondary analyses included diagnostic accuracy for biopsy-confirmed cases, biopsy and referral rates, review time, and diagnostic confidence.\n## RESULTS\nForty board-certified clinicians, including 20 PCPs (14 women [70.0%]; mean experience, 11.3 [range, 2-32] years) and 20 NPs (18 women [90.0%]; mean experience, 13.1 [range, 2-34] years) reviewed 1048 retrospective cases (672 female [64.2%]; median age, 43 [interquartile range, 30-56] years; 41\u202f920 total reviews) from a teledermatology practice serving 11 sites and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]). The PCPs were located across 12 states, and the NPs practiced in primary care without physician supervision across 9 states. The NPs had a mean of 13.1 (range, 2-34) years of experience and practiced in primary care without physician supervision across 9 states. Artificial intelligence assistance was significantly associated with higher agreement with reference diagnoses. For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P\u2009<\u2009.001), from 48% to 58%; for NPs, the increase was 12% (95% CI, 10%-14%; P\u2009<\u2009.001), from 46% to 58%. In secondary analyses, agreement with biopsy-obtained diagnosis categories of maglignant, precancerous, or benign increased by 3% (95% CI, -1% to 7%) for PCPs and by 8% (95% CI, 3%-13%) for NPs. Rates of desire for biopsies decreased by 1% (95% CI, 0-3%) for PCPs and 2% (95% CI, 1%-3%) for NPs; the rate of desire for referrals decreased by 3% (95% CI, 1%-4%) for PCPs and NPs. Diagnostic agreement on cases not indicated for a dermatologist referral increased by 10% (95% CI, 8%-12%) for PCPs and 12% (95% CI, 10%-14%) for NPs, and median review time increased slightly by 5 (95% CI, 0-8) seconds for PCPs and 7 (95% CI, 5-10) seconds for NPs per case.\n## CONCLUSIONS AND RELEVANCE\nArtificial intelligence assistance was associated with improved diagnoses by PCPs and NPs for 1 in every 8 to 10 cases, indicating potential for improving the quality of dermatologic care.\n",
        "methods": "### Methods\n\nThis study was approved by the Quorum Institutional Review Board, Seattle, Washington, and deemed exempt from informed consent because all data and images were deidentified. The [Standards for Reporting of Diagnostic Accuracy (STARD)[25]](http://www.equator-network.org/reporting-guidelines/stard/) reporting guideline was followed for this study.\n\n#### The AI Tool\nLiu et al[26] previously described an AI algorithm that provides a differential diagnosis given clinical photographs of skin conditions and the medical history (eTable 1 in the Supplement). Their AI model was developed using 16 114 cases and used a convolutional neural network to output prediction scores across 419 skin conditions. In the present study, we created a web-based tool using the AI model described by Liu et al by incorporating user experience insights (Figure 1).\n\nThe tool provides information about the case, including demographic information, history of present illness, and other elements of the patient\u2019s medical history. For each case, 1 to 6 images were available for review (median, 4), and readers could toggle between or zoom in on images. Primary care physicians and NPs reviewed these cases using a laptop and could consult additional resources as they would in clinical practice.\n\nThe AI assistance component of the web-based tool was only available during the assisted mode of the study (described below). At the top of the panel, the interface displayed the skin conditions that were output by the AI, sorted in order of the AI\u2019s predicted likelihood scores. Artificial intelligence predictions with low scores (<0.05) were removed, and the list was limited to 5 skin conditions to avoid presenting extraneous information. Each condition could be clicked on to display additional information (Figure 1 and eFigure 1 and the AI Tool Interface section in the eMethods in the Supplement).\n\n#### Study Design\nTo evaluate whether this tool could assist primary care clinicians in diagnosing skin conditions, we conducted a multiple-reader, multiple-case diagnostic study with 20 PCPs and 20 NPs (Figure 1). The characteristics of the clinicians are described in the Reader Characteristics section of the eMethods and eFigures 2 and 3 in the Supplement. Before reviewing the study cases, each reader was presented with materials describing how to use the AI assistant and given the opportunity to practice using the AI assistant with 2 sample cases (independent of the study cases). Additional details of this training[27] can be found in the Onboarding Process section in the eMethods in the Supplement.\n\nThe study used cases from 2 retrospective data sets from California and Hawaii previously used to validate the AI algorithm.[26] Specifically, the prior study used a validation set A and a subset (validation set B) enriched for rarer conditions via random sampling stratified by condition. Validation set B (963 cases) was included in its entirety. From validation set A, all 85 cases for which biopsy results were available were also included to yield a total of 1048 cases (Table). None of the PCPs or NPs in this study previously reviewed these cases, and the AI algorithm used was identical to the one used in the previous study.[26]\n\nEach reader was randomly assigned to 1 of 2 reader cohorts. The 2 reader cohorts read the same cases but with the opposite assistance modalities (ie, unassisted vs AI assisted) for each case. To reduce effects associated with switching modalities, the 1048 cases were divided into batches of 50 cases (except the last 48 cases, which were divided into 2 batches of 24 cases), and the assistance modality switched after each batch of cases. For the first batch of 50 cases, reader cohort 1 reviewed these cases with AI assistance, whereas reader cohort 2 reviewed the same cases unassisted. The next batch of cases were reviewed in the opposite modality (Figure 1). By ensuring each reader reviewed each case only once in either the assisted or unassisted modality, this design eliminated any memory effect associated with a crossover study (where memorable cases may inflate the diagnostic performance when reviewed a second time by the same readers).\n\nDuring the case reviews, the readers either provided their top differential diagnoses or indicated that they were unable to diagnose a case. They also answered a few questions on their intended clinical next steps for each case (see the Study End Points section below). Reviews were performed without time constraint. These reviews occurred from February 21 to April 28, 2020.\n\n#### Reference Diagnoses\nReference diagnoses were provided by a panel of dermatologists.[26] Briefly, 3 US board-certified dermatologists (from a pool of 12) independently reviewed each case. The dermatologists participated in the study via Advanced Clinical, Deerfield, Illinois; had 5 to 13 years of experience (mean [SD], 7.2 [2.7] years); and practiced in multiple states, including Colorado, Hawaii, Iowa, Maryland, New York, South Carolina, Tennessee, and Texas. Reference diagnoses were obtained using a previously described collective intelligence approach, which results in more reproducible diagnoses than diagnoses obtained by individual dermatologist review (eTable 2 in the Supplement). This approach assigns a vote to each diagnosis based on its ranking: the first diagnosis in a dermatologist\u2019s differential was given a weight of 1/1 = 1; the secondary diagnosis was given a weight of 1/2 = 0.5. The votes for each diagnosis were summed across the 3 dermatologists, and the top-voted diagnosis was considered the primary diagnosis of the panel.\n\nAgreement was also assessed against biopsy-confirmed diagnoses when available. Diagnoses were extracted from pathology reports by the teledermatology service before transfer to study investigators. These diagnoses were then mapped to skin conditions by US board-certified dermatologists (including K.K. and S.J.H.). The case distribution across these diagnoses (both clinical and histologic) are presented in the Table; of 152 cases with available biopsy results, the diagnosis of 141 cases was growths.\n\n#### Study End Points\nOur study was designed to evaluate 2 prespecified primary end points: (1) the agreement rate of the primary differential diagnosis of the PCPs with the reference diagnosis and (2) the agreement rate of the primary differential diagnosis of the NPs with the reference diagnosis. Based on the relative frequencies of conditions in this data set, the chance agreement is 3.77%.\n\nSeveral secondary analyses were planned. First, for cases with biopsy results, diagnoses were classified as malignant, precancerous, or benign and were evaluated against biopsy-determined diagnoses. Clinicians were also asked to report whether they would have recommended a biopsy or referred the case to a dermatologist. For the subset of reads in which clinicians reported they would not opt for a referral, we assessed the diagnostic agreement rate. We also analyzed the time taken to review cases and self-reported diagnostic confidence.\n\nFinally, 2 additional metrics (top-3 agreement and average overlap) were used for more comprehensive evaluation of cases in which additional follow-up may be needed to arrive at a definitive diagnosis (Additional Evaluation Metrics section in the eMethods in the Supplement). An exploratory analysis also measured the effect of AI assistance on dermatologist agreement with reference diagnoses.\n\n#### Statistical Analysis\nData were analyzed from May 26, 2020, to January 27, 2021. To compare clinicians reviewing cases with AI assistance and reviewing cases without, we used a permutation test with 1000 iterations. In each iteration, we permuted the assignment of whether reads were assisted or unassisted (ie, one-half of the full set of assisted and unassisted reads per case were selected to be assisted and the other half unassisted). Sensitivity analysis using a permutation test that preserved the reader cohorts\u2019 structure and another statistical analysis via a generalized linear mixed model produced similar results (see the Alternative Statistical Analyses section in the eMethods in the Supplement).\n\nBecause this study had 2 prespecified primary end points (both 1-tailed superiority tests), we applied the Bonferroni correction, and P < .0125 was considered statistically significant (halved from \u03b1 = 0.05 owing to 1-tailed tests and halved again owing to having 2 primary end points). Confidence intervals were computed by bootstrapping across both cases and readers for each sampled case (1000 iterations; sampling both cases and reader with replacement in each iteration). Hypothesis tests were conducted in Python, version 3.6.7 (Python Software Foundation).",
        "results": "### Results\n\nThis study involved the participation of 40 board-certified clinicians, including 20 PCPs (14 women [70.0%] and 6 men [30.0%]; mean experience, 11.3 [range, 2-32] years) who were located across 12 states and 20 NPs (18 women [90.0%] and 2 men [10.0%]; mean experience, 13.1 [range, 2-34] years) who practiced in primary care without physician supervision across 9 states. These clinicians reviewed 1048 teledermatology cases (672 women [64.2%] and 375 men [35.8%], with 1 missing; median age, 43 [interquartile range, 30-56] years) from 11 sites (Table) and provided 0 to 5 differential diagnoses per case (mean [SD], 1.6 [0.7]), for a total of 41 920 case reviews. Every PCP and NP reviewed each case only once, either with or without AI assistance (Figure 1).\n\nArtificial intelligence assistance was associated with significantly higher top-1 agreement with the reference diagnosis (Figure 2A and eTable 3 in the Supplement). For PCPs, the increase in diagnostic agreement was 10% (95% CI, 8%-11%; P < .001), from 48% to 58%; for NPs, the improvement was 12% (95% CI, 10%-14%; P < .001), from 46% to 58%. Assistance was associated with improvements for all 40 readers, although the magnitude varied by reader (range, 2%-22%; median, 10%) (Figure 2B). Similar improvements were observed beyond the primary diagnosis based on the top-3 agreement, average overlap, per-condition sensitivity, and \u03ba value (eFigures 4 and 5 and eTable 3 in the Supplement). In an exploratory analysis, 2 dermatologists\u2019 agreement with the reference diagnosis remained largely unchanged with AI assistance, increasing by 2% (95% CI, \u22121% to 5%), from 63% to 66% (eFigures 4 and 5 in the Supplement).\n\nFor cases with available biopsy diagnoses (n = 141), the readers\u2019 accuracy at classifying lesions as malignant, precancerous, or benign trended upward by 3% for PCPs (95% CI, \u22121% to 7%) from 64% to 67% and by 8% for NPs (95% CI, 3%-13%) from 60% to 68% (Figure 2C-D). Subgroup analysis further found that sensitivity for malignant lesions, precancerous lesions, infectious skin diseases, and categories of hair loss trended upward or remained similar with assistance for both NPs and PCPs, with improvements ranging from \u22121% to 36% (eTable 4 in the Supplement).\n\nOn the subset of cases in which the top prediction of AI was accurate (63% of cases), the use of assistance was associated with an increased top-1 agreement with reference diagnosis of 18% (95% CI, 16%-20%) for PCPs and 21% (95% CI, 19%-23%) for NPs. On the contrary, when none of the AI tool\u2019s predictions was correct (13% of cases), the agreement was 8% lower (95% CI, 5%-12%) for PCPs and 9% lower (95% CI, 6%-12%) for NPs. The effects were intermediate when the correct diagnosis was in the second or third position instead of the first (see the Impact of AI Accuracy on Assistance section of eMethods and eFigures 6 and 7 in the Supplement). An exploratory analysis also suggested that assistance was particularly beneficial for less ambiguous cases. For example, in the subset of cases in which the dermatologist panel had unanimous agreement, the use of AI assistance was associated with a top-1 agreement increase of 13% (95% CI, 10%-15%) for PCPs and of 16% (95% CI, 14%-19%) for NPs (eFigure 8 in the Supplement). Subanalyses also indicated that assistance-associated benefits were consistent during the study and across several skin types (eFigures 9 and 10 in the Supplement).\n\nArtificial intelligence assistance was also associated with changes in several simulated clinical decisions (Figure 3A-B). The rates of indicating a need for biopsy were 1% lower (95% CI, 0%-3%) for PCPs and 2% lower (95% CI, 1%-3%) for NPs; the rate of desire was 3% lower (95% CI, 1%-4%) for both PCPs and NPs (eTable 5 in the Supplement). For cases in which readers indicated referrals were unnecessary, their top-1 agreement rate with dermatologists was higher by 10% for PCPs (95% CI, 8%-12%), from 51% to 61%, and by 12% for NPs (95% CI, 10%-14%), from 47% to 59%, with a similar effect on referred cases (Figure 3C-D and eFigure 11 in the Supplement). In related findings, self-reported diagnostic confidence was substantially higher with AI assistance for both reader cohorts (Figure 4A). The top-1 agreement of cases that were rated with more than 90% confidence was substantially higher (73% vs 64% for PCPs and 68% vs 58% for NPs) (eFigure 12 in the Supplement).\n\nIn terms of review time per case, AI assistance was associated with a slightly increased median review time. A difference of 5 (95% CI, 0-8) seconds, from 89 to 94 seconds, was observed for PCPs and a difference of 7 (95% CI, 5-10) seconds, from 77 to 84 seconds, was observed for NPs (Figure 4B and eFigure 9D in the Supplement). We also present representative examples of cases in which AI assistance was associated with the largest increases or decreases in agreement with reference diagnoses (eFigures 13 and 14 in the Supplement) and results of follow-up surveys investigating the usefulness of various AI assistant features (eFigures 15-17 in the Supplement).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Performance of a deep neural network in teledermatology: a single-centre prospective diagnostic study.",
        "abstract": "## BACKGROUND\nThe use of artificial intelligence (AI) algorithms for the diagnosis of skin diseases has shown promise in experimental settings but has not been yet tested in real-life conditions.\n## OBJECTIVE\nTo assess the diagnostic performance and potential clinical utility of a 174-multiclass AI algorithm in a real-life telemedicine setting.\n## METHODS\nProspective, diagnostic accuracy study including consecutive patients who submitted images for teledermatology evaluation. The treating dermatologist chose a single image to upload to a web application during teleconsultation. A follow-up reader study including nine healthcare providers (3 dermatologists, 3 dermatology residents and 3 general practitioners) was performed.\n## RESULTS\nA total of 340 cases from 281 patients met study inclusion criteria. The mean (SD) age of patients was 33.7 (17.5) years; 63% (n\u00a0=\u00a0177) were female. Exposure to the AI algorithm results was considered useful in 11.8% of visits (n\u00a0=\u00a040) and the teledermatologist correctly modified the real-time diagnosis in 0.6% (n\u00a0=\u00a02) of cases. The overall top-1 accuracy of the algorithm (41.2%) was lower than that of the dermatologists (60.1%), residents (57.8%) and general practitioners (49.3%) (all comparisons P\u00a0<\u00a00.05, in the reader study). When the analysis was limited to the diagnoses on which the algorithm had been explicitly trained, the balanced top-1 accuracy of the algorithm (47.6%) was comparable to the dermatologists (49.7%) and residents (47.7%) but superior to the general practitioners (39.7%; P\u00a0=\u00a00.049). Algorithm performance was associated with patient skin type and image quality.\n## CONCLUSIONS\nA 174-disease class AI algorithm appears to be a promising tool in the triage and evaluation of lesions with patient-taken photographs via telemedicine.\n",
        "methods": "### Methods\n\nThis was an IRB-approved, single-centre, prospective, diagnostic accuracy study conducted between 27 March and 30 April 2020. All data collection was explicitly planned prior to the performance of the index tests and reference standard and the study was reported following the STARD 2015 guidelines. Consecutive patients who submitted clinical images of one or more new skin conditions for evaluation in our telemedicine dermatology clinic were considered eligible for inclusion. Both new and existing patients were eligible with no age restrictions. Cases were excluded if no images were submitted by patients, images were rated as insufficient quality, patients submitted images but did not present for telemedicine visit, patients had a telemedicine visit but did not subsequently adhere to the recommendation for an in-person visit or diagnostic testing, or upon evaluation, they had no skin condition present for examination.\n\n**Telemedicine clinical workflow**\n\nPatients submitted \u2265 1 clinical images of the skin condition to the consulting dermatologist prior to, or during the telemedicine visit. All patients received standardized imaging instructions, which provided lighting and flash guidelines in addition to recommended poses. Images were sent to the dermatologist via secure email. If necessary, the dermatologist could request an additional set of images. Quality of images was rated by the dermatologist as \u2018insufficient\u2019 vs. \u2018sufficient\u2019. Images of sufficient quality were further rated as \u2018high quality\u2019 or \u2018average quality\u2019. All telemedicine visits were real-time video encounters conducted using a video conferencing platform (Zoom Video Communication, Inc.). After reviewing the images and interviewing the patient, the dermatologist recorded their telemedicine clinical diagnosis and treatment plan into a study-specific data collection form. Patient demographics (age and sex), Fitzpatrick skin type (I-VI) and medical history were recorded. If the dermatologist was unable to reach a diagnosis, the patient was directed for an in-person visit, skin biopsy and/or additional laboratory workup.\n\n**Artificial intelligence index test**\n\nDuring the telemedicine visit, a validated, public, web-based AI algorithm (available online at http://modelderm.com) was used as the AI index test. Briefly, the algorithm was trained by one of the authors (S.S.H.) using the ASAN dataset that consisted of 220,680 images. It was validated using part of the ASAN dataset (17,125 images). The test set consisted of images from Korea (2201 images from the Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital), and the Edinburgh dataset (1300 images). The network structure includes SENet, SE-ResNet-50, and visual geometry group (VGG-19) parallelly concatenated and trained with 174 disease classes. These neural networks (SENet, SE-ResNet-50, and VGG-19) are convolutional neural networks, and the number of hidden layers is 154, 50 and 19, respectively. ImageNet pretrained models of each network were finetuned end-to-end separately. Details can be found in Han et al.\n\nThe AI web application allowed uploading images at different magnifications. As per the instructions for the algorithm, images were uploaded with the skin condition occupying at least 80% of the field of view and with the region of interest (ROI) centred. The dermatologist chose a single image to upload to the web application. Dermatologists were instructed to use the most-representative and highest-quality image for each skin condition. The algorithm output consisted of three diagnoses (from a possible list of 174 skin conditions), ranked in order of probability (ranging from 0 to 1). All algorithm outputs were recorded. This procedure was conducted in real-time, during the visit, with the potential of affecting the outcome of the visit. Given the nature and setting of the AI index test, the reference standard as well as the composition of test classes was inherently blinded. Finally, dermatologists recorded if exposure to the AI algorithm: (a) was clinically useful to them during the telemedicine visit by either increasing their diagnostic confidence and/or expanding their differential diagnosis (yes, no) and (b) led to a change in their telemedicine clinical diagnosis (yes, no).\n\n**Healthcare provider index test**\n\nSubsequent to the telemedicine visit, clinical images were independently evaluated by nine healthcare providers in a reader study. We invited three board-certified dermatologists, three dermatology residents, and three general practitioners (GPs), and all agreed to participate. The dermatologists had a median (min-max) of 3 (2\u20135) years of postresident clinical experience. One resident was in their second year and two were in their third year of dermatology training. The GPs had no formal training in dermatology but had an interest in skin diseases. GPs had a median (min-max) of 7 (1\u201311) years of postmedical school clinical experience. In this experiment, readers were only exposed to the images submitted by the patient and were blinded to clinical history/metadata as well as the results of the AI index test and the reference standard. There were no time restrictions.\n\n**Reference standard**\n\nThe reference standard was defined in one of two ways. First, if the patient was recommended to return for an in-person clinic visit, the diagnosis from this visit (and any associated laboratory testing or skin biopsies) was used as the reference standard. Second, if no in-person clinic visit was performed, a panel of 6 dermatologists evaluated the case and established the reference standard based on consensus agreement. The panel dermatologists had a median (min-max) of 8 (5\u201313) years of postresident clinical experience; there was no overlap between dermatologists who participated in the reader study and the reference standard. During the consensus evaluation, the clinical images and medical history were shown to the panel of dermatologists. All index test results were unavailable during consensus evaluations. Only one diagnosis was permitted for the reference standard. There were no cases of disagreement.\n\n**Statistical analysis**\n\nDescriptive statistics were used to characterize the study data. Diagnoses were grouped into 5 categories and 13 subcategories: (1) \u2018inflammatory\u2019 (subcategories: dermatitis, acne/rosacea, autoimmune, papulosquamous and other); (2) \u2018infectious\u2019 (subcategories: bacterial, viral, fungal and parasitic); (3) \u2018neoplastic\u2019 (subcategories: malignant and benign); (4) \u2018alopecia\u2019 (subcategories: scarring and non-scarring); and (5) \u2018other\u2019 (e.g. burn, scar, striae and among others), using a system adapted from Liu et al.\n\nThe primary outcomes were the \u2018top-1 accuracy\u2019 and the \u2018balanced top-1 accuracy\u2019. The \u2018top-1 accuracy\u2019 measures how frequently the top-1 prediction of the index test matched the reference standard diagnosis. Similar but not identical diagnoses to the reference standard (i.e. herpes simplex vs. herpes zoster) were considered incorrect. The \u2018balanced top-1 accuracy\u2019 is the top-1 accuracy computed for each unique diagnosis separately and then averaged. This was performed to account for the variant prevalence of certain diagnoses (e.g. acne vs hidradenitis suppurativa). Because algorithm fundamentally always predicted incorrect answers for the untrained cases (\u2018out of distribution\u2019), we performed sub-analyses for both measures restricted to the 174 conditions on which the index AI algorithm was explicitly trained (i.e. \u2018in distribution\u2019 vs. \u2018overall\u2019 conditions). A secondary outcome was the \u2018top-3 accuracy\u2019, which measures if the reference standard diagnosis was included in any of 3 diagnoses of the AI algorithm output. \n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.",
        "results": "### Results\n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.\n\nA total of 380 skin conditions were evaluated in the telemedicine clinic during the study period. After exclusion of 40 skin conditions (18 no clinical images, 8 insufficient quality images, 8 lost to follow-up after recommendation of in-person visit, and 6 no active disease), 340 skin conditions (87 unique diagnoses) from 281 patients met the study inclusion criteria. The mean (SD) age of the patients was 33.7 (17.5) years; 37% (n = 104) were male and 63% (n = 177) were female. Of the 340 skin conditions, 190 (55.9%) occurred in Fitzpatrick skin phototype III patients, 84 (24.7%) in type IV patients, 59 (17.3%) in type II patients and 7 (2.1%) in type I patients.\n\nIn 7.1% (n = 24) of visits, additional clinical images were requested and provided by the patients. Overall, 87.4% (n = 297) and 12.6% (n = 43) of images were rated as having \u2018high quality\u2019 and \u2018average quality\u2019, respectively. There were no differences in image quality by diagnostic category (P = 0.971) (Table S1). The reference standard was determined from in-person clinic visits (18.5%, n = 63), skin biopsy (13.8%, n = 47) and consensus agreement of panel dermatologists (81.5%, n = 277). Overall, per the reference standard, 225 of the skin conditions were inflammatory (66.2%), 52 infectious (15.3%), 32 neoplastic (9.4%), 15 alopecia (4.4%) and 16 other (4.7%). There were 87 unique diagnoses; the most frequent were acne (18.5%), contact dermatitis (8.2%), psoriasis (3.8%) and warts (3.6%) (Table 1 and Table S2).\n\n**AI performance**\n\nThe AI algorithm had overall top-1 and balanced top-1 accuracies of 41.2% and 35.1%, respectively (Table 2). Accuracy was associated with the Fitzpatrick skin type with better performance in darker skin types and was not associated with sex (Table S3). Of the 87 unique skin diagnoses included, 63 (72.4%) were part of the AI algorithm training dataset. Therefore, 305 out of the 340 (89.7%) skin conditions were \u2018in distribution\u2019 and 35 were \u2018out of distribution\u2019 images. When limiting the analysis to the \u2018in distribution\u2019 skin conditions, the top-1 and balanced top-1 accuracies increased to 45.3% and 47.6%, respectively (Table 3).\n\n**AI performance by disease category/subcategory**\n\nTop-1 accuracy was highest in the \u2018other\u2019 category (68.8%), followed by the \u2018infectious\u2019 (50%), \u2018inflammatory\u2019 (38.2%), \u2018neoplastic\u2019 (37.5%) and \u2018alopecia\u2019 (33.3%) categories (P = 0.089). For the subcategories, \u2018non-scarring alopecia\u2019 had the highest top-1 accuracy (71.4%), followed by \u2018papulosquamous\u2019 (52.9%), \u2018fungal\u2019 (52.4%), \u2018viral\u2019 (50%), \u2018bacterial\u2019 (50%), \u2018acne/rosacea\u2019 (47.4%), benign neoplasms (42.9%), \u2018dermatitis\u2019 (34.1%), \u2018autoimmune\u2019 (26.5%), \u2018parasitic\u2019 (0%), \u2018malignant neoplasms\u2019 (0%) and \u2018scarring alopecia\u2019 (0%) (P = 0.035) (Table 2).\n\n**AI vs. Physician performance**\n\nThe overall top-1 accuracy of the AI algorithm (41.2%) was lower than that of the dermatologists (60.1%; P < 0.001), residents (57.8%; P < 0.001) and GPs (49.3%; P = 0.034) (Table 2, Figure 1). The balanced top-1 accuracy of the AI algorithm was 35.1%, lower than the dermatologists (45.2%; P = 0.007) and the residents (42.6%: P = 0.045), but not different than the GPs (33.1%; P = 0.582).\n\nThe overall top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (45.3%) was lower than that of the dermatologists (63.1%) and residents (60.9%) (P < 0.001) but similar to GPs (52.9%, P = 0.061) (Table 3, Figure 1). The balanced top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (47.6%) was similar to that of the dermatologists (49.7%, P = 0.604) and residents (47.7%, P = 0.980) and superior than the GPs (39.7%, P = 0.049).\n\n**Accuracy by image quality**\n\nThe algorithm achieved a higher top-1 accuracy in images of \u2018high quality\u2019 than in images of \u2018average quality\u2019 (44.4% vs. 18.6%, P = 0.001). In contrast, there were no differences in the performance of the readers by image quality (Table S4).\n\n**AI utility**\n\nOverall, use of the AI algorithm was considered useful in 11.8% (n = 40) of cases, by increasing diagnostic confidence and/or expanding the differential diagnosis (Figure 2). The top-1 accuracy of these 40 photographs was 50%, similar to the overall top-1 accuracy of the 340 images (41.2%, P = 0.286). The top-3 accuracy was 75.9%. The \u2018balanced\u2019 top-1 accuracy for these 40 images was 54.9%, higher than the balanced top-1 accuracy of the 340 images (35.1%, P = 0.014). \u2018In distribution\u2019 pathologies included in these 40 cases were inflammatory (n = 17), infectious (n = 12) and neoplastic (n = 8) (Table S5). Exposure to the AI algorithm modified the real-time diagnosis of the teledermatologist in 0.9% (n = 3) of cases; in 2 cases the diagnosis was correctly modified (tinea corporis and perioral dermatitis) and in 1 case neither the clinician nor the algorithm matched the reference standard (sarcoidosis).",
        "column_title": "Risk of bias | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Performance of a deep neural network in teledermatology: a single-centre prospective diagnostic study.",
        "abstract": "## BACKGROUND\nThe use of artificial intelligence (AI) algorithms for the diagnosis of skin diseases has shown promise in experimental settings but has not been yet tested in real-life conditions.\n## OBJECTIVE\nTo assess the diagnostic performance and potential clinical utility of a 174-multiclass AI algorithm in a real-life telemedicine setting.\n## METHODS\nProspective, diagnostic accuracy study including consecutive patients who submitted images for teledermatology evaluation. The treating dermatologist chose a single image to upload to a web application during teleconsultation. A follow-up reader study including nine healthcare providers (3 dermatologists, 3 dermatology residents and 3 general practitioners) was performed.\n## RESULTS\nA total of 340 cases from 281 patients met study inclusion criteria. The mean (SD) age of patients was 33.7 (17.5) years; 63% (n\u00a0=\u00a0177) were female. Exposure to the AI algorithm results was considered useful in 11.8% of visits (n\u00a0=\u00a040) and the teledermatologist correctly modified the real-time diagnosis in 0.6% (n\u00a0=\u00a02) of cases. The overall top-1 accuracy of the algorithm (41.2%) was lower than that of the dermatologists (60.1%), residents (57.8%) and general practitioners (49.3%) (all comparisons P\u00a0<\u00a00.05, in the reader study). When the analysis was limited to the diagnoses on which the algorithm had been explicitly trained, the balanced top-1 accuracy of the algorithm (47.6%) was comparable to the dermatologists (49.7%) and residents (47.7%) but superior to the general practitioners (39.7%; P\u00a0=\u00a00.049). Algorithm performance was associated with patient skin type and image quality.\n## CONCLUSIONS\nA 174-disease class AI algorithm appears to be a promising tool in the triage and evaluation of lesions with patient-taken photographs via telemedicine.\n",
        "methods": "### Methods\n\nThis was an IRB-approved, single-centre, prospective, diagnostic accuracy study conducted between 27 March and 30 April 2020. All data collection was explicitly planned prior to the performance of the index tests and reference standard and the study was reported following the STARD 2015 guidelines. Consecutive patients who submitted clinical images of one or more new skin conditions for evaluation in our telemedicine dermatology clinic were considered eligible for inclusion. Both new and existing patients were eligible with no age restrictions. Cases were excluded if no images were submitted by patients, images were rated as insufficient quality, patients submitted images but did not present for telemedicine visit, patients had a telemedicine visit but did not subsequently adhere to the recommendation for an in-person visit or diagnostic testing, or upon evaluation, they had no skin condition present for examination.\n\n**Telemedicine clinical workflow**\n\nPatients submitted \u2265 1 clinical images of the skin condition to the consulting dermatologist prior to, or during the telemedicine visit. All patients received standardized imaging instructions, which provided lighting and flash guidelines in addition to recommended poses. Images were sent to the dermatologist via secure email. If necessary, the dermatologist could request an additional set of images. Quality of images was rated by the dermatologist as \u2018insufficient\u2019 vs. \u2018sufficient\u2019. Images of sufficient quality were further rated as \u2018high quality\u2019 or \u2018average quality\u2019. All telemedicine visits were real-time video encounters conducted using a video conferencing platform (Zoom Video Communication, Inc.). After reviewing the images and interviewing the patient, the dermatologist recorded their telemedicine clinical diagnosis and treatment plan into a study-specific data collection form. Patient demographics (age and sex), Fitzpatrick skin type (I-VI) and medical history were recorded. If the dermatologist was unable to reach a diagnosis, the patient was directed for an in-person visit, skin biopsy and/or additional laboratory workup.\n\n**Artificial intelligence index test**\n\nDuring the telemedicine visit, a validated, public, web-based AI algorithm (available online at http://modelderm.com) was used as the AI index test. Briefly, the algorithm was trained by one of the authors (S.S.H.) using the ASAN dataset that consisted of 220,680 images. It was validated using part of the ASAN dataset (17,125 images). The test set consisted of images from Korea (2201 images from the Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital), and the Edinburgh dataset (1300 images). The network structure includes SENet, SE-ResNet-50, and visual geometry group (VGG-19) parallelly concatenated and trained with 174 disease classes. These neural networks (SENet, SE-ResNet-50, and VGG-19) are convolutional neural networks, and the number of hidden layers is 154, 50 and 19, respectively. ImageNet pretrained models of each network were finetuned end-to-end separately. Details can be found in Han et al.\n\nThe AI web application allowed uploading images at different magnifications. As per the instructions for the algorithm, images were uploaded with the skin condition occupying at least 80% of the field of view and with the region of interest (ROI) centred. The dermatologist chose a single image to upload to the web application. Dermatologists were instructed to use the most-representative and highest-quality image for each skin condition. The algorithm output consisted of three diagnoses (from a possible list of 174 skin conditions), ranked in order of probability (ranging from 0 to 1). All algorithm outputs were recorded. This procedure was conducted in real-time, during the visit, with the potential of affecting the outcome of the visit. Given the nature and setting of the AI index test, the reference standard as well as the composition of test classes was inherently blinded. Finally, dermatologists recorded if exposure to the AI algorithm: (a) was clinically useful to them during the telemedicine visit by either increasing their diagnostic confidence and/or expanding their differential diagnosis (yes, no) and (b) led to a change in their telemedicine clinical diagnosis (yes, no).\n\n**Healthcare provider index test**\n\nSubsequent to the telemedicine visit, clinical images were independently evaluated by nine healthcare providers in a reader study. We invited three board-certified dermatologists, three dermatology residents, and three general practitioners (GPs), and all agreed to participate. The dermatologists had a median (min-max) of 3 (2\u20135) years of postresident clinical experience. One resident was in their second year and two were in their third year of dermatology training. The GPs had no formal training in dermatology but had an interest in skin diseases. GPs had a median (min-max) of 7 (1\u201311) years of postmedical school clinical experience. In this experiment, readers were only exposed to the images submitted by the patient and were blinded to clinical history/metadata as well as the results of the AI index test and the reference standard. There were no time restrictions.\n\n**Reference standard**\n\nThe reference standard was defined in one of two ways. First, if the patient was recommended to return for an in-person clinic visit, the diagnosis from this visit (and any associated laboratory testing or skin biopsies) was used as the reference standard. Second, if no in-person clinic visit was performed, a panel of 6 dermatologists evaluated the case and established the reference standard based on consensus agreement. The panel dermatologists had a median (min-max) of 8 (5\u201313) years of postresident clinical experience; there was no overlap between dermatologists who participated in the reader study and the reference standard. During the consensus evaluation, the clinical images and medical history were shown to the panel of dermatologists. All index test results were unavailable during consensus evaluations. Only one diagnosis was permitted for the reference standard. There were no cases of disagreement.\n\n**Statistical analysis**\n\nDescriptive statistics were used to characterize the study data. Diagnoses were grouped into 5 categories and 13 subcategories: (1) \u2018inflammatory\u2019 (subcategories: dermatitis, acne/rosacea, autoimmune, papulosquamous and other); (2) \u2018infectious\u2019 (subcategories: bacterial, viral, fungal and parasitic); (3) \u2018neoplastic\u2019 (subcategories: malignant and benign); (4) \u2018alopecia\u2019 (subcategories: scarring and non-scarring); and (5) \u2018other\u2019 (e.g. burn, scar, striae and among others), using a system adapted from Liu et al.\n\nThe primary outcomes were the \u2018top-1 accuracy\u2019 and the \u2018balanced top-1 accuracy\u2019. The \u2018top-1 accuracy\u2019 measures how frequently the top-1 prediction of the index test matched the reference standard diagnosis. Similar but not identical diagnoses to the reference standard (i.e. herpes simplex vs. herpes zoster) were considered incorrect. The \u2018balanced top-1 accuracy\u2019 is the top-1 accuracy computed for each unique diagnosis separately and then averaged. This was performed to account for the variant prevalence of certain diagnoses (e.g. acne vs hidradenitis suppurativa). Because algorithm fundamentally always predicted incorrect answers for the untrained cases (\u2018out of distribution\u2019), we performed sub-analyses for both measures restricted to the 174 conditions on which the index AI algorithm was explicitly trained (i.e. \u2018in distribution\u2019 vs. \u2018overall\u2019 conditions). A secondary outcome was the \u2018top-3 accuracy\u2019, which measures if the reference standard diagnosis was included in any of 3 diagnoses of the AI algorithm output. \n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.",
        "results": "### Results\n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.\n\nA total of 380 skin conditions were evaluated in the telemedicine clinic during the study period. After exclusion of 40 skin conditions (18 no clinical images, 8 insufficient quality images, 8 lost to follow-up after recommendation of in-person visit, and 6 no active disease), 340 skin conditions (87 unique diagnoses) from 281 patients met the study inclusion criteria. The mean (SD) age of the patients was 33.7 (17.5) years; 37% (n = 104) were male and 63% (n = 177) were female. Of the 340 skin conditions, 190 (55.9%) occurred in Fitzpatrick skin phototype III patients, 84 (24.7%) in type IV patients, 59 (17.3%) in type II patients and 7 (2.1%) in type I patients.\n\nIn 7.1% (n = 24) of visits, additional clinical images were requested and provided by the patients. Overall, 87.4% (n = 297) and 12.6% (n = 43) of images were rated as having \u2018high quality\u2019 and \u2018average quality\u2019, respectively. There were no differences in image quality by diagnostic category (P = 0.971) (Table S1). The reference standard was determined from in-person clinic visits (18.5%, n = 63), skin biopsy (13.8%, n = 47) and consensus agreement of panel dermatologists (81.5%, n = 277). Overall, per the reference standard, 225 of the skin conditions were inflammatory (66.2%), 52 infectious (15.3%), 32 neoplastic (9.4%), 15 alopecia (4.4%) and 16 other (4.7%). There were 87 unique diagnoses; the most frequent were acne (18.5%), contact dermatitis (8.2%), psoriasis (3.8%) and warts (3.6%) (Table 1 and Table S2).\n\n**AI performance**\n\nThe AI algorithm had overall top-1 and balanced top-1 accuracies of 41.2% and 35.1%, respectively (Table 2). Accuracy was associated with the Fitzpatrick skin type with better performance in darker skin types and was not associated with sex (Table S3). Of the 87 unique skin diagnoses included, 63 (72.4%) were part of the AI algorithm training dataset. Therefore, 305 out of the 340 (89.7%) skin conditions were \u2018in distribution\u2019 and 35 were \u2018out of distribution\u2019 images. When limiting the analysis to the \u2018in distribution\u2019 skin conditions, the top-1 and balanced top-1 accuracies increased to 45.3% and 47.6%, respectively (Table 3).\n\n**AI performance by disease category/subcategory**\n\nTop-1 accuracy was highest in the \u2018other\u2019 category (68.8%), followed by the \u2018infectious\u2019 (50%), \u2018inflammatory\u2019 (38.2%), \u2018neoplastic\u2019 (37.5%) and \u2018alopecia\u2019 (33.3%) categories (P = 0.089). For the subcategories, \u2018non-scarring alopecia\u2019 had the highest top-1 accuracy (71.4%), followed by \u2018papulosquamous\u2019 (52.9%), \u2018fungal\u2019 (52.4%), \u2018viral\u2019 (50%), \u2018bacterial\u2019 (50%), \u2018acne/rosacea\u2019 (47.4%), benign neoplasms (42.9%), \u2018dermatitis\u2019 (34.1%), \u2018autoimmune\u2019 (26.5%), \u2018parasitic\u2019 (0%), \u2018malignant neoplasms\u2019 (0%) and \u2018scarring alopecia\u2019 (0%) (P = 0.035) (Table 2).\n\n**AI vs. Physician performance**\n\nThe overall top-1 accuracy of the AI algorithm (41.2%) was lower than that of the dermatologists (60.1%; P < 0.001), residents (57.8%; P < 0.001) and GPs (49.3%; P = 0.034) (Table 2, Figure 1). The balanced top-1 accuracy of the AI algorithm was 35.1%, lower than the dermatologists (45.2%; P = 0.007) and the residents (42.6%: P = 0.045), but not different than the GPs (33.1%; P = 0.582).\n\nThe overall top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (45.3%) was lower than that of the dermatologists (63.1%) and residents (60.9%) (P < 0.001) but similar to GPs (52.9%, P = 0.061) (Table 3, Figure 1). The balanced top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (47.6%) was similar to that of the dermatologists (49.7%, P = 0.604) and residents (47.7%, P = 0.980) and superior than the GPs (39.7%, P = 0.049).\n\n**Accuracy by image quality**\n\nThe algorithm achieved a higher top-1 accuracy in images of \u2018high quality\u2019 than in images of \u2018average quality\u2019 (44.4% vs. 18.6%, P = 0.001). In contrast, there were no differences in the performance of the readers by image quality (Table S4).\n\n**AI utility**\n\nOverall, use of the AI algorithm was considered useful in 11.8% (n = 40) of cases, by increasing diagnostic confidence and/or expanding the differential diagnosis (Figure 2). The top-1 accuracy of these 40 photographs was 50%, similar to the overall top-1 accuracy of the 340 images (41.2%, P = 0.286). The top-3 accuracy was 75.9%. The \u2018balanced\u2019 top-1 accuracy for these 40 images was 54.9%, higher than the balanced top-1 accuracy of the 340 images (35.1%, P = 0.014). \u2018In distribution\u2019 pathologies included in these 40 cases were inflammatory (n = 17), infectious (n = 12) and neoplastic (n = 8) (Table S5). Exposure to the AI algorithm modified the real-time diagnosis of the teledermatologist in 0.9% (n = 3) of cases; in 2 cases the diagnosis was correctly modified (tinea corporis and perioral dermatitis) and in 1 case neither the clinician nor the algorithm matched the reference standard (sarcoidosis).",
        "column_title": "Risk of bias | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Performance of a deep neural network in teledermatology: a single-centre prospective diagnostic study.",
        "abstract": "## BACKGROUND\nThe use of artificial intelligence (AI) algorithms for the diagnosis of skin diseases has shown promise in experimental settings but has not been yet tested in real-life conditions.\n## OBJECTIVE\nTo assess the diagnostic performance and potential clinical utility of a 174-multiclass AI algorithm in a real-life telemedicine setting.\n## METHODS\nProspective, diagnostic accuracy study including consecutive patients who submitted images for teledermatology evaluation. The treating dermatologist chose a single image to upload to a web application during teleconsultation. A follow-up reader study including nine healthcare providers (3 dermatologists, 3 dermatology residents and 3 general practitioners) was performed.\n## RESULTS\nA total of 340 cases from 281 patients met study inclusion criteria. The mean (SD) age of patients was 33.7 (17.5) years; 63% (n\u00a0=\u00a0177) were female. Exposure to the AI algorithm results was considered useful in 11.8% of visits (n\u00a0=\u00a040) and the teledermatologist correctly modified the real-time diagnosis in 0.6% (n\u00a0=\u00a02) of cases. The overall top-1 accuracy of the algorithm (41.2%) was lower than that of the dermatologists (60.1%), residents (57.8%) and general practitioners (49.3%) (all comparisons P\u00a0<\u00a00.05, in the reader study). When the analysis was limited to the diagnoses on which the algorithm had been explicitly trained, the balanced top-1 accuracy of the algorithm (47.6%) was comparable to the dermatologists (49.7%) and residents (47.7%) but superior to the general practitioners (39.7%; P\u00a0=\u00a00.049). Algorithm performance was associated with patient skin type and image quality.\n## CONCLUSIONS\nA 174-disease class AI algorithm appears to be a promising tool in the triage and evaluation of lesions with patient-taken photographs via telemedicine.\n",
        "methods": "### Methods\n\nThis was an IRB-approved, single-centre, prospective, diagnostic accuracy study conducted between 27 March and 30 April 2020. All data collection was explicitly planned prior to the performance of the index tests and reference standard and the study was reported following the STARD 2015 guidelines. Consecutive patients who submitted clinical images of one or more new skin conditions for evaluation in our telemedicine dermatology clinic were considered eligible for inclusion. Both new and existing patients were eligible with no age restrictions. Cases were excluded if no images were submitted by patients, images were rated as insufficient quality, patients submitted images but did not present for telemedicine visit, patients had a telemedicine visit but did not subsequently adhere to the recommendation for an in-person visit or diagnostic testing, or upon evaluation, they had no skin condition present for examination.\n\n**Telemedicine clinical workflow**\n\nPatients submitted \u2265 1 clinical images of the skin condition to the consulting dermatologist prior to, or during the telemedicine visit. All patients received standardized imaging instructions, which provided lighting and flash guidelines in addition to recommended poses. Images were sent to the dermatologist via secure email. If necessary, the dermatologist could request an additional set of images. Quality of images was rated by the dermatologist as \u2018insufficient\u2019 vs. \u2018sufficient\u2019. Images of sufficient quality were further rated as \u2018high quality\u2019 or \u2018average quality\u2019. All telemedicine visits were real-time video encounters conducted using a video conferencing platform (Zoom Video Communication, Inc.). After reviewing the images and interviewing the patient, the dermatologist recorded their telemedicine clinical diagnosis and treatment plan into a study-specific data collection form. Patient demographics (age and sex), Fitzpatrick skin type (I-VI) and medical history were recorded. If the dermatologist was unable to reach a diagnosis, the patient was directed for an in-person visit, skin biopsy and/or additional laboratory workup.\n\n**Artificial intelligence index test**\n\nDuring the telemedicine visit, a validated, public, web-based AI algorithm (available online at http://modelderm.com) was used as the AI index test. Briefly, the algorithm was trained by one of the authors (S.S.H.) using the ASAN dataset that consisted of 220,680 images. It was validated using part of the ASAN dataset (17,125 images). The test set consisted of images from Korea (2201 images from the Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital), and the Edinburgh dataset (1300 images). The network structure includes SENet, SE-ResNet-50, and visual geometry group (VGG-19) parallelly concatenated and trained with 174 disease classes. These neural networks (SENet, SE-ResNet-50, and VGG-19) are convolutional neural networks, and the number of hidden layers is 154, 50 and 19, respectively. ImageNet pretrained models of each network were finetuned end-to-end separately. Details can be found in Han et al.\n\nThe AI web application allowed uploading images at different magnifications. As per the instructions for the algorithm, images were uploaded with the skin condition occupying at least 80% of the field of view and with the region of interest (ROI) centred. The dermatologist chose a single image to upload to the web application. Dermatologists were instructed to use the most-representative and highest-quality image for each skin condition. The algorithm output consisted of three diagnoses (from a possible list of 174 skin conditions), ranked in order of probability (ranging from 0 to 1). All algorithm outputs were recorded. This procedure was conducted in real-time, during the visit, with the potential of affecting the outcome of the visit. Given the nature and setting of the AI index test, the reference standard as well as the composition of test classes was inherently blinded. Finally, dermatologists recorded if exposure to the AI algorithm: (a) was clinically useful to them during the telemedicine visit by either increasing their diagnostic confidence and/or expanding their differential diagnosis (yes, no) and (b) led to a change in their telemedicine clinical diagnosis (yes, no).\n\n**Healthcare provider index test**\n\nSubsequent to the telemedicine visit, clinical images were independently evaluated by nine healthcare providers in a reader study. We invited three board-certified dermatologists, three dermatology residents, and three general practitioners (GPs), and all agreed to participate. The dermatologists had a median (min-max) of 3 (2\u20135) years of postresident clinical experience. One resident was in their second year and two were in their third year of dermatology training. The GPs had no formal training in dermatology but had an interest in skin diseases. GPs had a median (min-max) of 7 (1\u201311) years of postmedical school clinical experience. In this experiment, readers were only exposed to the images submitted by the patient and were blinded to clinical history/metadata as well as the results of the AI index test and the reference standard. There were no time restrictions.\n\n**Reference standard**\n\nThe reference standard was defined in one of two ways. First, if the patient was recommended to return for an in-person clinic visit, the diagnosis from this visit (and any associated laboratory testing or skin biopsies) was used as the reference standard. Second, if no in-person clinic visit was performed, a panel of 6 dermatologists evaluated the case and established the reference standard based on consensus agreement. The panel dermatologists had a median (min-max) of 8 (5\u201313) years of postresident clinical experience; there was no overlap between dermatologists who participated in the reader study and the reference standard. During the consensus evaluation, the clinical images and medical history were shown to the panel of dermatologists. All index test results were unavailable during consensus evaluations. Only one diagnosis was permitted for the reference standard. There were no cases of disagreement.\n\n**Statistical analysis**\n\nDescriptive statistics were used to characterize the study data. Diagnoses were grouped into 5 categories and 13 subcategories: (1) \u2018inflammatory\u2019 (subcategories: dermatitis, acne/rosacea, autoimmune, papulosquamous and other); (2) \u2018infectious\u2019 (subcategories: bacterial, viral, fungal and parasitic); (3) \u2018neoplastic\u2019 (subcategories: malignant and benign); (4) \u2018alopecia\u2019 (subcategories: scarring and non-scarring); and (5) \u2018other\u2019 (e.g. burn, scar, striae and among others), using a system adapted from Liu et al.\n\nThe primary outcomes were the \u2018top-1 accuracy\u2019 and the \u2018balanced top-1 accuracy\u2019. The \u2018top-1 accuracy\u2019 measures how frequently the top-1 prediction of the index test matched the reference standard diagnosis. Similar but not identical diagnoses to the reference standard (i.e. herpes simplex vs. herpes zoster) were considered incorrect. The \u2018balanced top-1 accuracy\u2019 is the top-1 accuracy computed for each unique diagnosis separately and then averaged. This was performed to account for the variant prevalence of certain diagnoses (e.g. acne vs hidradenitis suppurativa). Because algorithm fundamentally always predicted incorrect answers for the untrained cases (\u2018out of distribution\u2019), we performed sub-analyses for both measures restricted to the 174 conditions on which the index AI algorithm was explicitly trained (i.e. \u2018in distribution\u2019 vs. \u2018overall\u2019 conditions). A secondary outcome was the \u2018top-3 accuracy\u2019, which measures if the reference standard diagnosis was included in any of 3 diagnoses of the AI algorithm output. \n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.",
        "results": "### Results\n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.\n\nA total of 380 skin conditions were evaluated in the telemedicine clinic during the study period. After exclusion of 40 skin conditions (18 no clinical images, 8 insufficient quality images, 8 lost to follow-up after recommendation of in-person visit, and 6 no active disease), 340 skin conditions (87 unique diagnoses) from 281 patients met the study inclusion criteria. The mean (SD) age of the patients was 33.7 (17.5) years; 37% (n = 104) were male and 63% (n = 177) were female. Of the 340 skin conditions, 190 (55.9%) occurred in Fitzpatrick skin phototype III patients, 84 (24.7%) in type IV patients, 59 (17.3%) in type II patients and 7 (2.1%) in type I patients.\n\nIn 7.1% (n = 24) of visits, additional clinical images were requested and provided by the patients. Overall, 87.4% (n = 297) and 12.6% (n = 43) of images were rated as having \u2018high quality\u2019 and \u2018average quality\u2019, respectively. There were no differences in image quality by diagnostic category (P = 0.971) (Table S1). The reference standard was determined from in-person clinic visits (18.5%, n = 63), skin biopsy (13.8%, n = 47) and consensus agreement of panel dermatologists (81.5%, n = 277). Overall, per the reference standard, 225 of the skin conditions were inflammatory (66.2%), 52 infectious (15.3%), 32 neoplastic (9.4%), 15 alopecia (4.4%) and 16 other (4.7%). There were 87 unique diagnoses; the most frequent were acne (18.5%), contact dermatitis (8.2%), psoriasis (3.8%) and warts (3.6%) (Table 1 and Table S2).\n\n**AI performance**\n\nThe AI algorithm had overall top-1 and balanced top-1 accuracies of 41.2% and 35.1%, respectively (Table 2). Accuracy was associated with the Fitzpatrick skin type with better performance in darker skin types and was not associated with sex (Table S3). Of the 87 unique skin diagnoses included, 63 (72.4%) were part of the AI algorithm training dataset. Therefore, 305 out of the 340 (89.7%) skin conditions were \u2018in distribution\u2019 and 35 were \u2018out of distribution\u2019 images. When limiting the analysis to the \u2018in distribution\u2019 skin conditions, the top-1 and balanced top-1 accuracies increased to 45.3% and 47.6%, respectively (Table 3).\n\n**AI performance by disease category/subcategory**\n\nTop-1 accuracy was highest in the \u2018other\u2019 category (68.8%), followed by the \u2018infectious\u2019 (50%), \u2018inflammatory\u2019 (38.2%), \u2018neoplastic\u2019 (37.5%) and \u2018alopecia\u2019 (33.3%) categories (P = 0.089). For the subcategories, \u2018non-scarring alopecia\u2019 had the highest top-1 accuracy (71.4%), followed by \u2018papulosquamous\u2019 (52.9%), \u2018fungal\u2019 (52.4%), \u2018viral\u2019 (50%), \u2018bacterial\u2019 (50%), \u2018acne/rosacea\u2019 (47.4%), benign neoplasms (42.9%), \u2018dermatitis\u2019 (34.1%), \u2018autoimmune\u2019 (26.5%), \u2018parasitic\u2019 (0%), \u2018malignant neoplasms\u2019 (0%) and \u2018scarring alopecia\u2019 (0%) (P = 0.035) (Table 2).\n\n**AI vs. Physician performance**\n\nThe overall top-1 accuracy of the AI algorithm (41.2%) was lower than that of the dermatologists (60.1%; P < 0.001), residents (57.8%; P < 0.001) and GPs (49.3%; P = 0.034) (Table 2, Figure 1). The balanced top-1 accuracy of the AI algorithm was 35.1%, lower than the dermatologists (45.2%; P = 0.007) and the residents (42.6%: P = 0.045), but not different than the GPs (33.1%; P = 0.582).\n\nThe overall top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (45.3%) was lower than that of the dermatologists (63.1%) and residents (60.9%) (P < 0.001) but similar to GPs (52.9%, P = 0.061) (Table 3, Figure 1). The balanced top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (47.6%) was similar to that of the dermatologists (49.7%, P = 0.604) and residents (47.7%, P = 0.980) and superior than the GPs (39.7%, P = 0.049).\n\n**Accuracy by image quality**\n\nThe algorithm achieved a higher top-1 accuracy in images of \u2018high quality\u2019 than in images of \u2018average quality\u2019 (44.4% vs. 18.6%, P = 0.001). In contrast, there were no differences in the performance of the readers by image quality (Table S4).\n\n**AI utility**\n\nOverall, use of the AI algorithm was considered useful in 11.8% (n = 40) of cases, by increasing diagnostic confidence and/or expanding the differential diagnosis (Figure 2). The top-1 accuracy of these 40 photographs was 50%, similar to the overall top-1 accuracy of the 340 images (41.2%, P = 0.286). The top-3 accuracy was 75.9%. The \u2018balanced\u2019 top-1 accuracy for these 40 images was 54.9%, higher than the balanced top-1 accuracy of the 340 images (35.1%, P = 0.014). \u2018In distribution\u2019 pathologies included in these 40 cases were inflammatory (n = 17), infectious (n = 12) and neoplastic (n = 8) (Table S5). Exposure to the AI algorithm modified the real-time diagnosis of the teledermatologist in 0.9% (n = 3) of cases; in 2 cases the diagnosis was correctly modified (tinea corporis and perioral dermatitis) and in 1 case neither the clinician nor the algorithm matched the reference standard (sarcoidosis).",
        "column_title": "Risk of bias | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Performance of a deep neural network in teledermatology: a single-centre prospective diagnostic study.",
        "abstract": "## BACKGROUND\nThe use of artificial intelligence (AI) algorithms for the diagnosis of skin diseases has shown promise in experimental settings but has not been yet tested in real-life conditions.\n## OBJECTIVE\nTo assess the diagnostic performance and potential clinical utility of a 174-multiclass AI algorithm in a real-life telemedicine setting.\n## METHODS\nProspective, diagnostic accuracy study including consecutive patients who submitted images for teledermatology evaluation. The treating dermatologist chose a single image to upload to a web application during teleconsultation. A follow-up reader study including nine healthcare providers (3 dermatologists, 3 dermatology residents and 3 general practitioners) was performed.\n## RESULTS\nA total of 340 cases from 281 patients met study inclusion criteria. The mean (SD) age of patients was 33.7 (17.5) years; 63% (n\u00a0=\u00a0177) were female. Exposure to the AI algorithm results was considered useful in 11.8% of visits (n\u00a0=\u00a040) and the teledermatologist correctly modified the real-time diagnosis in 0.6% (n\u00a0=\u00a02) of cases. The overall top-1 accuracy of the algorithm (41.2%) was lower than that of the dermatologists (60.1%), residents (57.8%) and general practitioners (49.3%) (all comparisons P\u00a0<\u00a00.05, in the reader study). When the analysis was limited to the diagnoses on which the algorithm had been explicitly trained, the balanced top-1 accuracy of the algorithm (47.6%) was comparable to the dermatologists (49.7%) and residents (47.7%) but superior to the general practitioners (39.7%; P\u00a0=\u00a00.049). Algorithm performance was associated with patient skin type and image quality.\n## CONCLUSIONS\nA 174-disease class AI algorithm appears to be a promising tool in the triage and evaluation of lesions with patient-taken photographs via telemedicine.\n",
        "methods": "### Methods\n\nThis was an IRB-approved, single-centre, prospective, diagnostic accuracy study conducted between 27 March and 30 April 2020. All data collection was explicitly planned prior to the performance of the index tests and reference standard and the study was reported following the STARD 2015 guidelines. Consecutive patients who submitted clinical images of one or more new skin conditions for evaluation in our telemedicine dermatology clinic were considered eligible for inclusion. Both new and existing patients were eligible with no age restrictions. Cases were excluded if no images were submitted by patients, images were rated as insufficient quality, patients submitted images but did not present for telemedicine visit, patients had a telemedicine visit but did not subsequently adhere to the recommendation for an in-person visit or diagnostic testing, or upon evaluation, they had no skin condition present for examination.\n\n**Telemedicine clinical workflow**\n\nPatients submitted \u2265 1 clinical images of the skin condition to the consulting dermatologist prior to, or during the telemedicine visit. All patients received standardized imaging instructions, which provided lighting and flash guidelines in addition to recommended poses. Images were sent to the dermatologist via secure email. If necessary, the dermatologist could request an additional set of images. Quality of images was rated by the dermatologist as \u2018insufficient\u2019 vs. \u2018sufficient\u2019. Images of sufficient quality were further rated as \u2018high quality\u2019 or \u2018average quality\u2019. All telemedicine visits were real-time video encounters conducted using a video conferencing platform (Zoom Video Communication, Inc.). After reviewing the images and interviewing the patient, the dermatologist recorded their telemedicine clinical diagnosis and treatment plan into a study-specific data collection form. Patient demographics (age and sex), Fitzpatrick skin type (I-VI) and medical history were recorded. If the dermatologist was unable to reach a diagnosis, the patient was directed for an in-person visit, skin biopsy and/or additional laboratory workup.\n\n**Artificial intelligence index test**\n\nDuring the telemedicine visit, a validated, public, web-based AI algorithm (available online at http://modelderm.com) was used as the AI index test. Briefly, the algorithm was trained by one of the authors (S.S.H.) using the ASAN dataset that consisted of 220,680 images. It was validated using part of the ASAN dataset (17,125 images). The test set consisted of images from Korea (2201 images from the Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital), and the Edinburgh dataset (1300 images). The network structure includes SENet, SE-ResNet-50, and visual geometry group (VGG-19) parallelly concatenated and trained with 174 disease classes. These neural networks (SENet, SE-ResNet-50, and VGG-19) are convolutional neural networks, and the number of hidden layers is 154, 50 and 19, respectively. ImageNet pretrained models of each network were finetuned end-to-end separately. Details can be found in Han et al.\n\nThe AI web application allowed uploading images at different magnifications. As per the instructions for the algorithm, images were uploaded with the skin condition occupying at least 80% of the field of view and with the region of interest (ROI) centred. The dermatologist chose a single image to upload to the web application. Dermatologists were instructed to use the most-representative and highest-quality image for each skin condition. The algorithm output consisted of three diagnoses (from a possible list of 174 skin conditions), ranked in order of probability (ranging from 0 to 1). All algorithm outputs were recorded. This procedure was conducted in real-time, during the visit, with the potential of affecting the outcome of the visit. Given the nature and setting of the AI index test, the reference standard as well as the composition of test classes was inherently blinded. Finally, dermatologists recorded if exposure to the AI algorithm: (a) was clinically useful to them during the telemedicine visit by either increasing their diagnostic confidence and/or expanding their differential diagnosis (yes, no) and (b) led to a change in their telemedicine clinical diagnosis (yes, no).\n\n**Healthcare provider index test**\n\nSubsequent to the telemedicine visit, clinical images were independently evaluated by nine healthcare providers in a reader study. We invited three board-certified dermatologists, three dermatology residents, and three general practitioners (GPs), and all agreed to participate. The dermatologists had a median (min-max) of 3 (2\u20135) years of postresident clinical experience. One resident was in their second year and two were in their third year of dermatology training. The GPs had no formal training in dermatology but had an interest in skin diseases. GPs had a median (min-max) of 7 (1\u201311) years of postmedical school clinical experience. In this experiment, readers were only exposed to the images submitted by the patient and were blinded to clinical history/metadata as well as the results of the AI index test and the reference standard. There were no time restrictions.\n\n**Reference standard**\n\nThe reference standard was defined in one of two ways. First, if the patient was recommended to return for an in-person clinic visit, the diagnosis from this visit (and any associated laboratory testing or skin biopsies) was used as the reference standard. Second, if no in-person clinic visit was performed, a panel of 6 dermatologists evaluated the case and established the reference standard based on consensus agreement. The panel dermatologists had a median (min-max) of 8 (5\u201313) years of postresident clinical experience; there was no overlap between dermatologists who participated in the reader study and the reference standard. During the consensus evaluation, the clinical images and medical history were shown to the panel of dermatologists. All index test results were unavailable during consensus evaluations. Only one diagnosis was permitted for the reference standard. There were no cases of disagreement.\n\n**Statistical analysis**\n\nDescriptive statistics were used to characterize the study data. Diagnoses were grouped into 5 categories and 13 subcategories: (1) \u2018inflammatory\u2019 (subcategories: dermatitis, acne/rosacea, autoimmune, papulosquamous and other); (2) \u2018infectious\u2019 (subcategories: bacterial, viral, fungal and parasitic); (3) \u2018neoplastic\u2019 (subcategories: malignant and benign); (4) \u2018alopecia\u2019 (subcategories: scarring and non-scarring); and (5) \u2018other\u2019 (e.g. burn, scar, striae and among others), using a system adapted from Liu et al.\n\nThe primary outcomes were the \u2018top-1 accuracy\u2019 and the \u2018balanced top-1 accuracy\u2019. The \u2018top-1 accuracy\u2019 measures how frequently the top-1 prediction of the index test matched the reference standard diagnosis. Similar but not identical diagnoses to the reference standard (i.e. herpes simplex vs. herpes zoster) were considered incorrect. The \u2018balanced top-1 accuracy\u2019 is the top-1 accuracy computed for each unique diagnosis separately and then averaged. This was performed to account for the variant prevalence of certain diagnoses (e.g. acne vs hidradenitis suppurativa). Because algorithm fundamentally always predicted incorrect answers for the untrained cases (\u2018out of distribution\u2019), we performed sub-analyses for both measures restricted to the 174 conditions on which the index AI algorithm was explicitly trained (i.e. \u2018in distribution\u2019 vs. \u2018overall\u2019 conditions). A secondary outcome was the \u2018top-3 accuracy\u2019, which measures if the reference standard diagnosis was included in any of 3 diagnoses of the AI algorithm output. \n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.",
        "results": "### Results\n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.\n\nA total of 380 skin conditions were evaluated in the telemedicine clinic during the study period. After exclusion of 40 skin conditions (18 no clinical images, 8 insufficient quality images, 8 lost to follow-up after recommendation of in-person visit, and 6 no active disease), 340 skin conditions (87 unique diagnoses) from 281 patients met the study inclusion criteria. The mean (SD) age of the patients was 33.7 (17.5) years; 37% (n = 104) were male and 63% (n = 177) were female. Of the 340 skin conditions, 190 (55.9%) occurred in Fitzpatrick skin phototype III patients, 84 (24.7%) in type IV patients, 59 (17.3%) in type II patients and 7 (2.1%) in type I patients.\n\nIn 7.1% (n = 24) of visits, additional clinical images were requested and provided by the patients. Overall, 87.4% (n = 297) and 12.6% (n = 43) of images were rated as having \u2018high quality\u2019 and \u2018average quality\u2019, respectively. There were no differences in image quality by diagnostic category (P = 0.971) (Table S1). The reference standard was determined from in-person clinic visits (18.5%, n = 63), skin biopsy (13.8%, n = 47) and consensus agreement of panel dermatologists (81.5%, n = 277). Overall, per the reference standard, 225 of the skin conditions were inflammatory (66.2%), 52 infectious (15.3%), 32 neoplastic (9.4%), 15 alopecia (4.4%) and 16 other (4.7%). There were 87 unique diagnoses; the most frequent were acne (18.5%), contact dermatitis (8.2%), psoriasis (3.8%) and warts (3.6%) (Table 1 and Table S2).\n\n**AI performance**\n\nThe AI algorithm had overall top-1 and balanced top-1 accuracies of 41.2% and 35.1%, respectively (Table 2). Accuracy was associated with the Fitzpatrick skin type with better performance in darker skin types and was not associated with sex (Table S3). Of the 87 unique skin diagnoses included, 63 (72.4%) were part of the AI algorithm training dataset. Therefore, 305 out of the 340 (89.7%) skin conditions were \u2018in distribution\u2019 and 35 were \u2018out of distribution\u2019 images. When limiting the analysis to the \u2018in distribution\u2019 skin conditions, the top-1 and balanced top-1 accuracies increased to 45.3% and 47.6%, respectively (Table 3).\n\n**AI performance by disease category/subcategory**\n\nTop-1 accuracy was highest in the \u2018other\u2019 category (68.8%), followed by the \u2018infectious\u2019 (50%), \u2018inflammatory\u2019 (38.2%), \u2018neoplastic\u2019 (37.5%) and \u2018alopecia\u2019 (33.3%) categories (P = 0.089). For the subcategories, \u2018non-scarring alopecia\u2019 had the highest top-1 accuracy (71.4%), followed by \u2018papulosquamous\u2019 (52.9%), \u2018fungal\u2019 (52.4%), \u2018viral\u2019 (50%), \u2018bacterial\u2019 (50%), \u2018acne/rosacea\u2019 (47.4%), benign neoplasms (42.9%), \u2018dermatitis\u2019 (34.1%), \u2018autoimmune\u2019 (26.5%), \u2018parasitic\u2019 (0%), \u2018malignant neoplasms\u2019 (0%) and \u2018scarring alopecia\u2019 (0%) (P = 0.035) (Table 2).\n\n**AI vs. Physician performance**\n\nThe overall top-1 accuracy of the AI algorithm (41.2%) was lower than that of the dermatologists (60.1%; P < 0.001), residents (57.8%; P < 0.001) and GPs (49.3%; P = 0.034) (Table 2, Figure 1). The balanced top-1 accuracy of the AI algorithm was 35.1%, lower than the dermatologists (45.2%; P = 0.007) and the residents (42.6%: P = 0.045), but not different than the GPs (33.1%; P = 0.582).\n\nThe overall top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (45.3%) was lower than that of the dermatologists (63.1%) and residents (60.9%) (P < 0.001) but similar to GPs (52.9%, P = 0.061) (Table 3, Figure 1). The balanced top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (47.6%) was similar to that of the dermatologists (49.7%, P = 0.604) and residents (47.7%, P = 0.980) and superior than the GPs (39.7%, P = 0.049).\n\n**Accuracy by image quality**\n\nThe algorithm achieved a higher top-1 accuracy in images of \u2018high quality\u2019 than in images of \u2018average quality\u2019 (44.4% vs. 18.6%, P = 0.001). In contrast, there were no differences in the performance of the readers by image quality (Table S4).\n\n**AI utility**\n\nOverall, use of the AI algorithm was considered useful in 11.8% (n = 40) of cases, by increasing diagnostic confidence and/or expanding the differential diagnosis (Figure 2). The top-1 accuracy of these 40 photographs was 50%, similar to the overall top-1 accuracy of the 340 images (41.2%, P = 0.286). The top-3 accuracy was 75.9%. The \u2018balanced\u2019 top-1 accuracy for these 40 images was 54.9%, higher than the balanced top-1 accuracy of the 340 images (35.1%, P = 0.014). \u2018In distribution\u2019 pathologies included in these 40 cases were inflammatory (n = 17), infectious (n = 12) and neoplastic (n = 8) (Table S5). Exposure to the AI algorithm modified the real-time diagnosis of the teledermatologist in 0.9% (n = 3) of cases; in 2 cases the diagnosis was correctly modified (tinea corporis and perioral dermatitis) and in 1 case neither the clinician nor the algorithm matched the reference standard (sarcoidosis).",
        "column_title": "Risk of bias | Flow and timing",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Performance of a deep neural network in teledermatology: a single-centre prospective diagnostic study.",
        "abstract": "## BACKGROUND\nThe use of artificial intelligence (AI) algorithms for the diagnosis of skin diseases has shown promise in experimental settings but has not been yet tested in real-life conditions.\n## OBJECTIVE\nTo assess the diagnostic performance and potential clinical utility of a 174-multiclass AI algorithm in a real-life telemedicine setting.\n## METHODS\nProspective, diagnostic accuracy study including consecutive patients who submitted images for teledermatology evaluation. The treating dermatologist chose a single image to upload to a web application during teleconsultation. A follow-up reader study including nine healthcare providers (3 dermatologists, 3 dermatology residents and 3 general practitioners) was performed.\n## RESULTS\nA total of 340 cases from 281 patients met study inclusion criteria. The mean (SD) age of patients was 33.7 (17.5) years; 63% (n\u00a0=\u00a0177) were female. Exposure to the AI algorithm results was considered useful in 11.8% of visits (n\u00a0=\u00a040) and the teledermatologist correctly modified the real-time diagnosis in 0.6% (n\u00a0=\u00a02) of cases. The overall top-1 accuracy of the algorithm (41.2%) was lower than that of the dermatologists (60.1%), residents (57.8%) and general practitioners (49.3%) (all comparisons P\u00a0<\u00a00.05, in the reader study). When the analysis was limited to the diagnoses on which the algorithm had been explicitly trained, the balanced top-1 accuracy of the algorithm (47.6%) was comparable to the dermatologists (49.7%) and residents (47.7%) but superior to the general practitioners (39.7%; P\u00a0=\u00a00.049). Algorithm performance was associated with patient skin type and image quality.\n## CONCLUSIONS\nA 174-disease class AI algorithm appears to be a promising tool in the triage and evaluation of lesions with patient-taken photographs via telemedicine.\n",
        "methods": "### Methods\n\nThis was an IRB-approved, single-centre, prospective, diagnostic accuracy study conducted between 27 March and 30 April 2020. All data collection was explicitly planned prior to the performance of the index tests and reference standard and the study was reported following the STARD 2015 guidelines. Consecutive patients who submitted clinical images of one or more new skin conditions for evaluation in our telemedicine dermatology clinic were considered eligible for inclusion. Both new and existing patients were eligible with no age restrictions. Cases were excluded if no images were submitted by patients, images were rated as insufficient quality, patients submitted images but did not present for telemedicine visit, patients had a telemedicine visit but did not subsequently adhere to the recommendation for an in-person visit or diagnostic testing, or upon evaluation, they had no skin condition present for examination.\n\n**Telemedicine clinical workflow**\n\nPatients submitted \u2265 1 clinical images of the skin condition to the consulting dermatologist prior to, or during the telemedicine visit. All patients received standardized imaging instructions, which provided lighting and flash guidelines in addition to recommended poses. Images were sent to the dermatologist via secure email. If necessary, the dermatologist could request an additional set of images. Quality of images was rated by the dermatologist as \u2018insufficient\u2019 vs. \u2018sufficient\u2019. Images of sufficient quality were further rated as \u2018high quality\u2019 or \u2018average quality\u2019. All telemedicine visits were real-time video encounters conducted using a video conferencing platform (Zoom Video Communication, Inc.). After reviewing the images and interviewing the patient, the dermatologist recorded their telemedicine clinical diagnosis and treatment plan into a study-specific data collection form. Patient demographics (age and sex), Fitzpatrick skin type (I-VI) and medical history were recorded. If the dermatologist was unable to reach a diagnosis, the patient was directed for an in-person visit, skin biopsy and/or additional laboratory workup.\n\n**Artificial intelligence index test**\n\nDuring the telemedicine visit, a validated, public, web-based AI algorithm (available online at http://modelderm.com) was used as the AI index test. Briefly, the algorithm was trained by one of the authors (S.S.H.) using the ASAN dataset that consisted of 220,680 images. It was validated using part of the ASAN dataset (17,125 images). The test set consisted of images from Korea (2201 images from the Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital), and the Edinburgh dataset (1300 images). The network structure includes SENet, SE-ResNet-50, and visual geometry group (VGG-19) parallelly concatenated and trained with 174 disease classes. These neural networks (SENet, SE-ResNet-50, and VGG-19) are convolutional neural networks, and the number of hidden layers is 154, 50 and 19, respectively. ImageNet pretrained models of each network were finetuned end-to-end separately. Details can be found in Han et al.\n\nThe AI web application allowed uploading images at different magnifications. As per the instructions for the algorithm, images were uploaded with the skin condition occupying at least 80% of the field of view and with the region of interest (ROI) centred. The dermatologist chose a single image to upload to the web application. Dermatologists were instructed to use the most-representative and highest-quality image for each skin condition. The algorithm output consisted of three diagnoses (from a possible list of 174 skin conditions), ranked in order of probability (ranging from 0 to 1). All algorithm outputs were recorded. This procedure was conducted in real-time, during the visit, with the potential of affecting the outcome of the visit. Given the nature and setting of the AI index test, the reference standard as well as the composition of test classes was inherently blinded. Finally, dermatologists recorded if exposure to the AI algorithm: (a) was clinically useful to them during the telemedicine visit by either increasing their diagnostic confidence and/or expanding their differential diagnosis (yes, no) and (b) led to a change in their telemedicine clinical diagnosis (yes, no).\n\n**Healthcare provider index test**\n\nSubsequent to the telemedicine visit, clinical images were independently evaluated by nine healthcare providers in a reader study. We invited three board-certified dermatologists, three dermatology residents, and three general practitioners (GPs), and all agreed to participate. The dermatologists had a median (min-max) of 3 (2\u20135) years of postresident clinical experience. One resident was in their second year and two were in their third year of dermatology training. The GPs had no formal training in dermatology but had an interest in skin diseases. GPs had a median (min-max) of 7 (1\u201311) years of postmedical school clinical experience. In this experiment, readers were only exposed to the images submitted by the patient and were blinded to clinical history/metadata as well as the results of the AI index test and the reference standard. There were no time restrictions.\n\n**Reference standard**\n\nThe reference standard was defined in one of two ways. First, if the patient was recommended to return for an in-person clinic visit, the diagnosis from this visit (and any associated laboratory testing or skin biopsies) was used as the reference standard. Second, if no in-person clinic visit was performed, a panel of 6 dermatologists evaluated the case and established the reference standard based on consensus agreement. The panel dermatologists had a median (min-max) of 8 (5\u201313) years of postresident clinical experience; there was no overlap between dermatologists who participated in the reader study and the reference standard. During the consensus evaluation, the clinical images and medical history were shown to the panel of dermatologists. All index test results were unavailable during consensus evaluations. Only one diagnosis was permitted for the reference standard. There were no cases of disagreement.\n\n**Statistical analysis**\n\nDescriptive statistics were used to characterize the study data. Diagnoses were grouped into 5 categories and 13 subcategories: (1) \u2018inflammatory\u2019 (subcategories: dermatitis, acne/rosacea, autoimmune, papulosquamous and other); (2) \u2018infectious\u2019 (subcategories: bacterial, viral, fungal and parasitic); (3) \u2018neoplastic\u2019 (subcategories: malignant and benign); (4) \u2018alopecia\u2019 (subcategories: scarring and non-scarring); and (5) \u2018other\u2019 (e.g. burn, scar, striae and among others), using a system adapted from Liu et al.\n\nThe primary outcomes were the \u2018top-1 accuracy\u2019 and the \u2018balanced top-1 accuracy\u2019. The \u2018top-1 accuracy\u2019 measures how frequently the top-1 prediction of the index test matched the reference standard diagnosis. Similar but not identical diagnoses to the reference standard (i.e. herpes simplex vs. herpes zoster) were considered incorrect. The \u2018balanced top-1 accuracy\u2019 is the top-1 accuracy computed for each unique diagnosis separately and then averaged. This was performed to account for the variant prevalence of certain diagnoses (e.g. acne vs hidradenitis suppurativa). Because algorithm fundamentally always predicted incorrect answers for the untrained cases (\u2018out of distribution\u2019), we performed sub-analyses for both measures restricted to the 174 conditions on which the index AI algorithm was explicitly trained (i.e. \u2018in distribution\u2019 vs. \u2018overall\u2019 conditions). A secondary outcome was the \u2018top-3 accuracy\u2019, which measures if the reference standard diagnosis was included in any of 3 diagnoses of the AI algorithm output. \n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.",
        "results": "### Results\n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.\n\nA total of 380 skin conditions were evaluated in the telemedicine clinic during the study period. After exclusion of 40 skin conditions (18 no clinical images, 8 insufficient quality images, 8 lost to follow-up after recommendation of in-person visit, and 6 no active disease), 340 skin conditions (87 unique diagnoses) from 281 patients met the study inclusion criteria. The mean (SD) age of the patients was 33.7 (17.5) years; 37% (n = 104) were male and 63% (n = 177) were female. Of the 340 skin conditions, 190 (55.9%) occurred in Fitzpatrick skin phototype III patients, 84 (24.7%) in type IV patients, 59 (17.3%) in type II patients and 7 (2.1%) in type I patients.\n\nIn 7.1% (n = 24) of visits, additional clinical images were requested and provided by the patients. Overall, 87.4% (n = 297) and 12.6% (n = 43) of images were rated as having \u2018high quality\u2019 and \u2018average quality\u2019, respectively. There were no differences in image quality by diagnostic category (P = 0.971) (Table S1). The reference standard was determined from in-person clinic visits (18.5%, n = 63), skin biopsy (13.8%, n = 47) and consensus agreement of panel dermatologists (81.5%, n = 277). Overall, per the reference standard, 225 of the skin conditions were inflammatory (66.2%), 52 infectious (15.3%), 32 neoplastic (9.4%), 15 alopecia (4.4%) and 16 other (4.7%). There were 87 unique diagnoses; the most frequent were acne (18.5%), contact dermatitis (8.2%), psoriasis (3.8%) and warts (3.6%) (Table 1 and Table S2).\n\n**AI performance**\n\nThe AI algorithm had overall top-1 and balanced top-1 accuracies of 41.2% and 35.1%, respectively (Table 2). Accuracy was associated with the Fitzpatrick skin type with better performance in darker skin types and was not associated with sex (Table S3). Of the 87 unique skin diagnoses included, 63 (72.4%) were part of the AI algorithm training dataset. Therefore, 305 out of the 340 (89.7%) skin conditions were \u2018in distribution\u2019 and 35 were \u2018out of distribution\u2019 images. When limiting the analysis to the \u2018in distribution\u2019 skin conditions, the top-1 and balanced top-1 accuracies increased to 45.3% and 47.6%, respectively (Table 3).\n\n**AI performance by disease category/subcategory**\n\nTop-1 accuracy was highest in the \u2018other\u2019 category (68.8%), followed by the \u2018infectious\u2019 (50%), \u2018inflammatory\u2019 (38.2%), \u2018neoplastic\u2019 (37.5%) and \u2018alopecia\u2019 (33.3%) categories (P = 0.089). For the subcategories, \u2018non-scarring alopecia\u2019 had the highest top-1 accuracy (71.4%), followed by \u2018papulosquamous\u2019 (52.9%), \u2018fungal\u2019 (52.4%), \u2018viral\u2019 (50%), \u2018bacterial\u2019 (50%), \u2018acne/rosacea\u2019 (47.4%), benign neoplasms (42.9%), \u2018dermatitis\u2019 (34.1%), \u2018autoimmune\u2019 (26.5%), \u2018parasitic\u2019 (0%), \u2018malignant neoplasms\u2019 (0%) and \u2018scarring alopecia\u2019 (0%) (P = 0.035) (Table 2).\n\n**AI vs. Physician performance**\n\nThe overall top-1 accuracy of the AI algorithm (41.2%) was lower than that of the dermatologists (60.1%; P < 0.001), residents (57.8%; P < 0.001) and GPs (49.3%; P = 0.034) (Table 2, Figure 1). The balanced top-1 accuracy of the AI algorithm was 35.1%, lower than the dermatologists (45.2%; P = 0.007) and the residents (42.6%: P = 0.045), but not different than the GPs (33.1%; P = 0.582).\n\nThe overall top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (45.3%) was lower than that of the dermatologists (63.1%) and residents (60.9%) (P < 0.001) but similar to GPs (52.9%, P = 0.061) (Table 3, Figure 1). The balanced top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (47.6%) was similar to that of the dermatologists (49.7%, P = 0.604) and residents (47.7%, P = 0.980) and superior than the GPs (39.7%, P = 0.049).\n\n**Accuracy by image quality**\n\nThe algorithm achieved a higher top-1 accuracy in images of \u2018high quality\u2019 than in images of \u2018average quality\u2019 (44.4% vs. 18.6%, P = 0.001). In contrast, there were no differences in the performance of the readers by image quality (Table S4).\n\n**AI utility**\n\nOverall, use of the AI algorithm was considered useful in 11.8% (n = 40) of cases, by increasing diagnostic confidence and/or expanding the differential diagnosis (Figure 2). The top-1 accuracy of these 40 photographs was 50%, similar to the overall top-1 accuracy of the 340 images (41.2%, P = 0.286). The top-3 accuracy was 75.9%. The \u2018balanced\u2019 top-1 accuracy for these 40 images was 54.9%, higher than the balanced top-1 accuracy of the 340 images (35.1%, P = 0.014). \u2018In distribution\u2019 pathologies included in these 40 cases were inflammatory (n = 17), infectious (n = 12) and neoplastic (n = 8) (Table S5). Exposure to the AI algorithm modified the real-time diagnosis of the teledermatologist in 0.9% (n = 3) of cases; in 2 cases the diagnosis was correctly modified (tinea corporis and perioral dermatitis) and in 1 case neither the clinician nor the algorithm matched the reference standard (sarcoidosis).",
        "column_title": "Applicability concerns | Patient selection",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Performance of a deep neural network in teledermatology: a single-centre prospective diagnostic study.",
        "abstract": "## BACKGROUND\nThe use of artificial intelligence (AI) algorithms for the diagnosis of skin diseases has shown promise in experimental settings but has not been yet tested in real-life conditions.\n## OBJECTIVE\nTo assess the diagnostic performance and potential clinical utility of a 174-multiclass AI algorithm in a real-life telemedicine setting.\n## METHODS\nProspective, diagnostic accuracy study including consecutive patients who submitted images for teledermatology evaluation. The treating dermatologist chose a single image to upload to a web application during teleconsultation. A follow-up reader study including nine healthcare providers (3 dermatologists, 3 dermatology residents and 3 general practitioners) was performed.\n## RESULTS\nA total of 340 cases from 281 patients met study inclusion criteria. The mean (SD) age of patients was 33.7 (17.5) years; 63% (n\u00a0=\u00a0177) were female. Exposure to the AI algorithm results was considered useful in 11.8% of visits (n\u00a0=\u00a040) and the teledermatologist correctly modified the real-time diagnosis in 0.6% (n\u00a0=\u00a02) of cases. The overall top-1 accuracy of the algorithm (41.2%) was lower than that of the dermatologists (60.1%), residents (57.8%) and general practitioners (49.3%) (all comparisons P\u00a0<\u00a00.05, in the reader study). When the analysis was limited to the diagnoses on which the algorithm had been explicitly trained, the balanced top-1 accuracy of the algorithm (47.6%) was comparable to the dermatologists (49.7%) and residents (47.7%) but superior to the general practitioners (39.7%; P\u00a0=\u00a00.049). Algorithm performance was associated with patient skin type and image quality.\n## CONCLUSIONS\nA 174-disease class AI algorithm appears to be a promising tool in the triage and evaluation of lesions with patient-taken photographs via telemedicine.\n",
        "methods": "### Methods\n\nThis was an IRB-approved, single-centre, prospective, diagnostic accuracy study conducted between 27 March and 30 April 2020. All data collection was explicitly planned prior to the performance of the index tests and reference standard and the study was reported following the STARD 2015 guidelines. Consecutive patients who submitted clinical images of one or more new skin conditions for evaluation in our telemedicine dermatology clinic were considered eligible for inclusion. Both new and existing patients were eligible with no age restrictions. Cases were excluded if no images were submitted by patients, images were rated as insufficient quality, patients submitted images but did not present for telemedicine visit, patients had a telemedicine visit but did not subsequently adhere to the recommendation for an in-person visit or diagnostic testing, or upon evaluation, they had no skin condition present for examination.\n\n**Telemedicine clinical workflow**\n\nPatients submitted \u2265 1 clinical images of the skin condition to the consulting dermatologist prior to, or during the telemedicine visit. All patients received standardized imaging instructions, which provided lighting and flash guidelines in addition to recommended poses. Images were sent to the dermatologist via secure email. If necessary, the dermatologist could request an additional set of images. Quality of images was rated by the dermatologist as \u2018insufficient\u2019 vs. \u2018sufficient\u2019. Images of sufficient quality were further rated as \u2018high quality\u2019 or \u2018average quality\u2019. All telemedicine visits were real-time video encounters conducted using a video conferencing platform (Zoom Video Communication, Inc.). After reviewing the images and interviewing the patient, the dermatologist recorded their telemedicine clinical diagnosis and treatment plan into a study-specific data collection form. Patient demographics (age and sex), Fitzpatrick skin type (I-VI) and medical history were recorded. If the dermatologist was unable to reach a diagnosis, the patient was directed for an in-person visit, skin biopsy and/or additional laboratory workup.\n\n**Artificial intelligence index test**\n\nDuring the telemedicine visit, a validated, public, web-based AI algorithm (available online at http://modelderm.com) was used as the AI index test. Briefly, the algorithm was trained by one of the authors (S.S.H.) using the ASAN dataset that consisted of 220,680 images. It was validated using part of the ASAN dataset (17,125 images). The test set consisted of images from Korea (2201 images from the Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital), and the Edinburgh dataset (1300 images). The network structure includes SENet, SE-ResNet-50, and visual geometry group (VGG-19) parallelly concatenated and trained with 174 disease classes. These neural networks (SENet, SE-ResNet-50, and VGG-19) are convolutional neural networks, and the number of hidden layers is 154, 50 and 19, respectively. ImageNet pretrained models of each network were finetuned end-to-end separately. Details can be found in Han et al.\n\nThe AI web application allowed uploading images at different magnifications. As per the instructions for the algorithm, images were uploaded with the skin condition occupying at least 80% of the field of view and with the region of interest (ROI) centred. The dermatologist chose a single image to upload to the web application. Dermatologists were instructed to use the most-representative and highest-quality image for each skin condition. The algorithm output consisted of three diagnoses (from a possible list of 174 skin conditions), ranked in order of probability (ranging from 0 to 1). All algorithm outputs were recorded. This procedure was conducted in real-time, during the visit, with the potential of affecting the outcome of the visit. Given the nature and setting of the AI index test, the reference standard as well as the composition of test classes was inherently blinded. Finally, dermatologists recorded if exposure to the AI algorithm: (a) was clinically useful to them during the telemedicine visit by either increasing their diagnostic confidence and/or expanding their differential diagnosis (yes, no) and (b) led to a change in their telemedicine clinical diagnosis (yes, no).\n\n**Healthcare provider index test**\n\nSubsequent to the telemedicine visit, clinical images were independently evaluated by nine healthcare providers in a reader study. We invited three board-certified dermatologists, three dermatology residents, and three general practitioners (GPs), and all agreed to participate. The dermatologists had a median (min-max) of 3 (2\u20135) years of postresident clinical experience. One resident was in their second year and two were in their third year of dermatology training. The GPs had no formal training in dermatology but had an interest in skin diseases. GPs had a median (min-max) of 7 (1\u201311) years of postmedical school clinical experience. In this experiment, readers were only exposed to the images submitted by the patient and were blinded to clinical history/metadata as well as the results of the AI index test and the reference standard. There were no time restrictions.\n\n**Reference standard**\n\nThe reference standard was defined in one of two ways. First, if the patient was recommended to return for an in-person clinic visit, the diagnosis from this visit (and any associated laboratory testing or skin biopsies) was used as the reference standard. Second, if no in-person clinic visit was performed, a panel of 6 dermatologists evaluated the case and established the reference standard based on consensus agreement. The panel dermatologists had a median (min-max) of 8 (5\u201313) years of postresident clinical experience; there was no overlap between dermatologists who participated in the reader study and the reference standard. During the consensus evaluation, the clinical images and medical history were shown to the panel of dermatologists. All index test results were unavailable during consensus evaluations. Only one diagnosis was permitted for the reference standard. There were no cases of disagreement.\n\n**Statistical analysis**\n\nDescriptive statistics were used to characterize the study data. Diagnoses were grouped into 5 categories and 13 subcategories: (1) \u2018inflammatory\u2019 (subcategories: dermatitis, acne/rosacea, autoimmune, papulosquamous and other); (2) \u2018infectious\u2019 (subcategories: bacterial, viral, fungal and parasitic); (3) \u2018neoplastic\u2019 (subcategories: malignant and benign); (4) \u2018alopecia\u2019 (subcategories: scarring and non-scarring); and (5) \u2018other\u2019 (e.g. burn, scar, striae and among others), using a system adapted from Liu et al.\n\nThe primary outcomes were the \u2018top-1 accuracy\u2019 and the \u2018balanced top-1 accuracy\u2019. The \u2018top-1 accuracy\u2019 measures how frequently the top-1 prediction of the index test matched the reference standard diagnosis. Similar but not identical diagnoses to the reference standard (i.e. herpes simplex vs. herpes zoster) were considered incorrect. The \u2018balanced top-1 accuracy\u2019 is the top-1 accuracy computed for each unique diagnosis separately and then averaged. This was performed to account for the variant prevalence of certain diagnoses (e.g. acne vs hidradenitis suppurativa). Because algorithm fundamentally always predicted incorrect answers for the untrained cases (\u2018out of distribution\u2019), we performed sub-analyses for both measures restricted to the 174 conditions on which the index AI algorithm was explicitly trained (i.e. \u2018in distribution\u2019 vs. \u2018overall\u2019 conditions). A secondary outcome was the \u2018top-3 accuracy\u2019, which measures if the reference standard diagnosis was included in any of 3 diagnoses of the AI algorithm output. \n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.",
        "results": "### Results\n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.\n\nA total of 380 skin conditions were evaluated in the telemedicine clinic during the study period. After exclusion of 40 skin conditions (18 no clinical images, 8 insufficient quality images, 8 lost to follow-up after recommendation of in-person visit, and 6 no active disease), 340 skin conditions (87 unique diagnoses) from 281 patients met the study inclusion criteria. The mean (SD) age of the patients was 33.7 (17.5) years; 37% (n = 104) were male and 63% (n = 177) were female. Of the 340 skin conditions, 190 (55.9%) occurred in Fitzpatrick skin phototype III patients, 84 (24.7%) in type IV patients, 59 (17.3%) in type II patients and 7 (2.1%) in type I patients.\n\nIn 7.1% (n = 24) of visits, additional clinical images were requested and provided by the patients. Overall, 87.4% (n = 297) and 12.6% (n = 43) of images were rated as having \u2018high quality\u2019 and \u2018average quality\u2019, respectively. There were no differences in image quality by diagnostic category (P = 0.971) (Table S1). The reference standard was determined from in-person clinic visits (18.5%, n = 63), skin biopsy (13.8%, n = 47) and consensus agreement of panel dermatologists (81.5%, n = 277). Overall, per the reference standard, 225 of the skin conditions were inflammatory (66.2%), 52 infectious (15.3%), 32 neoplastic (9.4%), 15 alopecia (4.4%) and 16 other (4.7%). There were 87 unique diagnoses; the most frequent were acne (18.5%), contact dermatitis (8.2%), psoriasis (3.8%) and warts (3.6%) (Table 1 and Table S2).\n\n**AI performance**\n\nThe AI algorithm had overall top-1 and balanced top-1 accuracies of 41.2% and 35.1%, respectively (Table 2). Accuracy was associated with the Fitzpatrick skin type with better performance in darker skin types and was not associated with sex (Table S3). Of the 87 unique skin diagnoses included, 63 (72.4%) were part of the AI algorithm training dataset. Therefore, 305 out of the 340 (89.7%) skin conditions were \u2018in distribution\u2019 and 35 were \u2018out of distribution\u2019 images. When limiting the analysis to the \u2018in distribution\u2019 skin conditions, the top-1 and balanced top-1 accuracies increased to 45.3% and 47.6%, respectively (Table 3).\n\n**AI performance by disease category/subcategory**\n\nTop-1 accuracy was highest in the \u2018other\u2019 category (68.8%), followed by the \u2018infectious\u2019 (50%), \u2018inflammatory\u2019 (38.2%), \u2018neoplastic\u2019 (37.5%) and \u2018alopecia\u2019 (33.3%) categories (P = 0.089). For the subcategories, \u2018non-scarring alopecia\u2019 had the highest top-1 accuracy (71.4%), followed by \u2018papulosquamous\u2019 (52.9%), \u2018fungal\u2019 (52.4%), \u2018viral\u2019 (50%), \u2018bacterial\u2019 (50%), \u2018acne/rosacea\u2019 (47.4%), benign neoplasms (42.9%), \u2018dermatitis\u2019 (34.1%), \u2018autoimmune\u2019 (26.5%), \u2018parasitic\u2019 (0%), \u2018malignant neoplasms\u2019 (0%) and \u2018scarring alopecia\u2019 (0%) (P = 0.035) (Table 2).\n\n**AI vs. Physician performance**\n\nThe overall top-1 accuracy of the AI algorithm (41.2%) was lower than that of the dermatologists (60.1%; P < 0.001), residents (57.8%; P < 0.001) and GPs (49.3%; P = 0.034) (Table 2, Figure 1). The balanced top-1 accuracy of the AI algorithm was 35.1%, lower than the dermatologists (45.2%; P = 0.007) and the residents (42.6%: P = 0.045), but not different than the GPs (33.1%; P = 0.582).\n\nThe overall top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (45.3%) was lower than that of the dermatologists (63.1%) and residents (60.9%) (P < 0.001) but similar to GPs (52.9%, P = 0.061) (Table 3, Figure 1). The balanced top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (47.6%) was similar to that of the dermatologists (49.7%, P = 0.604) and residents (47.7%, P = 0.980) and superior than the GPs (39.7%, P = 0.049).\n\n**Accuracy by image quality**\n\nThe algorithm achieved a higher top-1 accuracy in images of \u2018high quality\u2019 than in images of \u2018average quality\u2019 (44.4% vs. 18.6%, P = 0.001). In contrast, there were no differences in the performance of the readers by image quality (Table S4).\n\n**AI utility**\n\nOverall, use of the AI algorithm was considered useful in 11.8% (n = 40) of cases, by increasing diagnostic confidence and/or expanding the differential diagnosis (Figure 2). The top-1 accuracy of these 40 photographs was 50%, similar to the overall top-1 accuracy of the 340 images (41.2%, P = 0.286). The top-3 accuracy was 75.9%. The \u2018balanced\u2019 top-1 accuracy for these 40 images was 54.9%, higher than the balanced top-1 accuracy of the 340 images (35.1%, P = 0.014). \u2018In distribution\u2019 pathologies included in these 40 cases were inflammatory (n = 17), infectious (n = 12) and neoplastic (n = 8) (Table S5). Exposure to the AI algorithm modified the real-time diagnosis of the teledermatologist in 0.9% (n = 3) of cases; in 2 cases the diagnosis was correctly modified (tinea corporis and perioral dermatitis) and in 1 case neither the clinician nor the algorithm matched the reference standard (sarcoidosis).",
        "column_title": "Applicability concerns | Index test",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "38594408",
        "meta_analysis_title": "Human-AI interaction in skin cancer diagnosis: a systematic review and meta-analysis.",
        "title": "Performance of a deep neural network in teledermatology: a single-centre prospective diagnostic study.",
        "abstract": "## BACKGROUND\nThe use of artificial intelligence (AI) algorithms for the diagnosis of skin diseases has shown promise in experimental settings but has not been yet tested in real-life conditions.\n## OBJECTIVE\nTo assess the diagnostic performance and potential clinical utility of a 174-multiclass AI algorithm in a real-life telemedicine setting.\n## METHODS\nProspective, diagnostic accuracy study including consecutive patients who submitted images for teledermatology evaluation. The treating dermatologist chose a single image to upload to a web application during teleconsultation. A follow-up reader study including nine healthcare providers (3 dermatologists, 3 dermatology residents and 3 general practitioners) was performed.\n## RESULTS\nA total of 340 cases from 281 patients met study inclusion criteria. The mean (SD) age of patients was 33.7 (17.5) years; 63% (n\u00a0=\u00a0177) were female. Exposure to the AI algorithm results was considered useful in 11.8% of visits (n\u00a0=\u00a040) and the teledermatologist correctly modified the real-time diagnosis in 0.6% (n\u00a0=\u00a02) of cases. The overall top-1 accuracy of the algorithm (41.2%) was lower than that of the dermatologists (60.1%), residents (57.8%) and general practitioners (49.3%) (all comparisons P\u00a0<\u00a00.05, in the reader study). When the analysis was limited to the diagnoses on which the algorithm had been explicitly trained, the balanced top-1 accuracy of the algorithm (47.6%) was comparable to the dermatologists (49.7%) and residents (47.7%) but superior to the general practitioners (39.7%; P\u00a0=\u00a00.049). Algorithm performance was associated with patient skin type and image quality.\n## CONCLUSIONS\nA 174-disease class AI algorithm appears to be a promising tool in the triage and evaluation of lesions with patient-taken photographs via telemedicine.\n",
        "methods": "### Methods\n\nThis was an IRB-approved, single-centre, prospective, diagnostic accuracy study conducted between 27 March and 30 April 2020. All data collection was explicitly planned prior to the performance of the index tests and reference standard and the study was reported following the STARD 2015 guidelines. Consecutive patients who submitted clinical images of one or more new skin conditions for evaluation in our telemedicine dermatology clinic were considered eligible for inclusion. Both new and existing patients were eligible with no age restrictions. Cases were excluded if no images were submitted by patients, images were rated as insufficient quality, patients submitted images but did not present for telemedicine visit, patients had a telemedicine visit but did not subsequently adhere to the recommendation for an in-person visit or diagnostic testing, or upon evaluation, they had no skin condition present for examination.\n\n**Telemedicine clinical workflow**\n\nPatients submitted \u2265 1 clinical images of the skin condition to the consulting dermatologist prior to, or during the telemedicine visit. All patients received standardized imaging instructions, which provided lighting and flash guidelines in addition to recommended poses. Images were sent to the dermatologist via secure email. If necessary, the dermatologist could request an additional set of images. Quality of images was rated by the dermatologist as \u2018insufficient\u2019 vs. \u2018sufficient\u2019. Images of sufficient quality were further rated as \u2018high quality\u2019 or \u2018average quality\u2019. All telemedicine visits were real-time video encounters conducted using a video conferencing platform (Zoom Video Communication, Inc.). After reviewing the images and interviewing the patient, the dermatologist recorded their telemedicine clinical diagnosis and treatment plan into a study-specific data collection form. Patient demographics (age and sex), Fitzpatrick skin type (I-VI) and medical history were recorded. If the dermatologist was unable to reach a diagnosis, the patient was directed for an in-person visit, skin biopsy and/or additional laboratory workup.\n\n**Artificial intelligence index test**\n\nDuring the telemedicine visit, a validated, public, web-based AI algorithm (available online at http://modelderm.com) was used as the AI index test. Briefly, the algorithm was trained by one of the authors (S.S.H.) using the ASAN dataset that consisted of 220,680 images. It was validated using part of the ASAN dataset (17,125 images). The test set consisted of images from Korea (2201 images from the Seoul National University Bundang Hospital, Inje University Sanggye Paik Hospital, and Hallym University Dongtan Hospital), and the Edinburgh dataset (1300 images). The network structure includes SENet, SE-ResNet-50, and visual geometry group (VGG-19) parallelly concatenated and trained with 174 disease classes. These neural networks (SENet, SE-ResNet-50, and VGG-19) are convolutional neural networks, and the number of hidden layers is 154, 50 and 19, respectively. ImageNet pretrained models of each network were finetuned end-to-end separately. Details can be found in Han et al.\n\nThe AI web application allowed uploading images at different magnifications. As per the instructions for the algorithm, images were uploaded with the skin condition occupying at least 80% of the field of view and with the region of interest (ROI) centred. The dermatologist chose a single image to upload to the web application. Dermatologists were instructed to use the most-representative and highest-quality image for each skin condition. The algorithm output consisted of three diagnoses (from a possible list of 174 skin conditions), ranked in order of probability (ranging from 0 to 1). All algorithm outputs were recorded. This procedure was conducted in real-time, during the visit, with the potential of affecting the outcome of the visit. Given the nature and setting of the AI index test, the reference standard as well as the composition of test classes was inherently blinded. Finally, dermatologists recorded if exposure to the AI algorithm: (a) was clinically useful to them during the telemedicine visit by either increasing their diagnostic confidence and/or expanding their differential diagnosis (yes, no) and (b) led to a change in their telemedicine clinical diagnosis (yes, no).\n\n**Healthcare provider index test**\n\nSubsequent to the telemedicine visit, clinical images were independently evaluated by nine healthcare providers in a reader study. We invited three board-certified dermatologists, three dermatology residents, and three general practitioners (GPs), and all agreed to participate. The dermatologists had a median (min-max) of 3 (2\u20135) years of postresident clinical experience. One resident was in their second year and two were in their third year of dermatology training. The GPs had no formal training in dermatology but had an interest in skin diseases. GPs had a median (min-max) of 7 (1\u201311) years of postmedical school clinical experience. In this experiment, readers were only exposed to the images submitted by the patient and were blinded to clinical history/metadata as well as the results of the AI index test and the reference standard. There were no time restrictions.\n\n**Reference standard**\n\nThe reference standard was defined in one of two ways. First, if the patient was recommended to return for an in-person clinic visit, the diagnosis from this visit (and any associated laboratory testing or skin biopsies) was used as the reference standard. Second, if no in-person clinic visit was performed, a panel of 6 dermatologists evaluated the case and established the reference standard based on consensus agreement. The panel dermatologists had a median (min-max) of 8 (5\u201313) years of postresident clinical experience; there was no overlap between dermatologists who participated in the reader study and the reference standard. During the consensus evaluation, the clinical images and medical history were shown to the panel of dermatologists. All index test results were unavailable during consensus evaluations. Only one diagnosis was permitted for the reference standard. There were no cases of disagreement.\n\n**Statistical analysis**\n\nDescriptive statistics were used to characterize the study data. Diagnoses were grouped into 5 categories and 13 subcategories: (1) \u2018inflammatory\u2019 (subcategories: dermatitis, acne/rosacea, autoimmune, papulosquamous and other); (2) \u2018infectious\u2019 (subcategories: bacterial, viral, fungal and parasitic); (3) \u2018neoplastic\u2019 (subcategories: malignant and benign); (4) \u2018alopecia\u2019 (subcategories: scarring and non-scarring); and (5) \u2018other\u2019 (e.g. burn, scar, striae and among others), using a system adapted from Liu et al.\n\nThe primary outcomes were the \u2018top-1 accuracy\u2019 and the \u2018balanced top-1 accuracy\u2019. The \u2018top-1 accuracy\u2019 measures how frequently the top-1 prediction of the index test matched the reference standard diagnosis. Similar but not identical diagnoses to the reference standard (i.e. herpes simplex vs. herpes zoster) were considered incorrect. The \u2018balanced top-1 accuracy\u2019 is the top-1 accuracy computed for each unique diagnosis separately and then averaged. This was performed to account for the variant prevalence of certain diagnoses (e.g. acne vs hidradenitis suppurativa). Because algorithm fundamentally always predicted incorrect answers for the untrained cases (\u2018out of distribution\u2019), we performed sub-analyses for both measures restricted to the 174 conditions on which the index AI algorithm was explicitly trained (i.e. \u2018in distribution\u2019 vs. \u2018overall\u2019 conditions). A secondary outcome was the \u2018top-3 accuracy\u2019, which measures if the reference standard diagnosis was included in any of 3 diagnoses of the AI algorithm output. \n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.",
        "results": "### Results\n\nShapiro\u2013Wilk test was used for tests of normality, Z-tests were used to compare proportions, and Chi-square and Fisher exact tests were used to evaluate the association of categorical variables. Statistical analysis was performed using STATA 14.0\u00ae. An alpha value below 0.05 was considered statistically significant.\n\nA total of 380 skin conditions were evaluated in the telemedicine clinic during the study period. After exclusion of 40 skin conditions (18 no clinical images, 8 insufficient quality images, 8 lost to follow-up after recommendation of in-person visit, and 6 no active disease), 340 skin conditions (87 unique diagnoses) from 281 patients met the study inclusion criteria. The mean (SD) age of the patients was 33.7 (17.5) years; 37% (n = 104) were male and 63% (n = 177) were female. Of the 340 skin conditions, 190 (55.9%) occurred in Fitzpatrick skin phototype III patients, 84 (24.7%) in type IV patients, 59 (17.3%) in type II patients and 7 (2.1%) in type I patients.\n\nIn 7.1% (n = 24) of visits, additional clinical images were requested and provided by the patients. Overall, 87.4% (n = 297) and 12.6% (n = 43) of images were rated as having \u2018high quality\u2019 and \u2018average quality\u2019, respectively. There were no differences in image quality by diagnostic category (P = 0.971) (Table S1). The reference standard was determined from in-person clinic visits (18.5%, n = 63), skin biopsy (13.8%, n = 47) and consensus agreement of panel dermatologists (81.5%, n = 277). Overall, per the reference standard, 225 of the skin conditions were inflammatory (66.2%), 52 infectious (15.3%), 32 neoplastic (9.4%), 15 alopecia (4.4%) and 16 other (4.7%). There were 87 unique diagnoses; the most frequent were acne (18.5%), contact dermatitis (8.2%), psoriasis (3.8%) and warts (3.6%) (Table 1 and Table S2).\n\n**AI performance**\n\nThe AI algorithm had overall top-1 and balanced top-1 accuracies of 41.2% and 35.1%, respectively (Table 2). Accuracy was associated with the Fitzpatrick skin type with better performance in darker skin types and was not associated with sex (Table S3). Of the 87 unique skin diagnoses included, 63 (72.4%) were part of the AI algorithm training dataset. Therefore, 305 out of the 340 (89.7%) skin conditions were \u2018in distribution\u2019 and 35 were \u2018out of distribution\u2019 images. When limiting the analysis to the \u2018in distribution\u2019 skin conditions, the top-1 and balanced top-1 accuracies increased to 45.3% and 47.6%, respectively (Table 3).\n\n**AI performance by disease category/subcategory**\n\nTop-1 accuracy was highest in the \u2018other\u2019 category (68.8%), followed by the \u2018infectious\u2019 (50%), \u2018inflammatory\u2019 (38.2%), \u2018neoplastic\u2019 (37.5%) and \u2018alopecia\u2019 (33.3%) categories (P = 0.089). For the subcategories, \u2018non-scarring alopecia\u2019 had the highest top-1 accuracy (71.4%), followed by \u2018papulosquamous\u2019 (52.9%), \u2018fungal\u2019 (52.4%), \u2018viral\u2019 (50%), \u2018bacterial\u2019 (50%), \u2018acne/rosacea\u2019 (47.4%), benign neoplasms (42.9%), \u2018dermatitis\u2019 (34.1%), \u2018autoimmune\u2019 (26.5%), \u2018parasitic\u2019 (0%), \u2018malignant neoplasms\u2019 (0%) and \u2018scarring alopecia\u2019 (0%) (P = 0.035) (Table 2).\n\n**AI vs. Physician performance**\n\nThe overall top-1 accuracy of the AI algorithm (41.2%) was lower than that of the dermatologists (60.1%; P < 0.001), residents (57.8%; P < 0.001) and GPs (49.3%; P = 0.034) (Table 2, Figure 1). The balanced top-1 accuracy of the AI algorithm was 35.1%, lower than the dermatologists (45.2%; P = 0.007) and the residents (42.6%: P = 0.045), but not different than the GPs (33.1%; P = 0.582).\n\nThe overall top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (45.3%) was lower than that of the dermatologists (63.1%) and residents (60.9%) (P < 0.001) but similar to GPs (52.9%, P = 0.061) (Table 3, Figure 1). The balanced top-1 \u2018in distribution\u2019 accuracy of the AI algorithm (47.6%) was similar to that of the dermatologists (49.7%, P = 0.604) and residents (47.7%, P = 0.980) and superior than the GPs (39.7%, P = 0.049).\n\n**Accuracy by image quality**\n\nThe algorithm achieved a higher top-1 accuracy in images of \u2018high quality\u2019 than in images of \u2018average quality\u2019 (44.4% vs. 18.6%, P = 0.001). In contrast, there were no differences in the performance of the readers by image quality (Table S4).\n\n**AI utility**\n\nOverall, use of the AI algorithm was considered useful in 11.8% (n = 40) of cases, by increasing diagnostic confidence and/or expanding the differential diagnosis (Figure 2). The top-1 accuracy of these 40 photographs was 50%, similar to the overall top-1 accuracy of the 340 images (41.2%, P = 0.286). The top-3 accuracy was 75.9%. The \u2018balanced\u2019 top-1 accuracy for these 40 images was 54.9%, higher than the balanced top-1 accuracy of the 340 images (35.1%, P = 0.014). \u2018In distribution\u2019 pathologies included in these 40 cases were inflammatory (n = 17), infectious (n = 12) and neoplastic (n = 8) (Table S5). Exposure to the AI algorithm modified the real-time diagnosis of the teledermatologist in 0.9% (n = 3) of cases; in 2 cases the diagnosis was correctly modified (tinea corporis and perioral dermatitis) and in 1 case neither the clinician nor the algorithm matched the reference standard (sarcoidosis).",
        "column_title": "Applicability concerns | Reference standard",
        "instruction": "Evaluate the given study and choose the most suitable label from the following:\n- Low\n- High\n- Unclear",
        "type": "label",
        "truth": "Low"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Beverage Intake and Metabolic Syndrome Risk Over 14 Years: The Study of Women's Health Across the Nation.",
        "abstract": "## BACKGROUND\nAlcohol and energy-dense beverages consumption have been implicated in cardiometabolic disease, albeit inconsistently.\n## OBJECTIVE\nThis study tested prospective associations between intakes of alcohol, energy-dense beverages, and low-calorie beverages and cardiometabolic risk in midlife women.\n## DESIGN\nThe Study of Women's Health Across the Nation is a 14-year, multisite prospective cohort study (1996-2011). Beverage intake and cardiometabolic risk factors that define the metabolic syndrome (hypertension, abdominal obesity, impaired fasting glucose, low high-density lipoprotein cholesterol level, and hypertriglyceridemia) were assessed throughout follow-up.\n## PARTICIPANTS/SETTING\nParticipants (N=1,448) were African American, Chinese, Japanese, and non-Hispanic white midlife women from six US cities.\n## MAIN OUTCOME MEASURES\nThe primary outcomes were incident metabolic syndrome and the individual metabolic syndrome components.\n## STATISTICAL ANALYSES PERFORMED\nGeneralized linear mixed models tested associations between intakes within each beverage category and odds of meeting criteria for metabolic syndrome and each of the metabolic syndrome components.\n## RESULTS\nEnergy-dense beverage consumption was highest among African-American women and lowest among women with college degrees. Non-Hispanic white women consumed the largest quantities of alcohol. Independent of energy intake and potential confounders, each additional 355 mL energy-dense beverages consumed per day was associated with higher odds of developing metabolic syndrome in each successive year of follow-up (odds ratio [OR] 1.05, 95% CI 1.02 to 1.08). Greater energy-dense beverage intake was associated with more rapidly increasing odds of developing hypertension (OR 1.06, 95% CI 1.02 to 1.11) and abdominal obesity (OR 1.10, 95% CI 1.03 to 1.16) over time, but not with the other metabolic syndrome components. Intakes of alcohol and low-calorie coffees, teas, and diet cola were not associated with metabolic syndrome risk.\n## CONCLUSIONS\nOver 14 years of follow-up, energy-dense nonalcoholic beverage consumption was associated with incident metabolic syndrome in midlife women. The observed differences in intakes by ethnicity/race and education suggest that consumption of these beverages may contribute to disparities in risk factors for diabetes and cardiovascular disease.",
        "methods": "### Methods\n\n**Participants**\n\nParticipants were midlife women from seven U.S. regions who were enrolled in the Study of Women's Health Across the Nation (SWAN), a multiethnic/racial, longitudinal study of aging and health determinants across the menopausal transition. SWAN's methodology is detailed elsewhere.[25] Procedures were approved by each site's institutional review board, and all participants provided written informed consent.\n\nWomen were recruited from five ethnic/racial groups: African-American (Boston, Chicago, Detroit area, and Pittsburgh), Chinese (Oakland area), Hispanic (Newark), Japanese (Los Angeles), and non-Hispanic white (all sites). Eligibility criteria included being aged 42-52 years at enrollment (1996-1997), having an intact uterus and at least one ovary, not being pregnant or lactating, not using oral contraceptives or reproductive hormone therapy, and reporting at least one menstrual cycle in the prior three months. This report includes data collected at the baseline assessment (1996-1997) and annual follow-up assessments that occurred through 2011. Assessments occurred either in clinic settings or participants' homes. However, participants from the Newark site were excluded from this analysis due to high attrition and lack of dietary data after visit 5. Of 2,870 enrolled women across the other 6 sites, 1,820 (63%) had at least nine years of follow-up data on both beverage intake and the metabolic syndrome components. Women who met criteria for the metabolic syndrome at baseline (n=372; 20%) were excluded as this analysis focused on incident cases. The final analytic sample included 1,448 women.\n\n**Measures**\n\n**Waist circumference\u2014Waist circumference was measured to the nearest 0.1 cm with a measuring tape placed horizontally around the participant at the narrowest part of the torso. Measurements were taken over light clothing at the end of a normal exhalation.**\n\n**Blood pressure\u2014Blood pressure was measured on the right arm in a seated position after at least 5 minutes of rest using a mercury sphygmomanometer. Respondents had not smoked or consumed caffeine within the prior 30 minutes. Appropriate cuff size was determined based on arm circumference. Two sequential blood pressure values were completed at least two minutes apart and averaged.**\n\n**Blood assays\u2014Blood was drawn after a 12-hour fast. Samples were frozen and shipped to either Medical Research Laboratories (Lexington, KY; baseline to visit 7) or University of Michigan Pathology Laboratory (after visit 7), which are certified by the National Heart Lung and Blood Institute, Centers for Disease Control Part III program.[26] Triglycerides were analyzed by enzymatic methods on a Hitachi 747 analyzer (Boehringer Mannheim Diagnostics, Indianapolis, IN). HDL was isolated using heparin-2M manganese chloride. Glucose was determined from fasting serum samples using a hexokinase-coupled reaction (Boehringer Mannheim Diagnostics, Indianapolis, IN).**\n\n**Definition of metabolic syndrome components and status\u2014Consistent with current diagnostic criteria,[1] metabolic syndrome was defined by the presence of at least three of the following five components: abdominal obesity (waist circumference \u226580 cm for Chinese and Japanese women, and \u226588 cm for other ethnic/racial groups), hypertension (systolic \u2265130 mmHg, diastolic \u226585 mmHg, or taking blood pressure medication), low HDL (<50 mg/dl), hypertriglyceridemia (fasting triglycerides \u2265150 mg/dl), and impaired fasting glucose (fasting glucose>100 mg/dl, or taking medication for impaired glucose regulation). Metabolic syndrome status and the total number of metabolic syndrome diagnostic components were determined at each time point.**\n\n**Beverage and total calorie intake\u2014The SWAN Food Frequency Questionnaire [FFQ[27]], which is an adaptation of the 1995 version of the Block Food Frequency Questionnaire,[28] was completed at baseline, visit 5, and visit 9. The FFQ was interviewer-administered. Missing and double-marked items were not scored. The SWAN FFQ assesses the usual frequency and portion sizes consumed for 83 foods and 20 beverages over the past year or so. Due to the multiethnic/racial nature of SWAN, the FFQ was administered in English, Spanish, Chinese (Cantonese), and Japanese. The Cantonese and Japanese FFQ versions contained the same 103 core items as the English version, plus 12-16 additional ethnic foods specific to each group. To avoid under-reporting of ethnic foods, the Chinese and Japanese ethnic food items were administered to all ethnic/racial groups at visits 5 and 9. The FFQ yields an estimate of total daily energy intake. Additionally, typical intakes of 20 beverages and beverage condiments were estimated by multiplying the reported portion sizes typically consumed by the reported frequency of consumption. To account for the contribution of beverage condiments such as cream, milk, and sugar (3 FFQ items) on the energy densities of coffees and teas, calories from beverage condiments were evenly distributed into the total coffee and tea consumed by each participant. Low-calorie coffees and teas were defined by an energy density[29] below 0.165 kcal/g (approximately \u226459 kcal in a 12 fluid ounce serving), which corresponds to a natural breakpoint in the distribution of energy-densities of beverages consumed by free-living adult women, and below which fall artificially-sweetened beverages and coffee and tea drinks with minimal beverage condiments.[19] Beverage intakes were aggregated into three non-overlapping categories: 1) energy-dense, non-alcoholic beverages (Kool-aid, Hi-C, or other drinks with added vitamin C; Snapple, Calistoga, or sweetened bottled waters or iced teas; regular cola soft drinks; coffee or tea consumed with condiments that collectively yield an energy density \u22650.165 kcal/g), 2) low-calorie teas, coffees, and diet cola (caffeinated coffee; green tea; black, English or Chinese tea; Chinese herbs made into or added to soup or tea; diet cola soft drinks), and 3) alcohol (ethanol contained in beer, wine or wine coolers, and liquor or mixed drinks). Though alcoholic beverages are also energy-dense, they were categorized separately in order to maintain orthogonality in the exposure variables and due to the fact that ethanol is the putative mechanism linking alcoholic beverage consumption to cardiometabolic risk.**\n\n**Covariates\u2014Covariates were chosen based on their potential to confound hypothesized associations. Study site (Pittsburgh, PA, Chicago, IL, Oakland, CA, Los Angeles, CA, Boston, MA, and Detroit-area Michigan), ethnicity/race (African-American, Chinese, Japanese, non-Hispanic white), age (derived from birth date), and education (less than high school, high school, some college, bachelor's degree, graduate degree) were documented at baseline. Time-varying covariates, assessed at annual follow-up visits, included smoking status (no/yes), the presence of clinically-elevated depressive symptoms based on a Center for Epidemiological Studies -Depression Scale[30] score \u226516, annual income ($0 - $19,999; $20,000 - $34,999; $35,000 -$49,999; $50,000 - $74,999; $75,000 - $99,999; $100,000 and higher), menopausal status (premenopausal or early perimenopausal vs. late perimenopausal or postmenopausal) determined through menstrual bleeding patterns,[31] current hormone therapy use (yes/no), and non-occupational physical activity (assessed on 5-point Likert and ordinal quantitative scales with total scores ranging from 3-15 and higher values indicating more frequent engagement in physical activity).[32,33]**\n\n**Statistical Analyses**\n\nAll analyses were performed in Stata 13.1 (StataCorp, LLC, College Station, TX; 2013)[34] with a type I error rate of \u03b1=.05. At each dietary assessment (baseline, visit 5, visit 9), intakes of energy-dense and low-calorie beverage categories were quantified in units of 355 ml servings (approximately 12 fluid ounces) per day, and alcohol was quantified in units of grams per day. To adjust for total energy intake, each beverage intake variable was regressed on total energy intake in the concurrent year, and the residuals from these models were used as measures of dietary exposure to beverage intake in that year.[35,36] To summarize total dietary exposure from baseline to visit 9, the area under the curve (AUC) of residualized beverage intake across time was calculated based on the trapezoidal method using Stata's \u201cpkcollapse\u201d command. This approach has the advantages of accommodating different numbers of complete follow-up assessments, and different time intervals between assessments. AUC was calculated based on either all three FFQ assessments (n=1,281), or the baseline and visit 9 assessments only (n=167). Finally, as total follow-up duration from baseline to visit 9 varied across participants, beverage intake values were converted to a rate by dividing the AUC values by each participant's follow-up duration. The resulting values corresponded to rates of energy-adjusted usual beverage intake, interpolated across 9 years of follow-up, in units of 355 ml servings/d for energy-dense and low-calorie beverage and g/d for alcohol. As analyses also adjusted for total energy intake across follow-up, the AUC for total energy intake was calculated in the same fashion. To assess the potential relevance of beverage intake to disparities in cardiometabolic risk, ANOVA was used to test for differences in beverage and total energy intake by ethnicity/race and education.\n\nThree sets of analyses tested associations between beverage intake and cardiometabolic risk markers, which were assessed at baseline and visits 1, 3, 5, 7, and 12. Generalized linear mixed models were used to model the odds of incident metabolic syndrome from the beverage intake by time interaction, along with the lower-order terms for beverage intake (AUC) and time, and covariates. Interaction terms involving the beverage intake variables and ethnicity/race, education, and physical activity levels were separately entered into the above models to test effect modification. Sensitivity analyses testing for non-linear (quadratic) associations involving alcohol intake were conducted. Finally, for any beverage intake category associated with incident metabolic syndrome, additional analyses tested associations with odds of developing each of the five individual metabolic syndrome components. These analyses included only women who did not meet criteria at baseline for the specific metabolic syndrome component being modeled. All models utilized mean and variance adaptive Gauss\u2013Hermite quadrature integration with seven integration points, and included random effects for both intercepts and the effect of time (in years since baseline). Results from both crude and adjusted models are reported.",
        "results": "### Results\n\nAnalyses included 8,256 observations from 1,448 women without metabolic syndrome at baseline. Sample characteristics are summarized in Table 1. The average follow-up duration in the analytic sample was 13.33 (SD=1.69) years. By the end of the follow-up period, 18.5% of women had developed metabolic syndrome. The average number of metabolic syndrome diagnostic components increased (p<.001) from baseline (mean=0.78, SD=0.79; median=1, interquartile range: 0-1) to visit 12 (mean=1.38, SD=1.23, median=1, interquartile range: 0-2).\n\nBaseline beverage intakes are shown in Table 2. Overall, women consumed almost twice as much low-calorie tea, coffee, and diet cola than they did energy-dense, non-alcoholic beverages. However, significant ethnic/racial differences in baseline beverage intakes were present (Table 2). African-American women consumed substantially more energy-dense, non-alcoholic beverages, and less low-calorie beverages and alcohol, than the reference category of non-Hispanic white women. Chinese women had significantly lower intakes of all three beverage categories than non-Hispanic white women. Japanese women consumed more low-calorie beverages and less alcohol than non-Hispanic white women. Associations between education and beverage intake are also shown in Table 2. Relative to the reference category of women with some college education, women with a 4-year college or graduate degree had lower baseline intake of energy-dense, non-alcoholic beverages, and women with graduate degrees had higher intake of alcohol. At baseline, physical activity exhibited a weak negative association with energy-dense beverage intake (r=-.05, p=.04) and a weak positive association with alcohol intake (r=.08, p<.01), but was unrelated to intake of low-calorie beverages (r=.03, p=.21).\n\nIn correlation analyses with the energy-adjusted beverage intake variables (data not shown), usual intake of energy-dense non-alcoholic beverages was inversely associated with intake of low-calorie teas, coffee, and diet cola (r=-.23, p<.0001), and weakly associated with alcohol intake (r=-.08, p<.01). Intake of low-calorie teas, coffee, and diet cola was positively correlated with alcohol intake (r=.09, p<.001). Total energy intake was positively correlated with energy-dense non-alcoholic beverage intake (r=.24, p<.0001) and alcohol intake (r=.11, p<.0001), but was not correlated with low-calorie beverage intake (r=.05, p=.08).\n\nOdds of incident metabolic syndrome increased across the duration of follow-up (Table 3). However, as reflected in a significant time by beverage intake interaction term, the odds of developing metabolic syndrome increased by 1.05 per year with each additional 355 ml of energy-dense beverages typically consumed (Table 3). Temporal changes in the odds of developing metabolic syndrome were unrelated to intakes of low-calorie beverages and alcohol (Table 3). This was also true when examining a quadratic term for alcohol intake (OR=1.00, 95% CI: 0.99, 1.01). There was no evidence of effect modification by ethnicity/race, education, and physical activity, as the three-way interactions of these variables with time and beverage intake were non-significant in all models (results not shown). No associations were observed in exploratory analyses predicting cardiometabolic outcomes solely from intake of diet cola (results not shown).\n\nSecondary analyses tested whether intake of energy-dense, non-alcoholic beverages was associated with odds of developing each of the five metabolic syndrome components among women without these risk factors at baseline (Table 4). In both unadjusted and adjusted models, odds of developing abdominal obesity and hypertension increased at a faster rate across follow-up in women with higher intakes of energy-dense, non-alcoholic beverages. Beverage intake was unrelated to time-related changes in the odds of developing hypertriglyceridemia, low HDL, or impaired fasting glucose (Table 4).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "4"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Beverage Intake and Metabolic Syndrome Risk Over 14 Years: The Study of Women's Health Across the Nation.",
        "abstract": "## BACKGROUND\nAlcohol and energy-dense beverages consumption have been implicated in cardiometabolic disease, albeit inconsistently.\n## OBJECTIVE\nThis study tested prospective associations between intakes of alcohol, energy-dense beverages, and low-calorie beverages and cardiometabolic risk in midlife women.\n## DESIGN\nThe Study of Women's Health Across the Nation is a 14-year, multisite prospective cohort study (1996-2011). Beverage intake and cardiometabolic risk factors that define the metabolic syndrome (hypertension, abdominal obesity, impaired fasting glucose, low high-density lipoprotein cholesterol level, and hypertriglyceridemia) were assessed throughout follow-up.\n## PARTICIPANTS/SETTING\nParticipants (N=1,448) were African American, Chinese, Japanese, and non-Hispanic white midlife women from six US cities.\n## MAIN OUTCOME MEASURES\nThe primary outcomes were incident metabolic syndrome and the individual metabolic syndrome components.\n## STATISTICAL ANALYSES PERFORMED\nGeneralized linear mixed models tested associations between intakes within each beverage category and odds of meeting criteria for metabolic syndrome and each of the metabolic syndrome components.\n## RESULTS\nEnergy-dense beverage consumption was highest among African-American women and lowest among women with college degrees. Non-Hispanic white women consumed the largest quantities of alcohol. Independent of energy intake and potential confounders, each additional 355 mL energy-dense beverages consumed per day was associated with higher odds of developing metabolic syndrome in each successive year of follow-up (odds ratio [OR] 1.05, 95% CI 1.02 to 1.08). Greater energy-dense beverage intake was associated with more rapidly increasing odds of developing hypertension (OR 1.06, 95% CI 1.02 to 1.11) and abdominal obesity (OR 1.10, 95% CI 1.03 to 1.16) over time, but not with the other metabolic syndrome components. Intakes of alcohol and low-calorie coffees, teas, and diet cola were not associated with metabolic syndrome risk.\n## CONCLUSIONS\nOver 14 years of follow-up, energy-dense nonalcoholic beverage consumption was associated with incident metabolic syndrome in midlife women. The observed differences in intakes by ethnicity/race and education suggest that consumption of these beverages may contribute to disparities in risk factors for diabetes and cardiovascular disease.",
        "methods": "### Methods\n\n**Participants**\n\nParticipants were midlife women from seven U.S. regions who were enrolled in the Study of Women's Health Across the Nation (SWAN), a multiethnic/racial, longitudinal study of aging and health determinants across the menopausal transition. SWAN's methodology is detailed elsewhere.[25] Procedures were approved by each site's institutional review board, and all participants provided written informed consent.\n\nWomen were recruited from five ethnic/racial groups: African-American (Boston, Chicago, Detroit area, and Pittsburgh), Chinese (Oakland area), Hispanic (Newark), Japanese (Los Angeles), and non-Hispanic white (all sites). Eligibility criteria included being aged 42-52 years at enrollment (1996-1997), having an intact uterus and at least one ovary, not being pregnant or lactating, not using oral contraceptives or reproductive hormone therapy, and reporting at least one menstrual cycle in the prior three months. This report includes data collected at the baseline assessment (1996-1997) and annual follow-up assessments that occurred through 2011. Assessments occurred either in clinic settings or participants' homes. However, participants from the Newark site were excluded from this analysis due to high attrition and lack of dietary data after visit 5. Of 2,870 enrolled women across the other 6 sites, 1,820 (63%) had at least nine years of follow-up data on both beverage intake and the metabolic syndrome components. Women who met criteria for the metabolic syndrome at baseline (n=372; 20%) were excluded as this analysis focused on incident cases. The final analytic sample included 1,448 women.\n\n**Measures**\n\n**Waist circumference\u2014Waist circumference was measured to the nearest 0.1 cm with a measuring tape placed horizontally around the participant at the narrowest part of the torso. Measurements were taken over light clothing at the end of a normal exhalation.**\n\n**Blood pressure\u2014Blood pressure was measured on the right arm in a seated position after at least 5 minutes of rest using a mercury sphygmomanometer. Respondents had not smoked or consumed caffeine within the prior 30 minutes. Appropriate cuff size was determined based on arm circumference. Two sequential blood pressure values were completed at least two minutes apart and averaged.**\n\n**Blood assays\u2014Blood was drawn after a 12-hour fast. Samples were frozen and shipped to either Medical Research Laboratories (Lexington, KY; baseline to visit 7) or University of Michigan Pathology Laboratory (after visit 7), which are certified by the National Heart Lung and Blood Institute, Centers for Disease Control Part III program.[26] Triglycerides were analyzed by enzymatic methods on a Hitachi 747 analyzer (Boehringer Mannheim Diagnostics, Indianapolis, IN). HDL was isolated using heparin-2M manganese chloride. Glucose was determined from fasting serum samples using a hexokinase-coupled reaction (Boehringer Mannheim Diagnostics, Indianapolis, IN).**\n\n**Definition of metabolic syndrome components and status\u2014Consistent with current diagnostic criteria,[1] metabolic syndrome was defined by the presence of at least three of the following five components: abdominal obesity (waist circumference \u226580 cm for Chinese and Japanese women, and \u226588 cm for other ethnic/racial groups), hypertension (systolic \u2265130 mmHg, diastolic \u226585 mmHg, or taking blood pressure medication), low HDL (<50 mg/dl), hypertriglyceridemia (fasting triglycerides \u2265150 mg/dl), and impaired fasting glucose (fasting glucose>100 mg/dl, or taking medication for impaired glucose regulation). Metabolic syndrome status and the total number of metabolic syndrome diagnostic components were determined at each time point.**\n\n**Beverage and total calorie intake\u2014The SWAN Food Frequency Questionnaire [FFQ[27]], which is an adaptation of the 1995 version of the Block Food Frequency Questionnaire,[28] was completed at baseline, visit 5, and visit 9. The FFQ was interviewer-administered. Missing and double-marked items were not scored. The SWAN FFQ assesses the usual frequency and portion sizes consumed for 83 foods and 20 beverages over the past year or so. Due to the multiethnic/racial nature of SWAN, the FFQ was administered in English, Spanish, Chinese (Cantonese), and Japanese. The Cantonese and Japanese FFQ versions contained the same 103 core items as the English version, plus 12-16 additional ethnic foods specific to each group. To avoid under-reporting of ethnic foods, the Chinese and Japanese ethnic food items were administered to all ethnic/racial groups at visits 5 and 9. The FFQ yields an estimate of total daily energy intake. Additionally, typical intakes of 20 beverages and beverage condiments were estimated by multiplying the reported portion sizes typically consumed by the reported frequency of consumption. To account for the contribution of beverage condiments such as cream, milk, and sugar (3 FFQ items) on the energy densities of coffees and teas, calories from beverage condiments were evenly distributed into the total coffee and tea consumed by each participant. Low-calorie coffees and teas were defined by an energy density[29] below 0.165 kcal/g (approximately \u226459 kcal in a 12 fluid ounce serving), which corresponds to a natural breakpoint in the distribution of energy-densities of beverages consumed by free-living adult women, and below which fall artificially-sweetened beverages and coffee and tea drinks with minimal beverage condiments.[19] Beverage intakes were aggregated into three non-overlapping categories: 1) energy-dense, non-alcoholic beverages (Kool-aid, Hi-C, or other drinks with added vitamin C; Snapple, Calistoga, or sweetened bottled waters or iced teas; regular cola soft drinks; coffee or tea consumed with condiments that collectively yield an energy density \u22650.165 kcal/g), 2) low-calorie teas, coffees, and diet cola (caffeinated coffee; green tea; black, English or Chinese tea; Chinese herbs made into or added to soup or tea; diet cola soft drinks), and 3) alcohol (ethanol contained in beer, wine or wine coolers, and liquor or mixed drinks). Though alcoholic beverages are also energy-dense, they were categorized separately in order to maintain orthogonality in the exposure variables and due to the fact that ethanol is the putative mechanism linking alcoholic beverage consumption to cardiometabolic risk.**\n\n**Covariates\u2014Covariates were chosen based on their potential to confound hypothesized associations. Study site (Pittsburgh, PA, Chicago, IL, Oakland, CA, Los Angeles, CA, Boston, MA, and Detroit-area Michigan), ethnicity/race (African-American, Chinese, Japanese, non-Hispanic white), age (derived from birth date), and education (less than high school, high school, some college, bachelor's degree, graduate degree) were documented at baseline. Time-varying covariates, assessed at annual follow-up visits, included smoking status (no/yes), the presence of clinically-elevated depressive symptoms based on a Center for Epidemiological Studies -Depression Scale[30] score \u226516, annual income ($0 - $19,999; $20,000 - $34,999; $35,000 -$49,999; $50,000 - $74,999; $75,000 - $99,999; $100,000 and higher), menopausal status (premenopausal or early perimenopausal vs. late perimenopausal or postmenopausal) determined through menstrual bleeding patterns,[31] current hormone therapy use (yes/no), and non-occupational physical activity (assessed on 5-point Likert and ordinal quantitative scales with total scores ranging from 3-15 and higher values indicating more frequent engagement in physical activity).[32,33]**\n\n**Statistical Analyses**\n\nAll analyses were performed in Stata 13.1 (StataCorp, LLC, College Station, TX; 2013)[34] with a type I error rate of \u03b1=.05. At each dietary assessment (baseline, visit 5, visit 9), intakes of energy-dense and low-calorie beverage categories were quantified in units of 355 ml servings (approximately 12 fluid ounces) per day, and alcohol was quantified in units of grams per day. To adjust for total energy intake, each beverage intake variable was regressed on total energy intake in the concurrent year, and the residuals from these models were used as measures of dietary exposure to beverage intake in that year.[35,36] To summarize total dietary exposure from baseline to visit 9, the area under the curve (AUC) of residualized beverage intake across time was calculated based on the trapezoidal method using Stata's \u201cpkcollapse\u201d command. This approach has the advantages of accommodating different numbers of complete follow-up assessments, and different time intervals between assessments. AUC was calculated based on either all three FFQ assessments (n=1,281), or the baseline and visit 9 assessments only (n=167). Finally, as total follow-up duration from baseline to visit 9 varied across participants, beverage intake values were converted to a rate by dividing the AUC values by each participant's follow-up duration. The resulting values corresponded to rates of energy-adjusted usual beverage intake, interpolated across 9 years of follow-up, in units of 355 ml servings/d for energy-dense and low-calorie beverage and g/d for alcohol. As analyses also adjusted for total energy intake across follow-up, the AUC for total energy intake was calculated in the same fashion. To assess the potential relevance of beverage intake to disparities in cardiometabolic risk, ANOVA was used to test for differences in beverage and total energy intake by ethnicity/race and education.\n\nThree sets of analyses tested associations between beverage intake and cardiometabolic risk markers, which were assessed at baseline and visits 1, 3, 5, 7, and 12. Generalized linear mixed models were used to model the odds of incident metabolic syndrome from the beverage intake by time interaction, along with the lower-order terms for beverage intake (AUC) and time, and covariates. Interaction terms involving the beverage intake variables and ethnicity/race, education, and physical activity levels were separately entered into the above models to test effect modification. Sensitivity analyses testing for non-linear (quadratic) associations involving alcohol intake were conducted. Finally, for any beverage intake category associated with incident metabolic syndrome, additional analyses tested associations with odds of developing each of the five individual metabolic syndrome components. These analyses included only women who did not meet criteria at baseline for the specific metabolic syndrome component being modeled. All models utilized mean and variance adaptive Gauss\u2013Hermite quadrature integration with seven integration points, and included random effects for both intercepts and the effect of time (in years since baseline). Results from both crude and adjusted models are reported.",
        "results": "### Results\n\nAnalyses included 8,256 observations from 1,448 women without metabolic syndrome at baseline. Sample characteristics are summarized in Table 1. The average follow-up duration in the analytic sample was 13.33 (SD=1.69) years. By the end of the follow-up period, 18.5% of women had developed metabolic syndrome. The average number of metabolic syndrome diagnostic components increased (p<.001) from baseline (mean=0.78, SD=0.79; median=1, interquartile range: 0-1) to visit 12 (mean=1.38, SD=1.23, median=1, interquartile range: 0-2).\n\nBaseline beverage intakes are shown in Table 2. Overall, women consumed almost twice as much low-calorie tea, coffee, and diet cola than they did energy-dense, non-alcoholic beverages. However, significant ethnic/racial differences in baseline beverage intakes were present (Table 2). African-American women consumed substantially more energy-dense, non-alcoholic beverages, and less low-calorie beverages and alcohol, than the reference category of non-Hispanic white women. Chinese women had significantly lower intakes of all three beverage categories than non-Hispanic white women. Japanese women consumed more low-calorie beverages and less alcohol than non-Hispanic white women. Associations between education and beverage intake are also shown in Table 2. Relative to the reference category of women with some college education, women with a 4-year college or graduate degree had lower baseline intake of energy-dense, non-alcoholic beverages, and women with graduate degrees had higher intake of alcohol. At baseline, physical activity exhibited a weak negative association with energy-dense beverage intake (r=-.05, p=.04) and a weak positive association with alcohol intake (r=.08, p<.01), but was unrelated to intake of low-calorie beverages (r=.03, p=.21).\n\nIn correlation analyses with the energy-adjusted beverage intake variables (data not shown), usual intake of energy-dense non-alcoholic beverages was inversely associated with intake of low-calorie teas, coffee, and diet cola (r=-.23, p<.0001), and weakly associated with alcohol intake (r=-.08, p<.01). Intake of low-calorie teas, coffee, and diet cola was positively correlated with alcohol intake (r=.09, p<.001). Total energy intake was positively correlated with energy-dense non-alcoholic beverage intake (r=.24, p<.0001) and alcohol intake (r=.11, p<.0001), but was not correlated with low-calorie beverage intake (r=.05, p=.08).\n\nOdds of incident metabolic syndrome increased across the duration of follow-up (Table 3). However, as reflected in a significant time by beverage intake interaction term, the odds of developing metabolic syndrome increased by 1.05 per year with each additional 355 ml of energy-dense beverages typically consumed (Table 3). Temporal changes in the odds of developing metabolic syndrome were unrelated to intakes of low-calorie beverages and alcohol (Table 3). This was also true when examining a quadratic term for alcohol intake (OR=1.00, 95% CI: 0.99, 1.01). There was no evidence of effect modification by ethnicity/race, education, and physical activity, as the three-way interactions of these variables with time and beverage intake were non-significant in all models (results not shown). No associations were observed in exploratory analyses predicting cardiometabolic outcomes solely from intake of diet cola (results not shown).\n\nSecondary analyses tested whether intake of energy-dense, non-alcoholic beverages was associated with odds of developing each of the five metabolic syndrome components among women without these risk factors at baseline (Table 4). In both unadjusted and adjusted models, odds of developing abdominal obesity and hypertension increased at a faster rate across follow-up in women with higher intakes of energy-dense, non-alcoholic beverages. Beverage intake was unrelated to time-related changes in the odds of developing hypertriglyceridemia, low HDL, or impaired fasting glucose (Table 4).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Beverage Intake and Metabolic Syndrome Risk Over 14 Years: The Study of Women's Health Across the Nation.",
        "abstract": "## BACKGROUND\nAlcohol and energy-dense beverages consumption have been implicated in cardiometabolic disease, albeit inconsistently.\n## OBJECTIVE\nThis study tested prospective associations between intakes of alcohol, energy-dense beverages, and low-calorie beverages and cardiometabolic risk in midlife women.\n## DESIGN\nThe Study of Women's Health Across the Nation is a 14-year, multisite prospective cohort study (1996-2011). Beverage intake and cardiometabolic risk factors that define the metabolic syndrome (hypertension, abdominal obesity, impaired fasting glucose, low high-density lipoprotein cholesterol level, and hypertriglyceridemia) were assessed throughout follow-up.\n## PARTICIPANTS/SETTING\nParticipants (N=1,448) were African American, Chinese, Japanese, and non-Hispanic white midlife women from six US cities.\n## MAIN OUTCOME MEASURES\nThe primary outcomes were incident metabolic syndrome and the individual metabolic syndrome components.\n## STATISTICAL ANALYSES PERFORMED\nGeneralized linear mixed models tested associations between intakes within each beverage category and odds of meeting criteria for metabolic syndrome and each of the metabolic syndrome components.\n## RESULTS\nEnergy-dense beverage consumption was highest among African-American women and lowest among women with college degrees. Non-Hispanic white women consumed the largest quantities of alcohol. Independent of energy intake and potential confounders, each additional 355 mL energy-dense beverages consumed per day was associated with higher odds of developing metabolic syndrome in each successive year of follow-up (odds ratio [OR] 1.05, 95% CI 1.02 to 1.08). Greater energy-dense beverage intake was associated with more rapidly increasing odds of developing hypertension (OR 1.06, 95% CI 1.02 to 1.11) and abdominal obesity (OR 1.10, 95% CI 1.03 to 1.16) over time, but not with the other metabolic syndrome components. Intakes of alcohol and low-calorie coffees, teas, and diet cola were not associated with metabolic syndrome risk.\n## CONCLUSIONS\nOver 14 years of follow-up, energy-dense nonalcoholic beverage consumption was associated with incident metabolic syndrome in midlife women. The observed differences in intakes by ethnicity/race and education suggest that consumption of these beverages may contribute to disparities in risk factors for diabetes and cardiovascular disease.",
        "methods": "### Methods\n\n**Participants**\n\nParticipants were midlife women from seven U.S. regions who were enrolled in the Study of Women's Health Across the Nation (SWAN), a multiethnic/racial, longitudinal study of aging and health determinants across the menopausal transition. SWAN's methodology is detailed elsewhere.[25] Procedures were approved by each site's institutional review board, and all participants provided written informed consent.\n\nWomen were recruited from five ethnic/racial groups: African-American (Boston, Chicago, Detroit area, and Pittsburgh), Chinese (Oakland area), Hispanic (Newark), Japanese (Los Angeles), and non-Hispanic white (all sites). Eligibility criteria included being aged 42-52 years at enrollment (1996-1997), having an intact uterus and at least one ovary, not being pregnant or lactating, not using oral contraceptives or reproductive hormone therapy, and reporting at least one menstrual cycle in the prior three months. This report includes data collected at the baseline assessment (1996-1997) and annual follow-up assessments that occurred through 2011. Assessments occurred either in clinic settings or participants' homes. However, participants from the Newark site were excluded from this analysis due to high attrition and lack of dietary data after visit 5. Of 2,870 enrolled women across the other 6 sites, 1,820 (63%) had at least nine years of follow-up data on both beverage intake and the metabolic syndrome components. Women who met criteria for the metabolic syndrome at baseline (n=372; 20%) were excluded as this analysis focused on incident cases. The final analytic sample included 1,448 women.\n\n**Measures**\n\n**Waist circumference\u2014Waist circumference was measured to the nearest 0.1 cm with a measuring tape placed horizontally around the participant at the narrowest part of the torso. Measurements were taken over light clothing at the end of a normal exhalation.**\n\n**Blood pressure\u2014Blood pressure was measured on the right arm in a seated position after at least 5 minutes of rest using a mercury sphygmomanometer. Respondents had not smoked or consumed caffeine within the prior 30 minutes. Appropriate cuff size was determined based on arm circumference. Two sequential blood pressure values were completed at least two minutes apart and averaged.**\n\n**Blood assays\u2014Blood was drawn after a 12-hour fast. Samples were frozen and shipped to either Medical Research Laboratories (Lexington, KY; baseline to visit 7) or University of Michigan Pathology Laboratory (after visit 7), which are certified by the National Heart Lung and Blood Institute, Centers for Disease Control Part III program.[26] Triglycerides were analyzed by enzymatic methods on a Hitachi 747 analyzer (Boehringer Mannheim Diagnostics, Indianapolis, IN). HDL was isolated using heparin-2M manganese chloride. Glucose was determined from fasting serum samples using a hexokinase-coupled reaction (Boehringer Mannheim Diagnostics, Indianapolis, IN).**\n\n**Definition of metabolic syndrome components and status\u2014Consistent with current diagnostic criteria,[1] metabolic syndrome was defined by the presence of at least three of the following five components: abdominal obesity (waist circumference \u226580 cm for Chinese and Japanese women, and \u226588 cm for other ethnic/racial groups), hypertension (systolic \u2265130 mmHg, diastolic \u226585 mmHg, or taking blood pressure medication), low HDL (<50 mg/dl), hypertriglyceridemia (fasting triglycerides \u2265150 mg/dl), and impaired fasting glucose (fasting glucose>100 mg/dl, or taking medication for impaired glucose regulation). Metabolic syndrome status and the total number of metabolic syndrome diagnostic components were determined at each time point.**\n\n**Beverage and total calorie intake\u2014The SWAN Food Frequency Questionnaire [FFQ[27]], which is an adaptation of the 1995 version of the Block Food Frequency Questionnaire,[28] was completed at baseline, visit 5, and visit 9. The FFQ was interviewer-administered. Missing and double-marked items were not scored. The SWAN FFQ assesses the usual frequency and portion sizes consumed for 83 foods and 20 beverages over the past year or so. Due to the multiethnic/racial nature of SWAN, the FFQ was administered in English, Spanish, Chinese (Cantonese), and Japanese. The Cantonese and Japanese FFQ versions contained the same 103 core items as the English version, plus 12-16 additional ethnic foods specific to each group. To avoid under-reporting of ethnic foods, the Chinese and Japanese ethnic food items were administered to all ethnic/racial groups at visits 5 and 9. The FFQ yields an estimate of total daily energy intake. Additionally, typical intakes of 20 beverages and beverage condiments were estimated by multiplying the reported portion sizes typically consumed by the reported frequency of consumption. To account for the contribution of beverage condiments such as cream, milk, and sugar (3 FFQ items) on the energy densities of coffees and teas, calories from beverage condiments were evenly distributed into the total coffee and tea consumed by each participant. Low-calorie coffees and teas were defined by an energy density[29] below 0.165 kcal/g (approximately \u226459 kcal in a 12 fluid ounce serving), which corresponds to a natural breakpoint in the distribution of energy-densities of beverages consumed by free-living adult women, and below which fall artificially-sweetened beverages and coffee and tea drinks with minimal beverage condiments.[19] Beverage intakes were aggregated into three non-overlapping categories: 1) energy-dense, non-alcoholic beverages (Kool-aid, Hi-C, or other drinks with added vitamin C; Snapple, Calistoga, or sweetened bottled waters or iced teas; regular cola soft drinks; coffee or tea consumed with condiments that collectively yield an energy density \u22650.165 kcal/g), 2) low-calorie teas, coffees, and diet cola (caffeinated coffee; green tea; black, English or Chinese tea; Chinese herbs made into or added to soup or tea; diet cola soft drinks), and 3) alcohol (ethanol contained in beer, wine or wine coolers, and liquor or mixed drinks). Though alcoholic beverages are also energy-dense, they were categorized separately in order to maintain orthogonality in the exposure variables and due to the fact that ethanol is the putative mechanism linking alcoholic beverage consumption to cardiometabolic risk.**\n\n**Covariates\u2014Covariates were chosen based on their potential to confound hypothesized associations. Study site (Pittsburgh, PA, Chicago, IL, Oakland, CA, Los Angeles, CA, Boston, MA, and Detroit-area Michigan), ethnicity/race (African-American, Chinese, Japanese, non-Hispanic white), age (derived from birth date), and education (less than high school, high school, some college, bachelor's degree, graduate degree) were documented at baseline. Time-varying covariates, assessed at annual follow-up visits, included smoking status (no/yes), the presence of clinically-elevated depressive symptoms based on a Center for Epidemiological Studies -Depression Scale[30] score \u226516, annual income ($0 - $19,999; $20,000 - $34,999; $35,000 -$49,999; $50,000 - $74,999; $75,000 - $99,999; $100,000 and higher), menopausal status (premenopausal or early perimenopausal vs. late perimenopausal or postmenopausal) determined through menstrual bleeding patterns,[31] current hormone therapy use (yes/no), and non-occupational physical activity (assessed on 5-point Likert and ordinal quantitative scales with total scores ranging from 3-15 and higher values indicating more frequent engagement in physical activity).[32,33]**\n\n**Statistical Analyses**\n\nAll analyses were performed in Stata 13.1 (StataCorp, LLC, College Station, TX; 2013)[34] with a type I error rate of \u03b1=.05. At each dietary assessment (baseline, visit 5, visit 9), intakes of energy-dense and low-calorie beverage categories were quantified in units of 355 ml servings (approximately 12 fluid ounces) per day, and alcohol was quantified in units of grams per day. To adjust for total energy intake, each beverage intake variable was regressed on total energy intake in the concurrent year, and the residuals from these models were used as measures of dietary exposure to beverage intake in that year.[35,36] To summarize total dietary exposure from baseline to visit 9, the area under the curve (AUC) of residualized beverage intake across time was calculated based on the trapezoidal method using Stata's \u201cpkcollapse\u201d command. This approach has the advantages of accommodating different numbers of complete follow-up assessments, and different time intervals between assessments. AUC was calculated based on either all three FFQ assessments (n=1,281), or the baseline and visit 9 assessments only (n=167). Finally, as total follow-up duration from baseline to visit 9 varied across participants, beverage intake values were converted to a rate by dividing the AUC values by each participant's follow-up duration. The resulting values corresponded to rates of energy-adjusted usual beverage intake, interpolated across 9 years of follow-up, in units of 355 ml servings/d for energy-dense and low-calorie beverage and g/d for alcohol. As analyses also adjusted for total energy intake across follow-up, the AUC for total energy intake was calculated in the same fashion. To assess the potential relevance of beverage intake to disparities in cardiometabolic risk, ANOVA was used to test for differences in beverage and total energy intake by ethnicity/race and education.\n\nThree sets of analyses tested associations between beverage intake and cardiometabolic risk markers, which were assessed at baseline and visits 1, 3, 5, 7, and 12. Generalized linear mixed models were used to model the odds of incident metabolic syndrome from the beverage intake by time interaction, along with the lower-order terms for beverage intake (AUC) and time, and covariates. Interaction terms involving the beverage intake variables and ethnicity/race, education, and physical activity levels were separately entered into the above models to test effect modification. Sensitivity analyses testing for non-linear (quadratic) associations involving alcohol intake were conducted. Finally, for any beverage intake category associated with incident metabolic syndrome, additional analyses tested associations with odds of developing each of the five individual metabolic syndrome components. These analyses included only women who did not meet criteria at baseline for the specific metabolic syndrome component being modeled. All models utilized mean and variance adaptive Gauss\u2013Hermite quadrature integration with seven integration points, and included random effects for both intercepts and the effect of time (in years since baseline). Results from both crude and adjusted models are reported.",
        "results": "### Results\n\nAnalyses included 8,256 observations from 1,448 women without metabolic syndrome at baseline. Sample characteristics are summarized in Table 1. The average follow-up duration in the analytic sample was 13.33 (SD=1.69) years. By the end of the follow-up period, 18.5% of women had developed metabolic syndrome. The average number of metabolic syndrome diagnostic components increased (p<.001) from baseline (mean=0.78, SD=0.79; median=1, interquartile range: 0-1) to visit 12 (mean=1.38, SD=1.23, median=1, interquartile range: 0-2).\n\nBaseline beverage intakes are shown in Table 2. Overall, women consumed almost twice as much low-calorie tea, coffee, and diet cola than they did energy-dense, non-alcoholic beverages. However, significant ethnic/racial differences in baseline beverage intakes were present (Table 2). African-American women consumed substantially more energy-dense, non-alcoholic beverages, and less low-calorie beverages and alcohol, than the reference category of non-Hispanic white women. Chinese women had significantly lower intakes of all three beverage categories than non-Hispanic white women. Japanese women consumed more low-calorie beverages and less alcohol than non-Hispanic white women. Associations between education and beverage intake are also shown in Table 2. Relative to the reference category of women with some college education, women with a 4-year college or graduate degree had lower baseline intake of energy-dense, non-alcoholic beverages, and women with graduate degrees had higher intake of alcohol. At baseline, physical activity exhibited a weak negative association with energy-dense beverage intake (r=-.05, p=.04) and a weak positive association with alcohol intake (r=.08, p<.01), but was unrelated to intake of low-calorie beverages (r=.03, p=.21).\n\nIn correlation analyses with the energy-adjusted beverage intake variables (data not shown), usual intake of energy-dense non-alcoholic beverages was inversely associated with intake of low-calorie teas, coffee, and diet cola (r=-.23, p<.0001), and weakly associated with alcohol intake (r=-.08, p<.01). Intake of low-calorie teas, coffee, and diet cola was positively correlated with alcohol intake (r=.09, p<.001). Total energy intake was positively correlated with energy-dense non-alcoholic beverage intake (r=.24, p<.0001) and alcohol intake (r=.11, p<.0001), but was not correlated with low-calorie beverage intake (r=.05, p=.08).\n\nOdds of incident metabolic syndrome increased across the duration of follow-up (Table 3). However, as reflected in a significant time by beverage intake interaction term, the odds of developing metabolic syndrome increased by 1.05 per year with each additional 355 ml of energy-dense beverages typically consumed (Table 3). Temporal changes in the odds of developing metabolic syndrome were unrelated to intakes of low-calorie beverages and alcohol (Table 3). This was also true when examining a quadratic term for alcohol intake (OR=1.00, 95% CI: 0.99, 1.01). There was no evidence of effect modification by ethnicity/race, education, and physical activity, as the three-way interactions of these variables with time and beverage intake were non-significant in all models (results not shown). No associations were observed in exploratory analyses predicting cardiometabolic outcomes solely from intake of diet cola (results not shown).\n\nSecondary analyses tested whether intake of energy-dense, non-alcoholic beverages was associated with odds of developing each of the five metabolic syndrome components among women without these risk factors at baseline (Table 4). In both unadjusted and adjusted models, odds of developing abdominal obesity and hypertension increased at a faster rate across follow-up in women with higher intakes of energy-dense, non-alcoholic beverages. Beverage intake was unrelated to time-related changes in the odds of developing hypertriglyceridemia, low HDL, or impaired fasting glucose (Table 4).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Consumption of Yogurt, Low-Fat Milk, and Other Low-Fat Dairy Products Is Associated with Lower Risk of Metabolic Syndrome Incidence in an Elderly Mediterranean Population.",
        "abstract": "## BACKGROUND\nThe association between consumption of dairy products and the risk of developing metabolic syndrome (MetS) is unclear.\n## OBJECTIVE\nThe purpose of this study was to evaluate the associations between consumption of dairy products (total and different subtypes) and incident MetS in a Mediterranean population at high cardiovascular disease risk.\n## METHODS\nWe prospectively analyzed 1868 men and women (55-80 y old) without MetS at baseline, recruited from different PREDIMED (Prevenci\u00f3n con Dieta Mediterr\u00e1nea) centers between October 2003 and June 2009 and followed up until December 2010. MetS was defined according to updated, harmonized criteria. At baseline and yearly thereafter, we determined anthropometric variables, dietary habits by a 137-item validated food-frequency questionnaire, and blood biochemistry. Multivariable-adjusted HRs of MetS or its components were estimated for each of the 2 upper tertiles (vs. the lowest one) of mean consumption of dairy products during the follow-up.\n## RESULTS\nDuring a median follow-up of 3.2 y, we documented 930 incident MetS cases. In the multivariable-adjusted model, HRs (95% CIs) of MetS for the comparison of extreme tertiles of dairy product consumption were 0.72 (0.61, 0.86) for low-fat dairy, 0.73 (0.62, 0.86) for low-fat yogurt, 0.78 (0.66, 0.92) for whole-fat yogurt, and 0.80 (0.67, 0.95) for low-fat milk. The respective HR for cheese was 1.31 (1.10, 1.56).\n## CONCLUSIONS\nHigher consumption of low-fat dairy products, yogurt (total, low-fat, and whole-fat yogurt) and low-fat milk was associated with a reduced risk of MetS in individuals at high cardiovascular disease risk from a Mediterranean population. Conversely, higher consumption of cheese was related to a higher risk of MetS. This trial was registered at controlled-trials.com as ISRCTN35739639.",
        "methods": "### Methods\n\n##### Study design, participants, and outcome. The present study was conducted within the framework of the PREDIMED trial, the design of which has been described in detail elsewhere (23, 24). The PREDIMED study is a large, parallel-group, multicenter, randomized, controlled field trial aimed at assessing the effects of the Mediterranean diet on the primary prevention of CVD (25). The main results of the trial with respect to the primary endpoint have been published recently (26). Briefly, 7447 community-dwelling men (aged 55\u201380 y) and women (aged 60\u201380 y) with no previously documented CVD were recruited. They were eligible if they had either T2D or \u22653 of the following CVD risk factors: hypertension (systolic blood pressure \u2265140 mm Hg or diastolic blood pressure \u226590 mm Hg or on antihypertensive medication), high plasma LDL cholesterol (\u2265160 mg/dL), low plasma HDL cholesterol (<40 mg/dL in men; <50 mg/dL in women), overweight or obesity (BMI \u226525 kg/m\u00b2), current smoking, or a family history of premature coronary heart disease (23). From October 2003 to June 2009, participants were randomly assigned to 3 intervention groups (2 of which were advised to follow a Mediterranean diet supplemented with either 1 L/wk extra-virgin olive oil or 30 g/d mixed nuts, and were compared with the third group, which was advised to follow a control low-fat diet). The study follow-up ended in December 2010. All participants provided their informed consent and the protocol was approved by the institutional review boards of each recruitment center. In the present report, the data were analyzed assuming the design of an observational prospective cohort whose members were selected from all the PREDIMED recruiting centers with biochemical determinations available for a follow-up of \u22652 y (n = 5801). Because our aim was to explore the associations between the consumption of dairy foods and incident MetS, we excluded participants who had diagnoses of MetS at baseline (63.9%; n = 3707). We also excluded participants who had not completed the baseline FFQ or who reported an extreme total energy intake with values outside the prespecified limits (500\u20133500 kcal/d in women and 800\u20134000 kcal/d in men). A total of 2094 individuals were assessed. Of these, a total of 226 were excluded because of missing data that prevented the presence of MetS incidence from being determined. Thus, a total of 1868 participants were included in our longitudinal assessment for MetS incidence. The individual components of MetS\u2014abdominal obesity, hypertriglyceridemia, low HDL cholesterol, high blood pressure, and high fasting glucose concentration\u2014were analyzed for a total of 1386, 3539, 3745, 337, and 1844 participants (from 5801 participants initially considered with biochemical determinations available for a follow-up of \u22652 y), respectively. The primary endpoint of the PREDIMED trial was a combination of several major cardiovascular clinical events (myocardial infarction, stroke, or CVD death). In the present analysis, incident MetS and its components were considered as the outcome. MetS was defined in accordance with the updated harmonized criteria of the International Diabetes Federation and the AHA/National Heart, Lung, and Blood Institute (1). Individuals were diagnosed with MetS if they had \u22653 of the following components: hypertriglyceridemia concentrations of HDL cholesterol [<50 mg/dL (<1.3 mmol/L) and <40 mg/dL (<1.03 mmol/L) in women and men, respectively] or drug treatment for low HDL cholesterol; elevated blood pressure (systolic \u2265130 mmHg and/or diastolic \u226585 mmHg) or being treated for hypertension; high fasting plasma glucose [\u2265100 mg/dL (\u22655.5 mmol/L)] or drug treatment for hyperglycemia; and elevated waist circumference for European individuals (\u226588 cm in women and \u2265102 cm in men).\n\nDietary assessment. At baseline and yearly during follow-up, dietary intake was quantified by trained dietitians with a 137-item semi-quantitative FFQ validated for the PREDIMED study (27). In the validation study, the FFQ was administered twice to explore reproducibility at 1 y; 4 3-d dietary records for the different seasons of the year were used as gold standard. The reproducibility of the FFQ used in PREDIMED for food groups and energy and nutrient intake, explored by the Pearson correlation coefficient, ranged from 0.50 to 0.82, and the intraclass correlation coefficient (ICC) ranged from 0.63 to 0.90. The validity indexes of the FFQ in relation to the dietary records for food groups and energy and nutrient intake ranged from 0.24 to 0.72, whereas the range of the ICC was between 0.40 and 0.84. Regarding dairy product consumption, the reproducibility and validity of FFQs were 0.81 (ICC 0.89), and 0.72 (ICC 0.84), respectively. Dairy product consumption was assessed yearly with the use of 15 items from the FFQ. In order to assess habitual dietary intake over the previous year, frequencies of consumption were measured in 9 categories (ranging from never/almost never to >6 servings/d) for each food item. These responses to individual dairy items were then converted into mean daily consumption (grams per day) during the follow-up by multiplying the typical portion sizes (grams) by the consumption frequency for each food and making the appropriate division for the period assessed to obtain daily consumption. The total dairy foods category included low-fat/skim milk and skim yogurt, whole milk, condensed milk, whole yogurt, custard, and all types of cheeses, including petit Swiss, ricotta, cottage, and semicured/cured cheeses such as cheddar, Manchego, and Emmentaler). Low-fat dairy foods included low-fat/skim milk and skim yogurt, whole-fat dairy foods (whole milk and whole yogurt), and total dairy foods, including all of the above. Consumption of dairy products was also categorized by subtype, including milk (including total, low-fat and whole milk), yogurt (including total, low-fat and whole-fat yogurt), and cheese. Energy and nutrient intake and food groups were calculated from Spanish food composition tables (28, 29). We adjusted dairy consumption for total energy intake with the use of the nutrient residual method (30).\n\nMeasurements. Participants completed the following at baseline and yearly: 1) a questionnaire about lifestyle variables, medical history, and medication use; 2) a 14-item validated questionnaire (31) designed to assess adherence to the Mediterranean diet; 3) a validated 137-item semi-quantitative FFQ (27); and 4) the validated Spanish version (32) of the Minnesota Leisure-Time Physical Activity Questionnaire. In addition, anthropometric variables and blood pressure were determined by trained staff. Blood pressure was measured in triplicate with the use of a validated semiautomatic oscillometer with a 5 min interval between measurements and the subject in a sitting position (Omron HEM-705CP). Blood samples were collected after an overnight fast, coded, shipped to a central laboratory, and stored at -80\u00b0C until analysis. Biochemical analysis was performed in local laboratories. Glucose was measured by the glucose-oxidase method, cholesterol by esterase-oxidase-peroxidase, TGs by glycerol-phosphate oxidase-peroxidase, and HDL cholesterol by direct measurement. All local laboratories satisfied external quality-control requirements. When TGs were <300 mg/dL, LDL cholesterol was calculated with the use of the Friedewald formula so that the LDL cholesterol was not underestimated. A concordance study of 9 laboratories was conducted. From each study, a mean of 200 samples were analyzed for total cholesterol, HDL cholesterol, and TGs with the use of the Medical Research Institute of del Mar laboratory as reference. The Medical Research Institute of del Mar laboratory used ABX-Horiba commercial kits in a PENTRA-400 autoanalyzer (ABX-Horiba). One center was unable to provide samples for the concordance study. The analysis of concordance of lipid measurements showed an r\u00b2 and an ICC (95% CI) between 0.85 and 0.97, and 0.85 (0.77, 0.90) and 0.97 (0.81, 0.83) and 0.92 (0.89, 0.95) for HDL cholesterol, respectively; between 0.81 and 0.99, and 0.81 (0.73, 0.87) and 0.99 (0.99, 0.99) for triglycerides, respectively; and between 0.82 and 0.96, and 0.82 (0.74, 0.88) and 0.99 (0.99, 0.99) for glucose, respectively.\n\nStatistical analysis. We averaged the intake reported during the baseline interview and the yearly consumption during the follow-up. Then, participants were categorized into tertiles of the mean consumption of total dairy products and different subtypes during the follow-up. To better represent the long-term consumption of dairy products and to minimize within-person variation, we used the mean energy-adjusted dairy consumption for all analyses based on assessments of items from all FFQs, which were administered at baseline and yearly during the follow-up for those participants who did not develop MetS. For those who did develop MetS, and given that participants can alter their dietary pattern after developing MetS, we only used data from all the available FFQs until the year before MetS was diagnosed. The baseline characteristics of the participants are expressed as mean \u00b1 SD or median (IQR) for continuous variables and number and percentages for categorical variables. Chi-square and 1-factor ANOVA tests were used to assess differences in the baseline characteristics of the study population. Multivariable time-dependent Cox proportional regression models were fitted to assess the HRs of developing MetS and its components during follow-up according to tertiles of consumption of total, low-fat, and whole-fat dairy products; milk, low-fat milk, and whole milk; total yogurt, low-fat yogurt, and whole-fat yogurt; and cheese. Both upper tertiles were compared with the lowest tertile (reference). The assumption of proportional hazards was tested with the use of time-dependent covariates. The time variable was the interval between random assignment and the date of the last follow-up, or the last recorded clinical event (MetS incidence) of participants who were still alive, whichever occurred first. Participants who were free of MetS or who were lost during follow-up were censored at the date of the last visit. Three different Cox regression models were adjusted for potential confounding factors. Model 1 was adjusted for intervention group; sex; age (years); leisure time physical activity (metabolic equivalent task \u00d7 d); BMI (kilograms per meter squared); current smoker (yes/no); former smoker (yes/no); and hypoglycemic, hypolipidemic, antihypertensive, and insulin treatment at baseline. Model 2 was additionally adjusted for mean consumption during follow-up of vegetables, fruit, legumes, cereals, fish, red meat, cookies, olive oil, and nuts (all grams per day), as well as alcohol (grams per day and quadratic term). Model 3 was adjusted for model 2 plus the prevalence of MetS components at baseline, including abdominal obesity (yes/no), hypertriglyceridemia (yes/no), low HDL cholesterol (yes/no), hypertension (yes/no), and high fasting plasma glucose (yes/no). Statistical interaction between tertiles of total dairy consumption and its different subtypes, and potential effect-modifying variables, such as sex and intervention group, was assessed by including product terms in the models. To assess the linear trend, the median value of each tertile of dairy product and the different dairy subtypes was assigned and used as a continuous variable in the Cox regression models. The level of significance for all statistical tests was P < 0.05 for bilateral contrast. The Benjamini-Hochberg method was used to correct P values for multiple comparisons (33). We also conducted a sensitivity analysis of a number of MetS components at baseline to test the robustness of our results. The main analysis was stratified by those individuals who did not have any components of MetS or had one, or who had 2 components. This was done to prevent possible bias effects, because it is easier for those who already have 2 criteria to meet the diagnostic definition than those with one or none. All analyses were performed with SPSS software, version 19.0.",
        "results": "### Results\n\nAfter a median of 3.2 y of follow-up (IQR: 1.9, 5.8), 930 participants without MetS at baseline (53.8% women) developed new-onset MetS. Of those not showing the specific MetS components at baseline, 43.4% of 1040 participants developed abdominal obesity during follow-up, 27.7% of 1770 developed hypertriglyceridemia, 24.5% of 1810 developed low HDL cholesterol, and 41.4% of 1268 developed high fasting glucose concentration. The median consumption during the follow-up of total dairy products in the whole study population was 363 g/d (IQR: 257, 525 g/d), low-fat dairy products being the largest contributors to total dairy consumption (72.5%). The median of consumption of milk, yogurt, and cheese was 207 g/d, 70 g/d, and 30 g/d, respectively. The general characteristics of the study participants according to their mean consumption categories of total dairy products (tertiles) during the follow-up are shown in Table 1. Compared with those in the lowest tertile, participants in the top tertile were more likely to be older women and less likely to smoke, be physically active, and have lower serum concentrations of TGs and higher concentrations of HDL cholesterol. Participants in the highest tertile of dairy consumption also had lower total energy intake and consumed less red meat, fish, cereals, nuts, olive oil, and alcohol.\n\nConsumption of total dairy products and incidence of MetS. The multivariable-adjusted HRs (95% CIs) for MetS incidence across tertiles of consumption of total dairy products are shown in Table 2. After adjusting for several potential confounders, subjects in the top tertile of low-fat dairy consumption, but not total dairy or high-fat dairy, had a lower risk of incident MetS [HR: 0.72; 95% CI: 0.61, 0.86; P-trend = 0.001] compared with those in the bottom tertile.\n\nConsumption of dairy product subtypes and incidence of MetS. The HRs of incident MetS across tertiles of consumption of specific subtypes of dairy products (yogurt, cheese, and milk) also adjusted for potential confounders are shown in Table 3. Among the subtypes of dairy products, consumers in tertile 3 of low-fat yogurt [HR: 0.73; 95% CI: 0.62, 0.86; P-trend = 0.004], whole-fat yogurt [HR: 0.80; 95% CI: 0.67, 0.95); P-trend = 0.007] had a lower risk of developing MetS than did participants in the lowest tertile of consumption. Compared with participants in the lowest tertile of consumption of cheese, those in the highest tertile had an increased risk of incident MetS [HR: 1.31; 95% CI: 1.10, 1.56; P-trend < 0.001]. No statistical interactions were found between the consumption of total dairy products or subtypes and sex or intervention group.\n\nYogurt consumption and MetS and its components. The multivariable-adjusted HR of each MetS component in subjects who were initially free of MetS in extreme categories of total, whole-fat, and low-fat yogurt consumption (tertile 3 vs. tertile 1) are shown in Figure 1. With the exception of high blood pressure, participants in the highest tertile of total yogurt consumption had a significantly lower risk of developing each of the MetS components than those in the lowest tertile. However, the linear trend was significant only for high fasting glucose (P-trend = 0.004). Compared with participants in the lowest tertile, participants in the highest tertile of whole-fat yogurt consumption had a lower risk of several components of MetS, including abdominal obesity [HR: 0.80; 95% CI: 0.65, 0.98; P-trend = 0.048], hypertriglyceridemia [HR: 0.74; 95% CI: 0.64, 0.86; P-trend < 0.001], low HDL cholesterol [HR: 0.73; 95% CI: 0.63, 0.85; P-trend < 0.001], high blood pressure [HR: 0.62; 95% CI: 0.44, 0.86; P-trend = 0.001], and high fasting plasma glucose [HR: 0.80; 95% CI: 0.66, 0.94; P-trend = 0.005]. The associations with low-fat yogurt were in the same direction as those with total and whole-fat yogurt, but inverse associations were limited to hypertriglyceridemia [HR: 0.73; 95% CI: 0.63, 0.85; P-trend = 0.18], low HDL cholesterol [HR: 0.76; 95% CI: 0.66, 0.88; P-trend = 0.35], and high fasting plasma glucose [HR: 0.81; 95% CI: 0.68, 0.96; P-trend = 0.004].\n\nConsumption of dairy products, dairy product subtypes, and MetS components. HRs for the components of incident MetS across tertiles of consumption of total dairy products and subtypes are shown in Supplemental Table 1. An increased consumption of total milk and low-fat milk was significantly associated with a lower incidence of low HDL cholesterol and high fasting glucose. Low-fat dairy consumption was inversely associated with high fasting glucose, hypertriglyceridemia, and low HDL cholesterol (Supplemental Table 1).\n\nConsumption of dairy products, subtypes of dairy products, and incidence of MetS based on the number of MetS components at baseline. A sensitivity analysis based on the number of MetS components at baseline found that in those individuals who had only one component or none at all, there was no significant association between total dairy or its subtypes and MetS incidence, except for cheese (P-trend < 0.05). In those individuals with 2 MetS components at baseline, there were still significant inverse associations between low-fat dairy, low-fat yogurt, and whole-fat yogurt consumption and MetS incidence. A positive association was observed between cheese consumption and MetS development (Supplemental Table 2).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Consumption of Yogurt, Low-Fat Milk, and Other Low-Fat Dairy Products Is Associated with Lower Risk of Metabolic Syndrome Incidence in an Elderly Mediterranean Population.",
        "abstract": "## BACKGROUND\nThe association between consumption of dairy products and the risk of developing metabolic syndrome (MetS) is unclear.\n## OBJECTIVE\nThe purpose of this study was to evaluate the associations between consumption of dairy products (total and different subtypes) and incident MetS in a Mediterranean population at high cardiovascular disease risk.\n## METHODS\nWe prospectively analyzed 1868 men and women (55-80 y old) without MetS at baseline, recruited from different PREDIMED (Prevenci\u00f3n con Dieta Mediterr\u00e1nea) centers between October 2003 and June 2009 and followed up until December 2010. MetS was defined according to updated, harmonized criteria. At baseline and yearly thereafter, we determined anthropometric variables, dietary habits by a 137-item validated food-frequency questionnaire, and blood biochemistry. Multivariable-adjusted HRs of MetS or its components were estimated for each of the 2 upper tertiles (vs. the lowest one) of mean consumption of dairy products during the follow-up.\n## RESULTS\nDuring a median follow-up of 3.2 y, we documented 930 incident MetS cases. In the multivariable-adjusted model, HRs (95% CIs) of MetS for the comparison of extreme tertiles of dairy product consumption were 0.72 (0.61, 0.86) for low-fat dairy, 0.73 (0.62, 0.86) for low-fat yogurt, 0.78 (0.66, 0.92) for whole-fat yogurt, and 0.80 (0.67, 0.95) for low-fat milk. The respective HR for cheese was 1.31 (1.10, 1.56).\n## CONCLUSIONS\nHigher consumption of low-fat dairy products, yogurt (total, low-fat, and whole-fat yogurt) and low-fat milk was associated with a reduced risk of MetS in individuals at high cardiovascular disease risk from a Mediterranean population. Conversely, higher consumption of cheese was related to a higher risk of MetS. This trial was registered at controlled-trials.com as ISRCTN35739639.",
        "methods": "### Methods\n\n##### Study design, participants, and outcome. The present study was conducted within the framework of the PREDIMED trial, the design of which has been described in detail elsewhere (23, 24). The PREDIMED study is a large, parallel-group, multicenter, randomized, controlled field trial aimed at assessing the effects of the Mediterranean diet on the primary prevention of CVD (25). The main results of the trial with respect to the primary endpoint have been published recently (26). Briefly, 7447 community-dwelling men (aged 55\u201380 y) and women (aged 60\u201380 y) with no previously documented CVD were recruited. They were eligible if they had either T2D or \u22653 of the following CVD risk factors: hypertension (systolic blood pressure \u2265140 mm Hg or diastolic blood pressure \u226590 mm Hg or on antihypertensive medication), high plasma LDL cholesterol (\u2265160 mg/dL), low plasma HDL cholesterol (<40 mg/dL in men; <50 mg/dL in women), overweight or obesity (BMI \u226525 kg/m\u00b2), current smoking, or a family history of premature coronary heart disease (23). From October 2003 to June 2009, participants were randomly assigned to 3 intervention groups (2 of which were advised to follow a Mediterranean diet supplemented with either 1 L/wk extra-virgin olive oil or 30 g/d mixed nuts, and were compared with the third group, which was advised to follow a control low-fat diet). The study follow-up ended in December 2010. All participants provided their informed consent and the protocol was approved by the institutional review boards of each recruitment center. In the present report, the data were analyzed assuming the design of an observational prospective cohort whose members were selected from all the PREDIMED recruiting centers with biochemical determinations available for a follow-up of \u22652 y (n = 5801). Because our aim was to explore the associations between the consumption of dairy foods and incident MetS, we excluded participants who had diagnoses of MetS at baseline (63.9%; n = 3707). We also excluded participants who had not completed the baseline FFQ or who reported an extreme total energy intake with values outside the prespecified limits (500\u20133500 kcal/d in women and 800\u20134000 kcal/d in men). A total of 2094 individuals were assessed. Of these, a total of 226 were excluded because of missing data that prevented the presence of MetS incidence from being determined. Thus, a total of 1868 participants were included in our longitudinal assessment for MetS incidence. The individual components of MetS\u2014abdominal obesity, hypertriglyceridemia, low HDL cholesterol, high blood pressure, and high fasting glucose concentration\u2014were analyzed for a total of 1386, 3539, 3745, 337, and 1844 participants (from 5801 participants initially considered with biochemical determinations available for a follow-up of \u22652 y), respectively. The primary endpoint of the PREDIMED trial was a combination of several major cardiovascular clinical events (myocardial infarction, stroke, or CVD death). In the present analysis, incident MetS and its components were considered as the outcome. MetS was defined in accordance with the updated harmonized criteria of the International Diabetes Federation and the AHA/National Heart, Lung, and Blood Institute (1). Individuals were diagnosed with MetS if they had \u22653 of the following components: hypertriglyceridemia concentrations of HDL cholesterol [<50 mg/dL (<1.3 mmol/L) and <40 mg/dL (<1.03 mmol/L) in women and men, respectively] or drug treatment for low HDL cholesterol; elevated blood pressure (systolic \u2265130 mmHg and/or diastolic \u226585 mmHg) or being treated for hypertension; high fasting plasma glucose [\u2265100 mg/dL (\u22655.5 mmol/L)] or drug treatment for hyperglycemia; and elevated waist circumference for European individuals (\u226588 cm in women and \u2265102 cm in men).\n\nDietary assessment. At baseline and yearly during follow-up, dietary intake was quantified by trained dietitians with a 137-item semi-quantitative FFQ validated for the PREDIMED study (27). In the validation study, the FFQ was administered twice to explore reproducibility at 1 y; 4 3-d dietary records for the different seasons of the year were used as gold standard. The reproducibility of the FFQ used in PREDIMED for food groups and energy and nutrient intake, explored by the Pearson correlation coefficient, ranged from 0.50 to 0.82, and the intraclass correlation coefficient (ICC) ranged from 0.63 to 0.90. The validity indexes of the FFQ in relation to the dietary records for food groups and energy and nutrient intake ranged from 0.24 to 0.72, whereas the range of the ICC was between 0.40 and 0.84. Regarding dairy product consumption, the reproducibility and validity of FFQs were 0.81 (ICC 0.89), and 0.72 (ICC 0.84), respectively. Dairy product consumption was assessed yearly with the use of 15 items from the FFQ. In order to assess habitual dietary intake over the previous year, frequencies of consumption were measured in 9 categories (ranging from never/almost never to >6 servings/d) for each food item. These responses to individual dairy items were then converted into mean daily consumption (grams per day) during the follow-up by multiplying the typical portion sizes (grams) by the consumption frequency for each food and making the appropriate division for the period assessed to obtain daily consumption. The total dairy foods category included low-fat/skim milk and skim yogurt, whole milk, condensed milk, whole yogurt, custard, and all types of cheeses, including petit Swiss, ricotta, cottage, and semicured/cured cheeses such as cheddar, Manchego, and Emmentaler). Low-fat dairy foods included low-fat/skim milk and skim yogurt, whole-fat dairy foods (whole milk and whole yogurt), and total dairy foods, including all of the above. Consumption of dairy products was also categorized by subtype, including milk (including total, low-fat and whole milk), yogurt (including total, low-fat and whole-fat yogurt), and cheese. Energy and nutrient intake and food groups were calculated from Spanish food composition tables (28, 29). We adjusted dairy consumption for total energy intake with the use of the nutrient residual method (30).\n\nMeasurements. Participants completed the following at baseline and yearly: 1) a questionnaire about lifestyle variables, medical history, and medication use; 2) a 14-item validated questionnaire (31) designed to assess adherence to the Mediterranean diet; 3) a validated 137-item semi-quantitative FFQ (27); and 4) the validated Spanish version (32) of the Minnesota Leisure-Time Physical Activity Questionnaire. In addition, anthropometric variables and blood pressure were determined by trained staff. Blood pressure was measured in triplicate with the use of a validated semiautomatic oscillometer with a 5 min interval between measurements and the subject in a sitting position (Omron HEM-705CP). Blood samples were collected after an overnight fast, coded, shipped to a central laboratory, and stored at -80\u00b0C until analysis. Biochemical analysis was performed in local laboratories. Glucose was measured by the glucose-oxidase method, cholesterol by esterase-oxidase-peroxidase, TGs by glycerol-phosphate oxidase-peroxidase, and HDL cholesterol by direct measurement. All local laboratories satisfied external quality-control requirements. When TGs were <300 mg/dL, LDL cholesterol was calculated with the use of the Friedewald formula so that the LDL cholesterol was not underestimated. A concordance study of 9 laboratories was conducted. From each study, a mean of 200 samples were analyzed for total cholesterol, HDL cholesterol, and TGs with the use of the Medical Research Institute of del Mar laboratory as reference. The Medical Research Institute of del Mar laboratory used ABX-Horiba commercial kits in a PENTRA-400 autoanalyzer (ABX-Horiba). One center was unable to provide samples for the concordance study. The analysis of concordance of lipid measurements showed an r\u00b2 and an ICC (95% CI) between 0.85 and 0.97, and 0.85 (0.77, 0.90) and 0.97 (0.81, 0.83) and 0.92 (0.89, 0.95) for HDL cholesterol, respectively; between 0.81 and 0.99, and 0.81 (0.73, 0.87) and 0.99 (0.99, 0.99) for triglycerides, respectively; and between 0.82 and 0.96, and 0.82 (0.74, 0.88) and 0.99 (0.99, 0.99) for glucose, respectively.\n\nStatistical analysis. We averaged the intake reported during the baseline interview and the yearly consumption during the follow-up. Then, participants were categorized into tertiles of the mean consumption of total dairy products and different subtypes during the follow-up. To better represent the long-term consumption of dairy products and to minimize within-person variation, we used the mean energy-adjusted dairy consumption for all analyses based on assessments of items from all FFQs, which were administered at baseline and yearly during the follow-up for those participants who did not develop MetS. For those who did develop MetS, and given that participants can alter their dietary pattern after developing MetS, we only used data from all the available FFQs until the year before MetS was diagnosed. The baseline characteristics of the participants are expressed as mean \u00b1 SD or median (IQR) for continuous variables and number and percentages for categorical variables. Chi-square and 1-factor ANOVA tests were used to assess differences in the baseline characteristics of the study population. Multivariable time-dependent Cox proportional regression models were fitted to assess the HRs of developing MetS and its components during follow-up according to tertiles of consumption of total, low-fat, and whole-fat dairy products; milk, low-fat milk, and whole milk; total yogurt, low-fat yogurt, and whole-fat yogurt; and cheese. Both upper tertiles were compared with the lowest tertile (reference). The assumption of proportional hazards was tested with the use of time-dependent covariates. The time variable was the interval between random assignment and the date of the last follow-up, or the last recorded clinical event (MetS incidence) of participants who were still alive, whichever occurred first. Participants who were free of MetS or who were lost during follow-up were censored at the date of the last visit. Three different Cox regression models were adjusted for potential confounding factors. Model 1 was adjusted for intervention group; sex; age (years); leisure time physical activity (metabolic equivalent task \u00d7 d); BMI (kilograms per meter squared); current smoker (yes/no); former smoker (yes/no); and hypoglycemic, hypolipidemic, antihypertensive, and insulin treatment at baseline. Model 2 was additionally adjusted for mean consumption during follow-up of vegetables, fruit, legumes, cereals, fish, red meat, cookies, olive oil, and nuts (all grams per day), as well as alcohol (grams per day and quadratic term). Model 3 was adjusted for model 2 plus the prevalence of MetS components at baseline, including abdominal obesity (yes/no), hypertriglyceridemia (yes/no), low HDL cholesterol (yes/no), hypertension (yes/no), and high fasting plasma glucose (yes/no). Statistical interaction between tertiles of total dairy consumption and its different subtypes, and potential effect-modifying variables, such as sex and intervention group, was assessed by including product terms in the models. To assess the linear trend, the median value of each tertile of dairy product and the different dairy subtypes was assigned and used as a continuous variable in the Cox regression models. The level of significance for all statistical tests was P < 0.05 for bilateral contrast. The Benjamini-Hochberg method was used to correct P values for multiple comparisons (33). We also conducted a sensitivity analysis of a number of MetS components at baseline to test the robustness of our results. The main analysis was stratified by those individuals who did not have any components of MetS or had one, or who had 2 components. This was done to prevent possible bias effects, because it is easier for those who already have 2 criteria to meet the diagnostic definition than those with one or none. All analyses were performed with SPSS software, version 19.0.",
        "results": "### Results\n\nAfter a median of 3.2 y of follow-up (IQR: 1.9, 5.8), 930 participants without MetS at baseline (53.8% women) developed new-onset MetS. Of those not showing the specific MetS components at baseline, 43.4% of 1040 participants developed abdominal obesity during follow-up, 27.7% of 1770 developed hypertriglyceridemia, 24.5% of 1810 developed low HDL cholesterol, and 41.4% of 1268 developed high fasting glucose concentration. The median consumption during the follow-up of total dairy products in the whole study population was 363 g/d (IQR: 257, 525 g/d), low-fat dairy products being the largest contributors to total dairy consumption (72.5%). The median of consumption of milk, yogurt, and cheese was 207 g/d, 70 g/d, and 30 g/d, respectively. The general characteristics of the study participants according to their mean consumption categories of total dairy products (tertiles) during the follow-up are shown in Table 1. Compared with those in the lowest tertile, participants in the top tertile were more likely to be older women and less likely to smoke, be physically active, and have lower serum concentrations of TGs and higher concentrations of HDL cholesterol. Participants in the highest tertile of dairy consumption also had lower total energy intake and consumed less red meat, fish, cereals, nuts, olive oil, and alcohol.\n\nConsumption of total dairy products and incidence of MetS. The multivariable-adjusted HRs (95% CIs) for MetS incidence across tertiles of consumption of total dairy products are shown in Table 2. After adjusting for several potential confounders, subjects in the top tertile of low-fat dairy consumption, but not total dairy or high-fat dairy, had a lower risk of incident MetS [HR: 0.72; 95% CI: 0.61, 0.86; P-trend = 0.001] compared with those in the bottom tertile.\n\nConsumption of dairy product subtypes and incidence of MetS. The HRs of incident MetS across tertiles of consumption of specific subtypes of dairy products (yogurt, cheese, and milk) also adjusted for potential confounders are shown in Table 3. Among the subtypes of dairy products, consumers in tertile 3 of low-fat yogurt [HR: 0.73; 95% CI: 0.62, 0.86; P-trend = 0.004], whole-fat yogurt [HR: 0.80; 95% CI: 0.67, 0.95); P-trend = 0.007] had a lower risk of developing MetS than did participants in the lowest tertile of consumption. Compared with participants in the lowest tertile of consumption of cheese, those in the highest tertile had an increased risk of incident MetS [HR: 1.31; 95% CI: 1.10, 1.56; P-trend < 0.001]. No statistical interactions were found between the consumption of total dairy products or subtypes and sex or intervention group.\n\nYogurt consumption and MetS and its components. The multivariable-adjusted HR of each MetS component in subjects who were initially free of MetS in extreme categories of total, whole-fat, and low-fat yogurt consumption (tertile 3 vs. tertile 1) are shown in Figure 1. With the exception of high blood pressure, participants in the highest tertile of total yogurt consumption had a significantly lower risk of developing each of the MetS components than those in the lowest tertile. However, the linear trend was significant only for high fasting glucose (P-trend = 0.004). Compared with participants in the lowest tertile, participants in the highest tertile of whole-fat yogurt consumption had a lower risk of several components of MetS, including abdominal obesity [HR: 0.80; 95% CI: 0.65, 0.98; P-trend = 0.048], hypertriglyceridemia [HR: 0.74; 95% CI: 0.64, 0.86; P-trend < 0.001], low HDL cholesterol [HR: 0.73; 95% CI: 0.63, 0.85; P-trend < 0.001], high blood pressure [HR: 0.62; 95% CI: 0.44, 0.86; P-trend = 0.001], and high fasting plasma glucose [HR: 0.80; 95% CI: 0.66, 0.94; P-trend = 0.005]. The associations with low-fat yogurt were in the same direction as those with total and whole-fat yogurt, but inverse associations were limited to hypertriglyceridemia [HR: 0.73; 95% CI: 0.63, 0.85; P-trend = 0.18], low HDL cholesterol [HR: 0.76; 95% CI: 0.66, 0.88; P-trend = 0.35], and high fasting plasma glucose [HR: 0.81; 95% CI: 0.68, 0.96; P-trend = 0.004].\n\nConsumption of dairy products, dairy product subtypes, and MetS components. HRs for the components of incident MetS across tertiles of consumption of total dairy products and subtypes are shown in Supplemental Table 1. An increased consumption of total milk and low-fat milk was significantly associated with a lower incidence of low HDL cholesterol and high fasting glucose. Low-fat dairy consumption was inversely associated with high fasting glucose, hypertriglyceridemia, and low HDL cholesterol (Supplemental Table 1).\n\nConsumption of dairy products, subtypes of dairy products, and incidence of MetS based on the number of MetS components at baseline. A sensitivity analysis based on the number of MetS components at baseline found that in those individuals who had only one component or none at all, there was no significant association between total dairy or its subtypes and MetS incidence, except for cheese (P-trend < 0.05). In those individuals with 2 MetS components at baseline, there were still significant inverse associations between low-fat dairy, low-fat yogurt, and whole-fat yogurt consumption and MetS incidence. A positive association was observed between cheese consumption and MetS development (Supplemental Table 2).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Consumption of Yogurt, Low-Fat Milk, and Other Low-Fat Dairy Products Is Associated with Lower Risk of Metabolic Syndrome Incidence in an Elderly Mediterranean Population.",
        "abstract": "## BACKGROUND\nThe association between consumption of dairy products and the risk of developing metabolic syndrome (MetS) is unclear.\n## OBJECTIVE\nThe purpose of this study was to evaluate the associations between consumption of dairy products (total and different subtypes) and incident MetS in a Mediterranean population at high cardiovascular disease risk.\n## METHODS\nWe prospectively analyzed 1868 men and women (55-80 y old) without MetS at baseline, recruited from different PREDIMED (Prevenci\u00f3n con Dieta Mediterr\u00e1nea) centers between October 2003 and June 2009 and followed up until December 2010. MetS was defined according to updated, harmonized criteria. At baseline and yearly thereafter, we determined anthropometric variables, dietary habits by a 137-item validated food-frequency questionnaire, and blood biochemistry. Multivariable-adjusted HRs of MetS or its components were estimated for each of the 2 upper tertiles (vs. the lowest one) of mean consumption of dairy products during the follow-up.\n## RESULTS\nDuring a median follow-up of 3.2 y, we documented 930 incident MetS cases. In the multivariable-adjusted model, HRs (95% CIs) of MetS for the comparison of extreme tertiles of dairy product consumption were 0.72 (0.61, 0.86) for low-fat dairy, 0.73 (0.62, 0.86) for low-fat yogurt, 0.78 (0.66, 0.92) for whole-fat yogurt, and 0.80 (0.67, 0.95) for low-fat milk. The respective HR for cheese was 1.31 (1.10, 1.56).\n## CONCLUSIONS\nHigher consumption of low-fat dairy products, yogurt (total, low-fat, and whole-fat yogurt) and low-fat milk was associated with a reduced risk of MetS in individuals at high cardiovascular disease risk from a Mediterranean population. Conversely, higher consumption of cheese was related to a higher risk of MetS. This trial was registered at controlled-trials.com as ISRCTN35739639.",
        "methods": "### Methods\n\n##### Study design, participants, and outcome. The present study was conducted within the framework of the PREDIMED trial, the design of which has been described in detail elsewhere (23, 24). The PREDIMED study is a large, parallel-group, multicenter, randomized, controlled field trial aimed at assessing the effects of the Mediterranean diet on the primary prevention of CVD (25). The main results of the trial with respect to the primary endpoint have been published recently (26). Briefly, 7447 community-dwelling men (aged 55\u201380 y) and women (aged 60\u201380 y) with no previously documented CVD were recruited. They were eligible if they had either T2D or \u22653 of the following CVD risk factors: hypertension (systolic blood pressure \u2265140 mm Hg or diastolic blood pressure \u226590 mm Hg or on antihypertensive medication), high plasma LDL cholesterol (\u2265160 mg/dL), low plasma HDL cholesterol (<40 mg/dL in men; <50 mg/dL in women), overweight or obesity (BMI \u226525 kg/m\u00b2), current smoking, or a family history of premature coronary heart disease (23). From October 2003 to June 2009, participants were randomly assigned to 3 intervention groups (2 of which were advised to follow a Mediterranean diet supplemented with either 1 L/wk extra-virgin olive oil or 30 g/d mixed nuts, and were compared with the third group, which was advised to follow a control low-fat diet). The study follow-up ended in December 2010. All participants provided their informed consent and the protocol was approved by the institutional review boards of each recruitment center. In the present report, the data were analyzed assuming the design of an observational prospective cohort whose members were selected from all the PREDIMED recruiting centers with biochemical determinations available for a follow-up of \u22652 y (n = 5801). Because our aim was to explore the associations between the consumption of dairy foods and incident MetS, we excluded participants who had diagnoses of MetS at baseline (63.9%; n = 3707). We also excluded participants who had not completed the baseline FFQ or who reported an extreme total energy intake with values outside the prespecified limits (500\u20133500 kcal/d in women and 800\u20134000 kcal/d in men). A total of 2094 individuals were assessed. Of these, a total of 226 were excluded because of missing data that prevented the presence of MetS incidence from being determined. Thus, a total of 1868 participants were included in our longitudinal assessment for MetS incidence. The individual components of MetS\u2014abdominal obesity, hypertriglyceridemia, low HDL cholesterol, high blood pressure, and high fasting glucose concentration\u2014were analyzed for a total of 1386, 3539, 3745, 337, and 1844 participants (from 5801 participants initially considered with biochemical determinations available for a follow-up of \u22652 y), respectively. The primary endpoint of the PREDIMED trial was a combination of several major cardiovascular clinical events (myocardial infarction, stroke, or CVD death). In the present analysis, incident MetS and its components were considered as the outcome. MetS was defined in accordance with the updated harmonized criteria of the International Diabetes Federation and the AHA/National Heart, Lung, and Blood Institute (1). Individuals were diagnosed with MetS if they had \u22653 of the following components: hypertriglyceridemia concentrations of HDL cholesterol [<50 mg/dL (<1.3 mmol/L) and <40 mg/dL (<1.03 mmol/L) in women and men, respectively] or drug treatment for low HDL cholesterol; elevated blood pressure (systolic \u2265130 mmHg and/or diastolic \u226585 mmHg) or being treated for hypertension; high fasting plasma glucose [\u2265100 mg/dL (\u22655.5 mmol/L)] or drug treatment for hyperglycemia; and elevated waist circumference for European individuals (\u226588 cm in women and \u2265102 cm in men).\n\nDietary assessment. At baseline and yearly during follow-up, dietary intake was quantified by trained dietitians with a 137-item semi-quantitative FFQ validated for the PREDIMED study (27). In the validation study, the FFQ was administered twice to explore reproducibility at 1 y; 4 3-d dietary records for the different seasons of the year were used as gold standard. The reproducibility of the FFQ used in PREDIMED for food groups and energy and nutrient intake, explored by the Pearson correlation coefficient, ranged from 0.50 to 0.82, and the intraclass correlation coefficient (ICC) ranged from 0.63 to 0.90. The validity indexes of the FFQ in relation to the dietary records for food groups and energy and nutrient intake ranged from 0.24 to 0.72, whereas the range of the ICC was between 0.40 and 0.84. Regarding dairy product consumption, the reproducibility and validity of FFQs were 0.81 (ICC 0.89), and 0.72 (ICC 0.84), respectively. Dairy product consumption was assessed yearly with the use of 15 items from the FFQ. In order to assess habitual dietary intake over the previous year, frequencies of consumption were measured in 9 categories (ranging from never/almost never to >6 servings/d) for each food item. These responses to individual dairy items were then converted into mean daily consumption (grams per day) during the follow-up by multiplying the typical portion sizes (grams) by the consumption frequency for each food and making the appropriate division for the period assessed to obtain daily consumption. The total dairy foods category included low-fat/skim milk and skim yogurt, whole milk, condensed milk, whole yogurt, custard, and all types of cheeses, including petit Swiss, ricotta, cottage, and semicured/cured cheeses such as cheddar, Manchego, and Emmentaler). Low-fat dairy foods included low-fat/skim milk and skim yogurt, whole-fat dairy foods (whole milk and whole yogurt), and total dairy foods, including all of the above. Consumption of dairy products was also categorized by subtype, including milk (including total, low-fat and whole milk), yogurt (including total, low-fat and whole-fat yogurt), and cheese. Energy and nutrient intake and food groups were calculated from Spanish food composition tables (28, 29). We adjusted dairy consumption for total energy intake with the use of the nutrient residual method (30).\n\nMeasurements. Participants completed the following at baseline and yearly: 1) a questionnaire about lifestyle variables, medical history, and medication use; 2) a 14-item validated questionnaire (31) designed to assess adherence to the Mediterranean diet; 3) a validated 137-item semi-quantitative FFQ (27); and 4) the validated Spanish version (32) of the Minnesota Leisure-Time Physical Activity Questionnaire. In addition, anthropometric variables and blood pressure were determined by trained staff. Blood pressure was measured in triplicate with the use of a validated semiautomatic oscillometer with a 5 min interval between measurements and the subject in a sitting position (Omron HEM-705CP). Blood samples were collected after an overnight fast, coded, shipped to a central laboratory, and stored at -80\u00b0C until analysis. Biochemical analysis was performed in local laboratories. Glucose was measured by the glucose-oxidase method, cholesterol by esterase-oxidase-peroxidase, TGs by glycerol-phosphate oxidase-peroxidase, and HDL cholesterol by direct measurement. All local laboratories satisfied external quality-control requirements. When TGs were <300 mg/dL, LDL cholesterol was calculated with the use of the Friedewald formula so that the LDL cholesterol was not underestimated. A concordance study of 9 laboratories was conducted. From each study, a mean of 200 samples were analyzed for total cholesterol, HDL cholesterol, and TGs with the use of the Medical Research Institute of del Mar laboratory as reference. The Medical Research Institute of del Mar laboratory used ABX-Horiba commercial kits in a PENTRA-400 autoanalyzer (ABX-Horiba). One center was unable to provide samples for the concordance study. The analysis of concordance of lipid measurements showed an r\u00b2 and an ICC (95% CI) between 0.85 and 0.97, and 0.85 (0.77, 0.90) and 0.97 (0.81, 0.83) and 0.92 (0.89, 0.95) for HDL cholesterol, respectively; between 0.81 and 0.99, and 0.81 (0.73, 0.87) and 0.99 (0.99, 0.99) for triglycerides, respectively; and between 0.82 and 0.96, and 0.82 (0.74, 0.88) and 0.99 (0.99, 0.99) for glucose, respectively.\n\nStatistical analysis. We averaged the intake reported during the baseline interview and the yearly consumption during the follow-up. Then, participants were categorized into tertiles of the mean consumption of total dairy products and different subtypes during the follow-up. To better represent the long-term consumption of dairy products and to minimize within-person variation, we used the mean energy-adjusted dairy consumption for all analyses based on assessments of items from all FFQs, which were administered at baseline and yearly during the follow-up for those participants who did not develop MetS. For those who did develop MetS, and given that participants can alter their dietary pattern after developing MetS, we only used data from all the available FFQs until the year before MetS was diagnosed. The baseline characteristics of the participants are expressed as mean \u00b1 SD or median (IQR) for continuous variables and number and percentages for categorical variables. Chi-square and 1-factor ANOVA tests were used to assess differences in the baseline characteristics of the study population. Multivariable time-dependent Cox proportional regression models were fitted to assess the HRs of developing MetS and its components during follow-up according to tertiles of consumption of total, low-fat, and whole-fat dairy products; milk, low-fat milk, and whole milk; total yogurt, low-fat yogurt, and whole-fat yogurt; and cheese. Both upper tertiles were compared with the lowest tertile (reference). The assumption of proportional hazards was tested with the use of time-dependent covariates. The time variable was the interval between random assignment and the date of the last follow-up, or the last recorded clinical event (MetS incidence) of participants who were still alive, whichever occurred first. Participants who were free of MetS or who were lost during follow-up were censored at the date of the last visit. Three different Cox regression models were adjusted for potential confounding factors. Model 1 was adjusted for intervention group; sex; age (years); leisure time physical activity (metabolic equivalent task \u00d7 d); BMI (kilograms per meter squared); current smoker (yes/no); former smoker (yes/no); and hypoglycemic, hypolipidemic, antihypertensive, and insulin treatment at baseline. Model 2 was additionally adjusted for mean consumption during follow-up of vegetables, fruit, legumes, cereals, fish, red meat, cookies, olive oil, and nuts (all grams per day), as well as alcohol (grams per day and quadratic term). Model 3 was adjusted for model 2 plus the prevalence of MetS components at baseline, including abdominal obesity (yes/no), hypertriglyceridemia (yes/no), low HDL cholesterol (yes/no), hypertension (yes/no), and high fasting plasma glucose (yes/no). Statistical interaction between tertiles of total dairy consumption and its different subtypes, and potential effect-modifying variables, such as sex and intervention group, was assessed by including product terms in the models. To assess the linear trend, the median value of each tertile of dairy product and the different dairy subtypes was assigned and used as a continuous variable in the Cox regression models. The level of significance for all statistical tests was P < 0.05 for bilateral contrast. The Benjamini-Hochberg method was used to correct P values for multiple comparisons (33). We also conducted a sensitivity analysis of a number of MetS components at baseline to test the robustness of our results. The main analysis was stratified by those individuals who did not have any components of MetS or had one, or who had 2 components. This was done to prevent possible bias effects, because it is easier for those who already have 2 criteria to meet the diagnostic definition than those with one or none. All analyses were performed with SPSS software, version 19.0.",
        "results": "### Results\n\nAfter a median of 3.2 y of follow-up (IQR: 1.9, 5.8), 930 participants without MetS at baseline (53.8% women) developed new-onset MetS. Of those not showing the specific MetS components at baseline, 43.4% of 1040 participants developed abdominal obesity during follow-up, 27.7% of 1770 developed hypertriglyceridemia, 24.5% of 1810 developed low HDL cholesterol, and 41.4% of 1268 developed high fasting glucose concentration. The median consumption during the follow-up of total dairy products in the whole study population was 363 g/d (IQR: 257, 525 g/d), low-fat dairy products being the largest contributors to total dairy consumption (72.5%). The median of consumption of milk, yogurt, and cheese was 207 g/d, 70 g/d, and 30 g/d, respectively. The general characteristics of the study participants according to their mean consumption categories of total dairy products (tertiles) during the follow-up are shown in Table 1. Compared with those in the lowest tertile, participants in the top tertile were more likely to be older women and less likely to smoke, be physically active, and have lower serum concentrations of TGs and higher concentrations of HDL cholesterol. Participants in the highest tertile of dairy consumption also had lower total energy intake and consumed less red meat, fish, cereals, nuts, olive oil, and alcohol.\n\nConsumption of total dairy products and incidence of MetS. The multivariable-adjusted HRs (95% CIs) for MetS incidence across tertiles of consumption of total dairy products are shown in Table 2. After adjusting for several potential confounders, subjects in the top tertile of low-fat dairy consumption, but not total dairy or high-fat dairy, had a lower risk of incident MetS [HR: 0.72; 95% CI: 0.61, 0.86; P-trend = 0.001] compared with those in the bottom tertile.\n\nConsumption of dairy product subtypes and incidence of MetS. The HRs of incident MetS across tertiles of consumption of specific subtypes of dairy products (yogurt, cheese, and milk) also adjusted for potential confounders are shown in Table 3. Among the subtypes of dairy products, consumers in tertile 3 of low-fat yogurt [HR: 0.73; 95% CI: 0.62, 0.86; P-trend = 0.004], whole-fat yogurt [HR: 0.80; 95% CI: 0.67, 0.95); P-trend = 0.007] had a lower risk of developing MetS than did participants in the lowest tertile of consumption. Compared with participants in the lowest tertile of consumption of cheese, those in the highest tertile had an increased risk of incident MetS [HR: 1.31; 95% CI: 1.10, 1.56; P-trend < 0.001]. No statistical interactions were found between the consumption of total dairy products or subtypes and sex or intervention group.\n\nYogurt consumption and MetS and its components. The multivariable-adjusted HR of each MetS component in subjects who were initially free of MetS in extreme categories of total, whole-fat, and low-fat yogurt consumption (tertile 3 vs. tertile 1) are shown in Figure 1. With the exception of high blood pressure, participants in the highest tertile of total yogurt consumption had a significantly lower risk of developing each of the MetS components than those in the lowest tertile. However, the linear trend was significant only for high fasting glucose (P-trend = 0.004). Compared with participants in the lowest tertile, participants in the highest tertile of whole-fat yogurt consumption had a lower risk of several components of MetS, including abdominal obesity [HR: 0.80; 95% CI: 0.65, 0.98; P-trend = 0.048], hypertriglyceridemia [HR: 0.74; 95% CI: 0.64, 0.86; P-trend < 0.001], low HDL cholesterol [HR: 0.73; 95% CI: 0.63, 0.85; P-trend < 0.001], high blood pressure [HR: 0.62; 95% CI: 0.44, 0.86; P-trend = 0.001], and high fasting plasma glucose [HR: 0.80; 95% CI: 0.66, 0.94; P-trend = 0.005]. The associations with low-fat yogurt were in the same direction as those with total and whole-fat yogurt, but inverse associations were limited to hypertriglyceridemia [HR: 0.73; 95% CI: 0.63, 0.85; P-trend = 0.18], low HDL cholesterol [HR: 0.76; 95% CI: 0.66, 0.88; P-trend = 0.35], and high fasting plasma glucose [HR: 0.81; 95% CI: 0.68, 0.96; P-trend = 0.004].\n\nConsumption of dairy products, dairy product subtypes, and MetS components. HRs for the components of incident MetS across tertiles of consumption of total dairy products and subtypes are shown in Supplemental Table 1. An increased consumption of total milk and low-fat milk was significantly associated with a lower incidence of low HDL cholesterol and high fasting glucose. Low-fat dairy consumption was inversely associated with high fasting glucose, hypertriglyceridemia, and low HDL cholesterol (Supplemental Table 1).\n\nConsumption of dairy products, subtypes of dairy products, and incidence of MetS based on the number of MetS components at baseline. A sensitivity analysis based on the number of MetS components at baseline found that in those individuals who had only one component or none at all, there was no significant association between total dairy or its subtypes and MetS incidence, except for cheese (P-trend < 0.05). In those individuals with 2 MetS components at baseline, there were still significant inverse associations between low-fat dairy, low-fat yogurt, and whole-fat yogurt consumption and MetS incidence. A positive association was observed between cheese consumption and MetS development (Supplemental Table 2).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "The association between nutritional exposures and metabolic syndrome in the Tehran Lipid and Glucose Study (TLGS): a cohort study.",
        "abstract": "## OBJECTIVES\nThis study was conducted with the goal of investigating the effect of various food consumption on the incidence of metabolic syndrome (MetS).\n## STUDY DESIGN\nProspective cohort study.\n## METHODS\nThis study conducted on 3616 healthy adults aged \u226520\u00a0years, who were not affected with MetS. Nutritional intake was determined at the beginning of the study (2008-2011) by the Food Frequency Questionnaire (FFQ), and the incidence of MetS was investigated after a median of 24.6 months follow-up. Data were analyzed by multiple logistic regression, with 95% confidence interval.\n## RESULTS\nAfter adjusting the effect of other variables in the model many foods in the whole grains group such as Sangak bread (OR\u00a0=\u00a00.35, 95% CI: 0.11-1.56), multiple types of vegetables and fruits - such as peach (OR\u00a0=\u00a00.11, 95% CI: 0.01-0.75), and mushroom (OR\u00a0=\u00a00.28, 95% CI: 0.11-0.71) had protective effects against MetS. From the dairy group yoghurt (OR\u00a0=\u00a00.43, 95% CI: 0.18-1.01) and ice cream (OR\u00a0=\u00a00.35, 95% CI: 0.12-1.06) had similar such effects.\n## CONCLUSION\nBased on our findings, there was a significant rise in the incidence of MetS in Iran. The daily consumption of foods from the whole grains, dairy, vegetable and fruit groups can help reduce the odds of MetS.",
        "methods": "#### Methods\n\nThis prospective study is part of a larger study titled the Tehran Lipid and Glucose Study (TLGS). It is also part of a thesis that was conducted among Tehranian adults during three years of follow-up from 2011 to 2014. In this study, we have used data from the prospective TLGS cohort, which has been conducted on 15,005 persons aged 3\u201363 years old, from Tehran's district 13. The study began in 1998. We used the data collected in the fourth (2008\u20132011) and fifth (2011\u20132014) phases of the study. Data related to dietary intake and confounding variables were collected from the fourth phase, and new MetS cases were chosen from the fifth phase, which is considered as the follow-up phase.\n\n**Target population**\n\nAll healthy adults aged \u226520 years not affected with MetS at the beginning of the study (fourth phase) who had dietary information were selected. The newly affected MetS cases were 590 persons. The reason for selecting the \u226520 years age group was that the MetS definition holds true for adults; it is different for persons aged <20 years.\n\n**Inclusion criteria**\n\nThose who were eligible for the study included persons aged \u226520 years who had been followed up from the fourth to the fifth phase, and who possessed the following criteria: no history of chronic diseases (diabetes, stroke, thyroid problems and cancer); not having any specific dietary regime (such as a weight loss diet and the intake of fewer than 800 kcal or greater than 4000 kcal); and not being affected with the MetS.\n\n**Measurement of outcome**\n\nIn the TLGS study, persons with MetS were identified using the Adult Treatment Panel (ATP) Guidelines. The criteria included having three of the following five indicators simultaneously: (1) abdominal obesity (waist circumference > 90 cm in both genders); (2) lowered serum HDL levels (lower than 40 mg/dl in men and 50 mg/dl in women, or the consumption of HDL-elevating drugs); (3) hypertension (a systolic BP \u2265 130 mmHg or a diastolic BP \u2265 85 mmHg, or the consumption of antihypertensive drugs); (4) disordered FBS level (a fasting blood glucose/FBS \u2265100 mg/dl, or the consumption of hypoglycemic drugs); (5) hypertriglyceridemia (a serum triglyceride level \u2265150 mg/dl, or the consumption of triglyceride-lowering drugs). Among the five components of MetS, four components were measured only once (at the end of each phase, i.e. once every three years), including waist circumference, serum HDL, blood sugar, and serum triglycerides. Only hypertension was measured twice during each phase. Each time, blood pressure was measured twice with at least a 30-second interval between the two measurements. Thereafter the mean of the two measurements was considered as the participant's BP for that session.\n\n**Measurement of exposure**\n\nNutritional information on the participants' dietary intake was gathered using the Food Frequency Questionnaire (FFQ), which contains 147 food items. A nutritionist who had been trained in the field completed the questionnaires through a face-to-face interview. During the interview, the average size of each of the FFQ food items was explained to the participant, and then she/he was asked about the number of times each item was taken in the past year. The validity and reliability of this questionnaire has already been assessed during a pilot study on 133 samples, wherein Pearson's correlation coefficient was acceptable. The consumption frequency of each food item in the past year was assessed in the form of days, weeks, months and/or years, and then, using home scale guides, the amount consumed of each item was transformed into grams per day. The amount of intake of energy and macronutrients and micronutrients were determined using the food composition table. The data related to food intake in the fourth phase were considered as the exposure.\n\n**Other measured variables**\n\nHere, we measured the physical activity during the past year with the standard physical activity questionnaire. The frequency and duration of each physical activity in the past 12 months was asked, and was then calculated as Metabolic Equivalent (MET) hours per week. Data on physical activity status were obtained using the modified activity questionnaire (MAC). The MAQ determines the current (past year and past week) physical activity during occupation and leisure time. Total physical activities were explicit in hours of activity per week by adding leisure time physical activity to occupational activity. Other variables measured were as follows: weight (in kilograms: wearing minimum clothes, without shoes, using a seca weighing scale \u2013 with an accuracy of 0.5 kg); height (without shoes, using a meter fixed on the wall \u2013 with an accuracy of 1 cm); waist circumference (the circumference at the slimmest area between the lowest rib and the tip of the iliac crest); systolic and diastolic blood pressure (BP) (using a manual sphygmomanometer, the mean was considered as the person's BP and reported in mm Hg); and finally, a family history of diabetes. Moreover, the necessary data on age, gender, education (primary, intermediate, high school and high school graduate, academic education), narcotics use (never smoked, previous smoked, currently smoking), and drug or supplementary intake were collected using a general information questionnaire through an interview conducted by the nutritionist. A venous blood sample was taken after 10 h nocturnal fasting to measure serum HDL, glucose, triglycerides and cholesterol. Blood glucose was measured with glucose oxidase; serum triglycerides were measured with the Pars Azmoon Kit \u2013 with the AutoAnalyzer Selectra; and serum HDL was measured after precipitating `b` apolipoprotein with phosphotungstate.\n\n**Data analysis**\n\nWe used bi-variable logistic regression to estimate the crude effect of each of the nutritional and non-nutritional variables on MetS. We then used multiple logistic regression to estimate the adjusted effect of variables. The backward-selection method was applied to choose the variables in this model. To remove the variables from the model the P-value was set at 0.2. Seventy variables (three confounding variables of educational status, smoking history and history of weight loss, in addition to 68 food items) were removed at this stage. In addition, six items were removed from the model because of their high correlation. These food items included the following: jam (and sugar); plum (and peach); lemon juice (and lemon); apple juice (and apple); orange juice (and orange); cooked vegetables (and cooked carrots). Between the two similar food items that were correlated one that was consumed more (specified in parentheses) was kept for the final analysis. Moreover, 40 food items were removed as their consumption was very low in the individuals' daily diets. These items were as follows: biscuits; crackers; mung beans; burgers; salami; sheep gut (such as, stomach, tongue, brain, head and feet); whipped cream; margarine; curd; pumpkin; zucchini; long green pepper; fig; green plum; grapefruit; strawberry; raisins; fresh mulberry; dried mulberry; dried plums; green olives; olive oil; animal fat; peanuts; cookies; pastry; gaz; sohaan; chocolates; crystallized sugar; halva shekari (a sweet breakfast food in Iran); homemade halva; cake; canned fruit; salt; and spices. To be able to better interpret the effect of foodstuff on the incidence of MetS, valid references that had defined daily servings were used, and each food item's parameter was transformed from \u2018grams\u2019 to specified servings.\n\nEventually, data analysis was done with STATA 11 (StataCorp, College Station, TX, USA), at a confidence interval of 95%. To account for multiple testing, we used the Bonferroni correction and considered only those crude ORs for which P < 0.05/40 = 1.25 \u00d7 10^\u22123 as significant. This correction is known to be a conservative one and thus we \u2018over-corrected\u2019 the raw P-values. To estimate the effect of the food items and other confounding variables on the incidence of MetS we used multiple logistic regression and maximum likelihood estimation methods.",
        "results": "#### Results\n\nAdults aged \ufffd20 years who had completed the FFQ were included in the study and followed for three years. Overall, among the 590 new cases of MetS (291 males and 299 females), with a median follow-up time of 24.6 months, an incidence rate of 821.62 cases per 10,000 person-years (95% CI: 757.93e890.67) was calculated. The incidence rate of MetS was 1255.88 (95% CI: 998.05e1119.56) and 652.60 (95% CI: 582.67e730.93) cases per 10,000 person-years among males and females respectively. Based on our findings, the incidence was statistically significant lower in females than in males (P 0 001) I b th d th ff t d ith M tS\n\nolder, had higher systolic and diastolic BP, higher triglycerides, higher FBS, higher BMI and greater waist circumferences. Moreover, the consumption of antihypertensive, hypoglycemic, and lipid-lowering drugs was greater among them. The percentage of married individuals and those who had experienced a heart stroke were also higher. All the aforementioned differences were statistically significant (Table 1). We estimated the adjusted and the crude odds ratio of food intakes and other covariates on MetS (Table 2). In the multiple logistic regression section, for every 5-year increase in age, the odds of acquiring the MetS decreased by 3%, though this association was not statistically significant (95% CI: 0.84e1.12). People with cancer had 5.21 times greater chance of having MetS than those who did not have cancer; this association was not statistically significant either (95% CI: 0.82e32.88). Hospitalization in the past three months increased the odds of MetS by 3.84 times, which was statistically significant (95% CI: 1.49e9.88). For every hour of job-related physical activity, the odds of MetS reduced by 31%. However, this association was not statistically significant (95% CI: 0.44e1.08). Compared to men, women were 68% less likely to acquire MetS, which was statistically significant (95% CI: 0.14e0.70; Table 2). For every unit increase in BMI, the odds of being affected with MetS increased by 1.19 times, an association that was statistically significant (95% CI: 1.12e1.27). Married individuals were 4.38 times more likely to have MetS; this too was statistically significant (95% CI: 1.17e16.29). For every 100 g consumption of pizza per day, the odds of MetS increased by 4.42 times, which was significant (95% CI: 1.01e19.54). And, for every 100 g consumption of chicken meat per day, the odds of acquiring MetS increased by 4.54 times, an association that was very significant (95% CI: 1.53e11.74). For every tablespoon consumption of mayonnaise per day the chances of having MetS increased by 6.49 times, an association that was significant (95% CI: 1.18e35.52). For every tablespoon consumption of ketchup sauce per day the chances of having MetS increased by 5.06 times, an association that was significant (95% CI: 1.08e23.71). For every 30 g consumption of groats per day the chances of having MetS increased by 7.28 times, an association that was significant too (95% CI: 1.43e36.94). Likewise, for every tablespoon consumption of honey per day the chances of having MetS increased by 2.82 times, an association that was significant (95% CI: 1.08e7.30). The consumption of Sangak bread reduced the odds of MetS by 65%, which was statistically insignificant (95% CI: 0.11e1.56). The consumption of Lavash bread reduced the odds of MetS by 78%, which was statistically significant (95% CI: 0.04e0.99). The consumption of a peach a day reduced the incidence of MetS by 89%, which was statistically significant. For everyone serving consumption of yoghurt a day (equivalent to 200 g), the incidence of MetS drops by 57%, a finding that was marginally significant (95% CI: 0.18e1.01). The odds of having MetS drops by 62% for every 50 g consumption of lettuce per day; this finding too was marginally significant (95% CI: 0.14e1.05). For every one-third serving of lentils per day (equivalent to 50 g), the odds of MetS drops by 90%, which was very significant (95% CI: 0.02e0.34). Similarly, the odds is d d b 74% f i i l b d (95%\n\nThe consumption of 20 g of mushroom per day reduces the odds of MetS by 72%, which was statistically significant (95% CI: 0.11e0.71). Consuming one tablespoon of cooking oil reduced the incidence of MetS by 77%, a relationship that was statistically significant (95% CI: 0.11e0.51). For every 50 g of ice cream eaten per day, the odds of MetS reduced by 65%, which was marginally significant (95% CI: 0.12e1.06).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "4"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "The association between nutritional exposures and metabolic syndrome in the Tehran Lipid and Glucose Study (TLGS): a cohort study.",
        "abstract": "## OBJECTIVES\nThis study was conducted with the goal of investigating the effect of various food consumption on the incidence of metabolic syndrome (MetS).\n## STUDY DESIGN\nProspective cohort study.\n## METHODS\nThis study conducted on 3616 healthy adults aged \u226520\u00a0years, who were not affected with MetS. Nutritional intake was determined at the beginning of the study (2008-2011) by the Food Frequency Questionnaire (FFQ), and the incidence of MetS was investigated after a median of 24.6 months follow-up. Data were analyzed by multiple logistic regression, with 95% confidence interval.\n## RESULTS\nAfter adjusting the effect of other variables in the model many foods in the whole grains group such as Sangak bread (OR\u00a0=\u00a00.35, 95% CI: 0.11-1.56), multiple types of vegetables and fruits - such as peach (OR\u00a0=\u00a00.11, 95% CI: 0.01-0.75), and mushroom (OR\u00a0=\u00a00.28, 95% CI: 0.11-0.71) had protective effects against MetS. From the dairy group yoghurt (OR\u00a0=\u00a00.43, 95% CI: 0.18-1.01) and ice cream (OR\u00a0=\u00a00.35, 95% CI: 0.12-1.06) had similar such effects.\n## CONCLUSION\nBased on our findings, there was a significant rise in the incidence of MetS in Iran. The daily consumption of foods from the whole grains, dairy, vegetable and fruit groups can help reduce the odds of MetS.",
        "methods": "#### Methods\n\nThis prospective study is part of a larger study titled the Tehran Lipid and Glucose Study (TLGS). It is also part of a thesis that was conducted among Tehranian adults during three years of follow-up from 2011 to 2014. In this study, we have used data from the prospective TLGS cohort, which has been conducted on 15,005 persons aged 3\u201363 years old, from Tehran's district 13. The study began in 1998. We used the data collected in the fourth (2008\u20132011) and fifth (2011\u20132014) phases of the study. Data related to dietary intake and confounding variables were collected from the fourth phase, and new MetS cases were chosen from the fifth phase, which is considered as the follow-up phase.\n\n**Target population**\n\nAll healthy adults aged \u226520 years not affected with MetS at the beginning of the study (fourth phase) who had dietary information were selected. The newly affected MetS cases were 590 persons. The reason for selecting the \u226520 years age group was that the MetS definition holds true for adults; it is different for persons aged <20 years.\n\n**Inclusion criteria**\n\nThose who were eligible for the study included persons aged \u226520 years who had been followed up from the fourth to the fifth phase, and who possessed the following criteria: no history of chronic diseases (diabetes, stroke, thyroid problems and cancer); not having any specific dietary regime (such as a weight loss diet and the intake of fewer than 800 kcal or greater than 4000 kcal); and not being affected with the MetS.\n\n**Measurement of outcome**\n\nIn the TLGS study, persons with MetS were identified using the Adult Treatment Panel (ATP) Guidelines. The criteria included having three of the following five indicators simultaneously: (1) abdominal obesity (waist circumference > 90 cm in both genders); (2) lowered serum HDL levels (lower than 40 mg/dl in men and 50 mg/dl in women, or the consumption of HDL-elevating drugs); (3) hypertension (a systolic BP \u2265 130 mmHg or a diastolic BP \u2265 85 mmHg, or the consumption of antihypertensive drugs); (4) disordered FBS level (a fasting blood glucose/FBS \u2265100 mg/dl, or the consumption of hypoglycemic drugs); (5) hypertriglyceridemia (a serum triglyceride level \u2265150 mg/dl, or the consumption of triglyceride-lowering drugs). Among the five components of MetS, four components were measured only once (at the end of each phase, i.e. once every three years), including waist circumference, serum HDL, blood sugar, and serum triglycerides. Only hypertension was measured twice during each phase. Each time, blood pressure was measured twice with at least a 30-second interval between the two measurements. Thereafter the mean of the two measurements was considered as the participant's BP for that session.\n\n**Measurement of exposure**\n\nNutritional information on the participants' dietary intake was gathered using the Food Frequency Questionnaire (FFQ), which contains 147 food items. A nutritionist who had been trained in the field completed the questionnaires through a face-to-face interview. During the interview, the average size of each of the FFQ food items was explained to the participant, and then she/he was asked about the number of times each item was taken in the past year. The validity and reliability of this questionnaire has already been assessed during a pilot study on 133 samples, wherein Pearson's correlation coefficient was acceptable. The consumption frequency of each food item in the past year was assessed in the form of days, weeks, months and/or years, and then, using home scale guides, the amount consumed of each item was transformed into grams per day. The amount of intake of energy and macronutrients and micronutrients were determined using the food composition table. The data related to food intake in the fourth phase were considered as the exposure.\n\n**Other measured variables**\n\nHere, we measured the physical activity during the past year with the standard physical activity questionnaire. The frequency and duration of each physical activity in the past 12 months was asked, and was then calculated as Metabolic Equivalent (MET) hours per week. Data on physical activity status were obtained using the modified activity questionnaire (MAC). The MAQ determines the current (past year and past week) physical activity during occupation and leisure time. Total physical activities were explicit in hours of activity per week by adding leisure time physical activity to occupational activity. Other variables measured were as follows: weight (in kilograms: wearing minimum clothes, without shoes, using a seca weighing scale \u2013 with an accuracy of 0.5 kg); height (without shoes, using a meter fixed on the wall \u2013 with an accuracy of 1 cm); waist circumference (the circumference at the slimmest area between the lowest rib and the tip of the iliac crest); systolic and diastolic blood pressure (BP) (using a manual sphygmomanometer, the mean was considered as the person's BP and reported in mm Hg); and finally, a family history of diabetes. Moreover, the necessary data on age, gender, education (primary, intermediate, high school and high school graduate, academic education), narcotics use (never smoked, previous smoked, currently smoking), and drug or supplementary intake were collected using a general information questionnaire through an interview conducted by the nutritionist. A venous blood sample was taken after 10 h nocturnal fasting to measure serum HDL, glucose, triglycerides and cholesterol. Blood glucose was measured with glucose oxidase; serum triglycerides were measured with the Pars Azmoon Kit \u2013 with the AutoAnalyzer Selectra; and serum HDL was measured after precipitating `b` apolipoprotein with phosphotungstate.\n\n**Data analysis**\n\nWe used bi-variable logistic regression to estimate the crude effect of each of the nutritional and non-nutritional variables on MetS. We then used multiple logistic regression to estimate the adjusted effect of variables. The backward-selection method was applied to choose the variables in this model. To remove the variables from the model the P-value was set at 0.2. Seventy variables (three confounding variables of educational status, smoking history and history of weight loss, in addition to 68 food items) were removed at this stage. In addition, six items were removed from the model because of their high correlation. These food items included the following: jam (and sugar); plum (and peach); lemon juice (and lemon); apple juice (and apple); orange juice (and orange); cooked vegetables (and cooked carrots). Between the two similar food items that were correlated one that was consumed more (specified in parentheses) was kept for the final analysis. Moreover, 40 food items were removed as their consumption was very low in the individuals' daily diets. These items were as follows: biscuits; crackers; mung beans; burgers; salami; sheep gut (such as, stomach, tongue, brain, head and feet); whipped cream; margarine; curd; pumpkin; zucchini; long green pepper; fig; green plum; grapefruit; strawberry; raisins; fresh mulberry; dried mulberry; dried plums; green olives; olive oil; animal fat; peanuts; cookies; pastry; gaz; sohaan; chocolates; crystallized sugar; halva shekari (a sweet breakfast food in Iran); homemade halva; cake; canned fruit; salt; and spices. To be able to better interpret the effect of foodstuff on the incidence of MetS, valid references that had defined daily servings were used, and each food item's parameter was transformed from \u2018grams\u2019 to specified servings.\n\nEventually, data analysis was done with STATA 11 (StataCorp, College Station, TX, USA), at a confidence interval of 95%. To account for multiple testing, we used the Bonferroni correction and considered only those crude ORs for which P < 0.05/40 = 1.25 \u00d7 10^\u22123 as significant. This correction is known to be a conservative one and thus we \u2018over-corrected\u2019 the raw P-values. To estimate the effect of the food items and other confounding variables on the incidence of MetS we used multiple logistic regression and maximum likelihood estimation methods.",
        "results": "#### Results\n\nAdults aged \ufffd20 years who had completed the FFQ were included in the study and followed for three years. Overall, among the 590 new cases of MetS (291 males and 299 females), with a median follow-up time of 24.6 months, an incidence rate of 821.62 cases per 10,000 person-years (95% CI: 757.93e890.67) was calculated. The incidence rate of MetS was 1255.88 (95% CI: 998.05e1119.56) and 652.60 (95% CI: 582.67e730.93) cases per 10,000 person-years among males and females respectively. Based on our findings, the incidence was statistically significant lower in females than in males (P 0 001) I b th d th ff t d ith M tS\n\nolder, had higher systolic and diastolic BP, higher triglycerides, higher FBS, higher BMI and greater waist circumferences. Moreover, the consumption of antihypertensive, hypoglycemic, and lipid-lowering drugs was greater among them. The percentage of married individuals and those who had experienced a heart stroke were also higher. All the aforementioned differences were statistically significant (Table 1). We estimated the adjusted and the crude odds ratio of food intakes and other covariates on MetS (Table 2). In the multiple logistic regression section, for every 5-year increase in age, the odds of acquiring the MetS decreased by 3%, though this association was not statistically significant (95% CI: 0.84e1.12). People with cancer had 5.21 times greater chance of having MetS than those who did not have cancer; this association was not statistically significant either (95% CI: 0.82e32.88). Hospitalization in the past three months increased the odds of MetS by 3.84 times, which was statistically significant (95% CI: 1.49e9.88). For every hour of job-related physical activity, the odds of MetS reduced by 31%. However, this association was not statistically significant (95% CI: 0.44e1.08). Compared to men, women were 68% less likely to acquire MetS, which was statistically significant (95% CI: 0.14e0.70; Table 2). For every unit increase in BMI, the odds of being affected with MetS increased by 1.19 times, an association that was statistically significant (95% CI: 1.12e1.27). Married individuals were 4.38 times more likely to have MetS; this too was statistically significant (95% CI: 1.17e16.29). For every 100 g consumption of pizza per day, the odds of MetS increased by 4.42 times, which was significant (95% CI: 1.01e19.54). And, for every 100 g consumption of chicken meat per day, the odds of acquiring MetS increased by 4.54 times, an association that was very significant (95% CI: 1.53e11.74). For every tablespoon consumption of mayonnaise per day the chances of having MetS increased by 6.49 times, an association that was significant (95% CI: 1.18e35.52). For every tablespoon consumption of ketchup sauce per day the chances of having MetS increased by 5.06 times, an association that was significant (95% CI: 1.08e23.71). For every 30 g consumption of groats per day the chances of having MetS increased by 7.28 times, an association that was significant too (95% CI: 1.43e36.94). Likewise, for every tablespoon consumption of honey per day the chances of having MetS increased by 2.82 times, an association that was significant (95% CI: 1.08e7.30). The consumption of Sangak bread reduced the odds of MetS by 65%, which was statistically insignificant (95% CI: 0.11e1.56). The consumption of Lavash bread reduced the odds of MetS by 78%, which was statistically significant (95% CI: 0.04e0.99). The consumption of a peach a day reduced the incidence of MetS by 89%, which was statistically significant. For everyone serving consumption of yoghurt a day (equivalent to 200 g), the incidence of MetS drops by 57%, a finding that was marginally significant (95% CI: 0.18e1.01). The odds of having MetS drops by 62% for every 50 g consumption of lettuce per day; this finding too was marginally significant (95% CI: 0.14e1.05). For every one-third serving of lentils per day (equivalent to 50 g), the odds of MetS drops by 90%, which was very significant (95% CI: 0.02e0.34). Similarly, the odds is d d b 74% f i i l b d (95%\n\nThe consumption of 20 g of mushroom per day reduces the odds of MetS by 72%, which was statistically significant (95% CI: 0.11e0.71). Consuming one tablespoon of cooking oil reduced the incidence of MetS by 77%, a relationship that was statistically significant (95% CI: 0.11e0.51). For every 50 g of ice cream eaten per day, the odds of MetS reduced by 65%, which was marginally significant (95% CI: 0.12e1.06).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "The association between nutritional exposures and metabolic syndrome in the Tehran Lipid and Glucose Study (TLGS): a cohort study.",
        "abstract": "## OBJECTIVES\nThis study was conducted with the goal of investigating the effect of various food consumption on the incidence of metabolic syndrome (MetS).\n## STUDY DESIGN\nProspective cohort study.\n## METHODS\nThis study conducted on 3616 healthy adults aged \u226520\u00a0years, who were not affected with MetS. Nutritional intake was determined at the beginning of the study (2008-2011) by the Food Frequency Questionnaire (FFQ), and the incidence of MetS was investigated after a median of 24.6 months follow-up. Data were analyzed by multiple logistic regression, with 95% confidence interval.\n## RESULTS\nAfter adjusting the effect of other variables in the model many foods in the whole grains group such as Sangak bread (OR\u00a0=\u00a00.35, 95% CI: 0.11-1.56), multiple types of vegetables and fruits - such as peach (OR\u00a0=\u00a00.11, 95% CI: 0.01-0.75), and mushroom (OR\u00a0=\u00a00.28, 95% CI: 0.11-0.71) had protective effects against MetS. From the dairy group yoghurt (OR\u00a0=\u00a00.43, 95% CI: 0.18-1.01) and ice cream (OR\u00a0=\u00a00.35, 95% CI: 0.12-1.06) had similar such effects.\n## CONCLUSION\nBased on our findings, there was a significant rise in the incidence of MetS in Iran. The daily consumption of foods from the whole grains, dairy, vegetable and fruit groups can help reduce the odds of MetS.",
        "methods": "#### Methods\n\nThis prospective study is part of a larger study titled the Tehran Lipid and Glucose Study (TLGS). It is also part of a thesis that was conducted among Tehranian adults during three years of follow-up from 2011 to 2014. In this study, we have used data from the prospective TLGS cohort, which has been conducted on 15,005 persons aged 3\u201363 years old, from Tehran's district 13. The study began in 1998. We used the data collected in the fourth (2008\u20132011) and fifth (2011\u20132014) phases of the study. Data related to dietary intake and confounding variables were collected from the fourth phase, and new MetS cases were chosen from the fifth phase, which is considered as the follow-up phase.\n\n**Target population**\n\nAll healthy adults aged \u226520 years not affected with MetS at the beginning of the study (fourth phase) who had dietary information were selected. The newly affected MetS cases were 590 persons. The reason for selecting the \u226520 years age group was that the MetS definition holds true for adults; it is different for persons aged <20 years.\n\n**Inclusion criteria**\n\nThose who were eligible for the study included persons aged \u226520 years who had been followed up from the fourth to the fifth phase, and who possessed the following criteria: no history of chronic diseases (diabetes, stroke, thyroid problems and cancer); not having any specific dietary regime (such as a weight loss diet and the intake of fewer than 800 kcal or greater than 4000 kcal); and not being affected with the MetS.\n\n**Measurement of outcome**\n\nIn the TLGS study, persons with MetS were identified using the Adult Treatment Panel (ATP) Guidelines. The criteria included having three of the following five indicators simultaneously: (1) abdominal obesity (waist circumference > 90 cm in both genders); (2) lowered serum HDL levels (lower than 40 mg/dl in men and 50 mg/dl in women, or the consumption of HDL-elevating drugs); (3) hypertension (a systolic BP \u2265 130 mmHg or a diastolic BP \u2265 85 mmHg, or the consumption of antihypertensive drugs); (4) disordered FBS level (a fasting blood glucose/FBS \u2265100 mg/dl, or the consumption of hypoglycemic drugs); (5) hypertriglyceridemia (a serum triglyceride level \u2265150 mg/dl, or the consumption of triglyceride-lowering drugs). Among the five components of MetS, four components were measured only once (at the end of each phase, i.e. once every three years), including waist circumference, serum HDL, blood sugar, and serum triglycerides. Only hypertension was measured twice during each phase. Each time, blood pressure was measured twice with at least a 30-second interval between the two measurements. Thereafter the mean of the two measurements was considered as the participant's BP for that session.\n\n**Measurement of exposure**\n\nNutritional information on the participants' dietary intake was gathered using the Food Frequency Questionnaire (FFQ), which contains 147 food items. A nutritionist who had been trained in the field completed the questionnaires through a face-to-face interview. During the interview, the average size of each of the FFQ food items was explained to the participant, and then she/he was asked about the number of times each item was taken in the past year. The validity and reliability of this questionnaire has already been assessed during a pilot study on 133 samples, wherein Pearson's correlation coefficient was acceptable. The consumption frequency of each food item in the past year was assessed in the form of days, weeks, months and/or years, and then, using home scale guides, the amount consumed of each item was transformed into grams per day. The amount of intake of energy and macronutrients and micronutrients were determined using the food composition table. The data related to food intake in the fourth phase were considered as the exposure.\n\n**Other measured variables**\n\nHere, we measured the physical activity during the past year with the standard physical activity questionnaire. The frequency and duration of each physical activity in the past 12 months was asked, and was then calculated as Metabolic Equivalent (MET) hours per week. Data on physical activity status were obtained using the modified activity questionnaire (MAC). The MAQ determines the current (past year and past week) physical activity during occupation and leisure time. Total physical activities were explicit in hours of activity per week by adding leisure time physical activity to occupational activity. Other variables measured were as follows: weight (in kilograms: wearing minimum clothes, without shoes, using a seca weighing scale \u2013 with an accuracy of 0.5 kg); height (without shoes, using a meter fixed on the wall \u2013 with an accuracy of 1 cm); waist circumference (the circumference at the slimmest area between the lowest rib and the tip of the iliac crest); systolic and diastolic blood pressure (BP) (using a manual sphygmomanometer, the mean was considered as the person's BP and reported in mm Hg); and finally, a family history of diabetes. Moreover, the necessary data on age, gender, education (primary, intermediate, high school and high school graduate, academic education), narcotics use (never smoked, previous smoked, currently smoking), and drug or supplementary intake were collected using a general information questionnaire through an interview conducted by the nutritionist. A venous blood sample was taken after 10 h nocturnal fasting to measure serum HDL, glucose, triglycerides and cholesterol. Blood glucose was measured with glucose oxidase; serum triglycerides were measured with the Pars Azmoon Kit \u2013 with the AutoAnalyzer Selectra; and serum HDL was measured after precipitating `b` apolipoprotein with phosphotungstate.\n\n**Data analysis**\n\nWe used bi-variable logistic regression to estimate the crude effect of each of the nutritional and non-nutritional variables on MetS. We then used multiple logistic regression to estimate the adjusted effect of variables. The backward-selection method was applied to choose the variables in this model. To remove the variables from the model the P-value was set at 0.2. Seventy variables (three confounding variables of educational status, smoking history and history of weight loss, in addition to 68 food items) were removed at this stage. In addition, six items were removed from the model because of their high correlation. These food items included the following: jam (and sugar); plum (and peach); lemon juice (and lemon); apple juice (and apple); orange juice (and orange); cooked vegetables (and cooked carrots). Between the two similar food items that were correlated one that was consumed more (specified in parentheses) was kept for the final analysis. Moreover, 40 food items were removed as their consumption was very low in the individuals' daily diets. These items were as follows: biscuits; crackers; mung beans; burgers; salami; sheep gut (such as, stomach, tongue, brain, head and feet); whipped cream; margarine; curd; pumpkin; zucchini; long green pepper; fig; green plum; grapefruit; strawberry; raisins; fresh mulberry; dried mulberry; dried plums; green olives; olive oil; animal fat; peanuts; cookies; pastry; gaz; sohaan; chocolates; crystallized sugar; halva shekari (a sweet breakfast food in Iran); homemade halva; cake; canned fruit; salt; and spices. To be able to better interpret the effect of foodstuff on the incidence of MetS, valid references that had defined daily servings were used, and each food item's parameter was transformed from \u2018grams\u2019 to specified servings.\n\nEventually, data analysis was done with STATA 11 (StataCorp, College Station, TX, USA), at a confidence interval of 95%. To account for multiple testing, we used the Bonferroni correction and considered only those crude ORs for which P < 0.05/40 = 1.25 \u00d7 10^\u22123 as significant. This correction is known to be a conservative one and thus we \u2018over-corrected\u2019 the raw P-values. To estimate the effect of the food items and other confounding variables on the incidence of MetS we used multiple logistic regression and maximum likelihood estimation methods.",
        "results": "#### Results\n\nAdults aged \ufffd20 years who had completed the FFQ were included in the study and followed for three years. Overall, among the 590 new cases of MetS (291 males and 299 females), with a median follow-up time of 24.6 months, an incidence rate of 821.62 cases per 10,000 person-years (95% CI: 757.93e890.67) was calculated. The incidence rate of MetS was 1255.88 (95% CI: 998.05e1119.56) and 652.60 (95% CI: 582.67e730.93) cases per 10,000 person-years among males and females respectively. Based on our findings, the incidence was statistically significant lower in females than in males (P 0 001) I b th d th ff t d ith M tS\n\nolder, had higher systolic and diastolic BP, higher triglycerides, higher FBS, higher BMI and greater waist circumferences. Moreover, the consumption of antihypertensive, hypoglycemic, and lipid-lowering drugs was greater among them. The percentage of married individuals and those who had experienced a heart stroke were also higher. All the aforementioned differences were statistically significant (Table 1). We estimated the adjusted and the crude odds ratio of food intakes and other covariates on MetS (Table 2). In the multiple logistic regression section, for every 5-year increase in age, the odds of acquiring the MetS decreased by 3%, though this association was not statistically significant (95% CI: 0.84e1.12). People with cancer had 5.21 times greater chance of having MetS than those who did not have cancer; this association was not statistically significant either (95% CI: 0.82e32.88). Hospitalization in the past three months increased the odds of MetS by 3.84 times, which was statistically significant (95% CI: 1.49e9.88). For every hour of job-related physical activity, the odds of MetS reduced by 31%. However, this association was not statistically significant (95% CI: 0.44e1.08). Compared to men, women were 68% less likely to acquire MetS, which was statistically significant (95% CI: 0.14e0.70; Table 2). For every unit increase in BMI, the odds of being affected with MetS increased by 1.19 times, an association that was statistically significant (95% CI: 1.12e1.27). Married individuals were 4.38 times more likely to have MetS; this too was statistically significant (95% CI: 1.17e16.29). For every 100 g consumption of pizza per day, the odds of MetS increased by 4.42 times, which was significant (95% CI: 1.01e19.54). And, for every 100 g consumption of chicken meat per day, the odds of acquiring MetS increased by 4.54 times, an association that was very significant (95% CI: 1.53e11.74). For every tablespoon consumption of mayonnaise per day the chances of having MetS increased by 6.49 times, an association that was significant (95% CI: 1.18e35.52). For every tablespoon consumption of ketchup sauce per day the chances of having MetS increased by 5.06 times, an association that was significant (95% CI: 1.08e23.71). For every 30 g consumption of groats per day the chances of having MetS increased by 7.28 times, an association that was significant too (95% CI: 1.43e36.94). Likewise, for every tablespoon consumption of honey per day the chances of having MetS increased by 2.82 times, an association that was significant (95% CI: 1.08e7.30). The consumption of Sangak bread reduced the odds of MetS by 65%, which was statistically insignificant (95% CI: 0.11e1.56). The consumption of Lavash bread reduced the odds of MetS by 78%, which was statistically significant (95% CI: 0.04e0.99). The consumption of a peach a day reduced the incidence of MetS by 89%, which was statistically significant. For everyone serving consumption of yoghurt a day (equivalent to 200 g), the incidence of MetS drops by 57%, a finding that was marginally significant (95% CI: 0.18e1.01). The odds of having MetS drops by 62% for every 50 g consumption of lettuce per day; this finding too was marginally significant (95% CI: 0.14e1.05). For every one-third serving of lentils per day (equivalent to 50 g), the odds of MetS drops by 90%, which was very significant (95% CI: 0.02e0.34). Similarly, the odds is d d b 74% f i i l b d (95%\n\nThe consumption of 20 g of mushroom per day reduces the odds of MetS by 72%, which was statistically significant (95% CI: 0.11e0.71). Consuming one tablespoon of cooking oil reduced the incidence of MetS by 77%, a relationship that was statistically significant (95% CI: 0.11e0.51). For every 50 g of ice cream eaten per day, the odds of MetS reduced by 65%, which was marginally significant (95% CI: 0.12e1.06).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "0"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Drinking caloric beverages increases the risk of adverse cardiometabolic outcomes in the Coronary Artery Risk Development in Young Adults (CARDIA) Study.",
        "abstract": "## BACKGROUND\nIntake of caloric beverages is hypothesized to contribute to adverse health outcomes, but the beverages and populations studied vary considerably.\n## OBJECTIVE\nOur objective was to examine the relation between consumption of low- and whole-fat milk, fruit juice, and sugar-sweetened beverages (SSBs) and cardiometabolic risk factors.\n## DESIGN\nWe used data from a prospective 20-y cohort of 2774 adults. Data are taken from CARDIA (Coronary Artery Risk Development in Young Adults) Study examination years 0 (1985-1986), 7 (1992-1993), and 20 (2005-2006). Beverage intake was averaged across years 0 and 7, and continuous and categorical (quartile) distributions were used. Incident (year 20) high waist circumference (WC), high triglycerides, high LDL cholesterol, low HDL cholesterol, hypertension, and metabolic syndrome were examined by using multivariable-adjusted Poisson regression models.\n## RESULTS\nHigher SSB consumption (across quartiles) was associated with higher risk of high WC [adjusted relative risk (aRR): 1.09; 95% CI: 1.04, 1.14; P for trend < 0.001]; high LDL cholesterol (aRR: 1.18; 95% CI: 1.02, 1.35; P for trend = 0.018), high triglycerides (aRR: 1.06; 95% CI: 1.01, 1.13; P for trend = 0.033), and hypertension (aRR: 1.06; 95% CI: 1.01, 1.12; P for trend = 0.023). Whole-fat milk consumption was associated with lower risk of high triglycerides (aRR: 0.91; 95% CI: 0.81, 1.00; P for trend = 0.046). With the use of continuous beverage intake, results were similar. Consumers of whole-fat milk and SSBs were more likely to be younger, black, and male and to have lower levels of physical activity and higher total energy intake in comparison with nonconsumers (P < 0.05).\n## CONCLUSIONS\nOur findings suggest that higher SSB consumption is associated with cardiometabolic risk. Recommendations to limit consumption of these caloric beverages may help reduce the burden of these risk factors in US adult populations.",
        "methods": "### METHODS\n\n**Study sample**\n\nThe Coronary Artery Risk Development in Young Adults (CARDIA) Study is a prospective study of cardiovascular risk factors in 5115 persons aged 18\u201330 y at baseline examination (1985\u20131986). Detailed description of recruitment procedures are provided elsewhere (10). We used data from exam years 0 (1985\u20131986, baseline), 7 (1992\u20131993), and 20 (2005\u20132006) in which dietary intake data were collected. We excluded women who were pregnant at the time of interview (n = 75). Analysis procedures were in accordance with the ethical standards of the University of North Carolina at Chapel Hill (UNC-CH), Gillings School of Global Public Health, with UNC-CH Institutional Review Board approval.\n\n**Exposure and outcome measures**\n\n**Beverage intake**\n\nDietary intake was assessed by using a semiquantitative, interviewer-administered, validated (11) diet history food-frequency questionnaire, which consisted of a short questionnaire on general dietary practices followed by a comprehensive food-frequency questionnaire, which asked about the previous month\u2019s consumption. We focused on low-fat (skim and \u22642% fat) and whole-fat (\u22653% fat) milk (consumed as a beverage, not used in recipes), fruit juice, and sugar-sweetened soda and fruit drinks (SSBs) as grouped by the CARDIA diet coordinating center. For each beverage, baseline intake was defined as the average intake from exam year 0 and 7 among persons with complete data. Because diet soda does not provide calories, we do not examine patterns of its consumption in this article. Furthermore, there is a strong bimodal distribution of alcohol intake, and high, long-term consumption represents a complex behavioral issue that is outside the purview of this study. We examined 3 consumption trends: energy per capita (estimates that account for persons who did not consume the beverage), percentage consuming and energy per consumer (estimates that apply only to those individuals who reported consuming the beverage).\n\n**Anthropometric measurements, blood lipids, and metabolic syndrome**\n\nAnthropometric measurements and blood lipid samples were taken at each exam year (0, 7, and 20). Waist circumference (WC) was measured as the average of 2 measures at the minimum abdominal girth (nearest 0.5 cm) from participants standing upright. High WC was defined as WC >88 cm (women) or >102 cm (men). Seated blood pressure [millimeters of mercury (mm Hg)] was measured 3 times; we used the average of the last 2 measurements. We excluded individuals who fasted for <8 h. Hypertension was defined as systolic blood pressure (SBP) \u2265130 mm Hg, diastolic blood pressure (DBP) \u226585 mm Hg, or use of antihypertensive medication. Fasting blood samples were collected from which assays determined lipids, glucose, and insulin according to standardized CARDIA procedures (10). Briefly, fasting insulin and glucose were obtained by venous blood draw, with glucose measured by using hexokinase coupled to glucose-6-phosphate dehydrogenase. We created the following measures (which identify adverse incident year 20 health outcomes) among persons without prevalent outcome at year 0 or 7: high fasting glucose \u22656.1 mmol/L or use of diabetic medication; low HDL cholesterol, defined as <1.3 mmol/L in women or <1.03 mmol/L in men or use of cholesterol-lowering medication; high LDL cholesterol, defined as \u22654.1 mmol/L or use of cholesterol-lowering medication; and high triglyceride concentrations, defined as \u22651.7 mmol/L or use of cholesterol-lowering medication.\n\nMetabolic syndrome was defined according to Adult Treatment Panel III guidelines (12) as being present if \u22653 of the following factors were present at year 20 among those who did not have metabolic syndrome at year 0 or 7: abdominal obesity, high fasting glucose, low HDL cholesterol, hypertension, and/or high triglycerides.\n\n**Covariates**\n\nSociodemographic and selected health behavior covariates included the following: race and sex, smoking status (classified as never, former, or current), and highest education completed (classified as less than high school, completed high school, or more than high school). Physical activity was assessed by using the CARDIA physical activity questionnaire, a validated and reliable assessment of physical activity (13). A total score [exercise units (EU) per week] was computed by multiplying the frequency of participation by the intensity of the activity. As an example, a score of 100 EU is roughly equivalent to participation in a vigorous physical activity 2 or 3 h/wk for 6 mo of the year, calculated as [6 metabolic equivalent tasks (METs) \u00b7 (3 \u00b7 6 mo of high volume activity)]. A set of food groups measuring energy per day from meats, vegetables (except fried potatoes), fruit (except fried fruit), and nonmilk dairy products was created from the CARDIA diet history. Energy from food was calculated by subtracting all beverage energy from total energy intake.\n\n**Statistical analysis**\n\nAll analyses were conducted by using STATA (version 11.0; Stata Corp, College Station, TX). We examined differences in baseline sociodemographic covariates between consumers and nonconsumers of whole-fat milk and SSBs by using 2-sided t tests and chi-square tests for continuous and categorical variables, respectively. Within each of the 4 beverage groups, 2-sided t tests (per capita and per consumer) and chi-square tests (percentage consuming) were also used to determine statistical differences between years (with Bonferroni correction for multiple comparisons) for per capita/per consumer and percentage consuming estimates, respectively. To estimate the risk of incident year 20 outcomes, we used separate multivariate Poisson regression models (n = 7 models) of incident high WC (n = 2444), high fasting glucose (n = 2160), high triglycerides (n = 2627), high LDL cholesterol (n = 2640), low HDL cholesterol (n = 1837), hypertension (n = 2639), and metabolic syndrome (n = 3596) on the 4 beverages of interest (SSBs, fruit juice, low-fat milk, and whole-fat milk). All models used beverages as continuous and divided in quartiles, which allowed for tests of trend. Categorical results are presented here, with continuous results shown under \u201cSupplemental data\u201d in the online issue. We used robust estimators that allow the variance of the error term to be dependent on our independent variables (energy from beverages). Year 20 incidence models (ie, incident high WC) were run by using individuals without prevalent outcome (ie, high WC) at exam year 0 or 7. Models were adjusted for race and sex as well as year 0 age, weight, smoking status, total physical activity, calories from food, calories from the other beverages, calories from alcohol, and CARDIA exam center. Inclusion of a set of food groups (baseline servings per day of meats, vegetables, fruit, and nonmilk dairy products) as opposed to total food calories did not affect results so we included only the latter in the interest of building the most parsimonious model possible (ie, the one with the fewest variables). Interactions with baseline overweight and race were tested in multivariable linear models but were not statistically significant (likelihood ratio test, P > 0.05). P value < 0.05 was set for determining statistical significance in our continuous models.",
        "results": "### RESULTS\nBaseline (year 0) demographics are presented in Table 1. Consumers of whole-fat milk and SSBs were significantly more likely to be younger, black, male, and current smokers and to have lower levels (EU/wk) of physical activity, higher total daily energy intake, and higher SBP compared with nonconsumers (P, 0.05). Consumers of SSBs were also more likely to have lower levels of HDL cholesterol and higher diastolic blood pressure compared with nonconsumers (Table 1). There were no differences in any other anthropometric or blood lipid measure between consumers and nonconsumers. Differences between low-fat milk and fruit Juice consumers and nonconsumers are available on request from the corresponding author. Over 20 y, we observed a consistent increase in per capita calories for SSBs, adjusted for age and sex, whereas per capita intake of whole-fat and low-fat milk and fruit juice declined (Table 2). There was a slight (7%) increase between year 0 and year 7 followed by a considerable (14%) decrease between year 7 and year 20 in the percentage of the sample who consumed low-fat milk. The percentage of the population consuming SSBs steadily increased over the full time period, although only by a small percentage, whereas the percentage consuming fruit juice and whole-fat milk continuously declined. Energy per consumer from low- and whole-fat milk experienced consistent declines, whereas energy per consumer from SSBs increased steadily over the 20 y (Table 2). The most consistent adverse associations between beverage intake and incident cardiometabolic outcomes were observed for SSBs, in which higher (moving from one quartile to the next) baseline consumption was associated with a significant increase in the risk of incident high WC [adjusted relative risk (aRR): 1.09; 95% CI: 1.04, 1.14; P for trend, 0.001], high triglycerides (aRR: 1.06; 95% CI: 1.01, 1.13; P for trend = 0.033), high LDL cholesterol (aRR: 1.18; 95% CI: 1.02,1.36; P for trend = 0.018), and hypertension (aRR: 1.06; 95% CI: 1.01, 1.12; P for trend = 0.023; Table 3). Associations for SSB intake were in the same direction (associated with higher risk) for all other outcomes, although these estimates were not statistically significant (Table 3). Moving from one quartile to the next of whole-fat milk and fruit juice consumption was associated with lower risks of incident high triglycerides (aRR: 0.91; 95% CI: 0.81, 1.00; P for trend = 0.046) and hypertension (aRR: 0.89; 95% CI: 0.82, 0.97; P for trend= 0.007), respectively. Consumption of low-fat milk was not statistically significantly related to any incident cardiometabolic risk factor or to the metabolic syndrome (Table 3). Results using continuous beverage consumption were similar with the major difference that SSB intake was significantly associated with incident low HDL cholesterol (aRR: 1.06; 95% CI: 1.02, 1.10; P, 0.05) and not hypertension or incident high triglycerides, and fruit juice intake was not significantly associated with any incident cardiometabolic outcome (see the supplemental table under \u201cSupplemental data\u201d in the online issue).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "4"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Drinking caloric beverages increases the risk of adverse cardiometabolic outcomes in the Coronary Artery Risk Development in Young Adults (CARDIA) Study.",
        "abstract": "## BACKGROUND\nIntake of caloric beverages is hypothesized to contribute to adverse health outcomes, but the beverages and populations studied vary considerably.\n## OBJECTIVE\nOur objective was to examine the relation between consumption of low- and whole-fat milk, fruit juice, and sugar-sweetened beverages (SSBs) and cardiometabolic risk factors.\n## DESIGN\nWe used data from a prospective 20-y cohort of 2774 adults. Data are taken from CARDIA (Coronary Artery Risk Development in Young Adults) Study examination years 0 (1985-1986), 7 (1992-1993), and 20 (2005-2006). Beverage intake was averaged across years 0 and 7, and continuous and categorical (quartile) distributions were used. Incident (year 20) high waist circumference (WC), high triglycerides, high LDL cholesterol, low HDL cholesterol, hypertension, and metabolic syndrome were examined by using multivariable-adjusted Poisson regression models.\n## RESULTS\nHigher SSB consumption (across quartiles) was associated with higher risk of high WC [adjusted relative risk (aRR): 1.09; 95% CI: 1.04, 1.14; P for trend < 0.001]; high LDL cholesterol (aRR: 1.18; 95% CI: 1.02, 1.35; P for trend = 0.018), high triglycerides (aRR: 1.06; 95% CI: 1.01, 1.13; P for trend = 0.033), and hypertension (aRR: 1.06; 95% CI: 1.01, 1.12; P for trend = 0.023). Whole-fat milk consumption was associated with lower risk of high triglycerides (aRR: 0.91; 95% CI: 0.81, 1.00; P for trend = 0.046). With the use of continuous beverage intake, results were similar. Consumers of whole-fat milk and SSBs were more likely to be younger, black, and male and to have lower levels of physical activity and higher total energy intake in comparison with nonconsumers (P < 0.05).\n## CONCLUSIONS\nOur findings suggest that higher SSB consumption is associated with cardiometabolic risk. Recommendations to limit consumption of these caloric beverages may help reduce the burden of these risk factors in US adult populations.",
        "methods": "### METHODS\n\n**Study sample**\n\nThe Coronary Artery Risk Development in Young Adults (CARDIA) Study is a prospective study of cardiovascular risk factors in 5115 persons aged 18\u201330 y at baseline examination (1985\u20131986). Detailed description of recruitment procedures are provided elsewhere (10). We used data from exam years 0 (1985\u20131986, baseline), 7 (1992\u20131993), and 20 (2005\u20132006) in which dietary intake data were collected. We excluded women who were pregnant at the time of interview (n = 75). Analysis procedures were in accordance with the ethical standards of the University of North Carolina at Chapel Hill (UNC-CH), Gillings School of Global Public Health, with UNC-CH Institutional Review Board approval.\n\n**Exposure and outcome measures**\n\n**Beverage intake**\n\nDietary intake was assessed by using a semiquantitative, interviewer-administered, validated (11) diet history food-frequency questionnaire, which consisted of a short questionnaire on general dietary practices followed by a comprehensive food-frequency questionnaire, which asked about the previous month\u2019s consumption. We focused on low-fat (skim and \u22642% fat) and whole-fat (\u22653% fat) milk (consumed as a beverage, not used in recipes), fruit juice, and sugar-sweetened soda and fruit drinks (SSBs) as grouped by the CARDIA diet coordinating center. For each beverage, baseline intake was defined as the average intake from exam year 0 and 7 among persons with complete data. Because diet soda does not provide calories, we do not examine patterns of its consumption in this article. Furthermore, there is a strong bimodal distribution of alcohol intake, and high, long-term consumption represents a complex behavioral issue that is outside the purview of this study. We examined 3 consumption trends: energy per capita (estimates that account for persons who did not consume the beverage), percentage consuming and energy per consumer (estimates that apply only to those individuals who reported consuming the beverage).\n\n**Anthropometric measurements, blood lipids, and metabolic syndrome**\n\nAnthropometric measurements and blood lipid samples were taken at each exam year (0, 7, and 20). Waist circumference (WC) was measured as the average of 2 measures at the minimum abdominal girth (nearest 0.5 cm) from participants standing upright. High WC was defined as WC >88 cm (women) or >102 cm (men). Seated blood pressure [millimeters of mercury (mm Hg)] was measured 3 times; we used the average of the last 2 measurements. We excluded individuals who fasted for <8 h. Hypertension was defined as systolic blood pressure (SBP) \u2265130 mm Hg, diastolic blood pressure (DBP) \u226585 mm Hg, or use of antihypertensive medication. Fasting blood samples were collected from which assays determined lipids, glucose, and insulin according to standardized CARDIA procedures (10). Briefly, fasting insulin and glucose were obtained by venous blood draw, with glucose measured by using hexokinase coupled to glucose-6-phosphate dehydrogenase. We created the following measures (which identify adverse incident year 20 health outcomes) among persons without prevalent outcome at year 0 or 7: high fasting glucose \u22656.1 mmol/L or use of diabetic medication; low HDL cholesterol, defined as <1.3 mmol/L in women or <1.03 mmol/L in men or use of cholesterol-lowering medication; high LDL cholesterol, defined as \u22654.1 mmol/L or use of cholesterol-lowering medication; and high triglyceride concentrations, defined as \u22651.7 mmol/L or use of cholesterol-lowering medication.\n\nMetabolic syndrome was defined according to Adult Treatment Panel III guidelines (12) as being present if \u22653 of the following factors were present at year 20 among those who did not have metabolic syndrome at year 0 or 7: abdominal obesity, high fasting glucose, low HDL cholesterol, hypertension, and/or high triglycerides.\n\n**Covariates**\n\nSociodemographic and selected health behavior covariates included the following: race and sex, smoking status (classified as never, former, or current), and highest education completed (classified as less than high school, completed high school, or more than high school). Physical activity was assessed by using the CARDIA physical activity questionnaire, a validated and reliable assessment of physical activity (13). A total score [exercise units (EU) per week] was computed by multiplying the frequency of participation by the intensity of the activity. As an example, a score of 100 EU is roughly equivalent to participation in a vigorous physical activity 2 or 3 h/wk for 6 mo of the year, calculated as [6 metabolic equivalent tasks (METs) \u00b7 (3 \u00b7 6 mo of high volume activity)]. A set of food groups measuring energy per day from meats, vegetables (except fried potatoes), fruit (except fried fruit), and nonmilk dairy products was created from the CARDIA diet history. Energy from food was calculated by subtracting all beverage energy from total energy intake.\n\n**Statistical analysis**\n\nAll analyses were conducted by using STATA (version 11.0; Stata Corp, College Station, TX). We examined differences in baseline sociodemographic covariates between consumers and nonconsumers of whole-fat milk and SSBs by using 2-sided t tests and chi-square tests for continuous and categorical variables, respectively. Within each of the 4 beverage groups, 2-sided t tests (per capita and per consumer) and chi-square tests (percentage consuming) were also used to determine statistical differences between years (with Bonferroni correction for multiple comparisons) for per capita/per consumer and percentage consuming estimates, respectively. To estimate the risk of incident year 20 outcomes, we used separate multivariate Poisson regression models (n = 7 models) of incident high WC (n = 2444), high fasting glucose (n = 2160), high triglycerides (n = 2627), high LDL cholesterol (n = 2640), low HDL cholesterol (n = 1837), hypertension (n = 2639), and metabolic syndrome (n = 3596) on the 4 beverages of interest (SSBs, fruit juice, low-fat milk, and whole-fat milk). All models used beverages as continuous and divided in quartiles, which allowed for tests of trend. Categorical results are presented here, with continuous results shown under \u201cSupplemental data\u201d in the online issue. We used robust estimators that allow the variance of the error term to be dependent on our independent variables (energy from beverages). Year 20 incidence models (ie, incident high WC) were run by using individuals without prevalent outcome (ie, high WC) at exam year 0 or 7. Models were adjusted for race and sex as well as year 0 age, weight, smoking status, total physical activity, calories from food, calories from the other beverages, calories from alcohol, and CARDIA exam center. Inclusion of a set of food groups (baseline servings per day of meats, vegetables, fruit, and nonmilk dairy products) as opposed to total food calories did not affect results so we included only the latter in the interest of building the most parsimonious model possible (ie, the one with the fewest variables). Interactions with baseline overweight and race were tested in multivariable linear models but were not statistically significant (likelihood ratio test, P > 0.05). P value < 0.05 was set for determining statistical significance in our continuous models.",
        "results": "### RESULTS\nBaseline (year 0) demographics are presented in Table 1. Consumers of whole-fat milk and SSBs were significantly more likely to be younger, black, male, and current smokers and to have lower levels (EU/wk) of physical activity, higher total daily energy intake, and higher SBP compared with nonconsumers (P, 0.05). Consumers of SSBs were also more likely to have lower levels of HDL cholesterol and higher diastolic blood pressure compared with nonconsumers (Table 1). There were no differences in any other anthropometric or blood lipid measure between consumers and nonconsumers. Differences between low-fat milk and fruit Juice consumers and nonconsumers are available on request from the corresponding author. Over 20 y, we observed a consistent increase in per capita calories for SSBs, adjusted for age and sex, whereas per capita intake of whole-fat and low-fat milk and fruit juice declined (Table 2). There was a slight (7%) increase between year 0 and year 7 followed by a considerable (14%) decrease between year 7 and year 20 in the percentage of the sample who consumed low-fat milk. The percentage of the population consuming SSBs steadily increased over the full time period, although only by a small percentage, whereas the percentage consuming fruit juice and whole-fat milk continuously declined. Energy per consumer from low- and whole-fat milk experienced consistent declines, whereas energy per consumer from SSBs increased steadily over the 20 y (Table 2). The most consistent adverse associations between beverage intake and incident cardiometabolic outcomes were observed for SSBs, in which higher (moving from one quartile to the next) baseline consumption was associated with a significant increase in the risk of incident high WC [adjusted relative risk (aRR): 1.09; 95% CI: 1.04, 1.14; P for trend, 0.001], high triglycerides (aRR: 1.06; 95% CI: 1.01, 1.13; P for trend = 0.033), high LDL cholesterol (aRR: 1.18; 95% CI: 1.02,1.36; P for trend = 0.018), and hypertension (aRR: 1.06; 95% CI: 1.01, 1.12; P for trend = 0.023; Table 3). Associations for SSB intake were in the same direction (associated with higher risk) for all other outcomes, although these estimates were not statistically significant (Table 3). Moving from one quartile to the next of whole-fat milk and fruit juice consumption was associated with lower risks of incident high triglycerides (aRR: 0.91; 95% CI: 0.81, 1.00; P for trend = 0.046) and hypertension (aRR: 0.89; 95% CI: 0.82, 0.97; P for trend= 0.007), respectively. Consumption of low-fat milk was not statistically significantly related to any incident cardiometabolic risk factor or to the metabolic syndrome (Table 3). Results using continuous beverage consumption were similar with the major difference that SSB intake was significantly associated with incident low HDL cholesterol (aRR: 1.06; 95% CI: 1.02, 1.10; P, 0.05) and not hypertension or incident high triglycerides, and fruit juice intake was not significantly associated with any incident cardiometabolic outcome (see the supplemental table under \u201cSupplemental data\u201d in the online issue).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Drinking caloric beverages increases the risk of adverse cardiometabolic outcomes in the Coronary Artery Risk Development in Young Adults (CARDIA) Study.",
        "abstract": "## BACKGROUND\nIntake of caloric beverages is hypothesized to contribute to adverse health outcomes, but the beverages and populations studied vary considerably.\n## OBJECTIVE\nOur objective was to examine the relation between consumption of low- and whole-fat milk, fruit juice, and sugar-sweetened beverages (SSBs) and cardiometabolic risk factors.\n## DESIGN\nWe used data from a prospective 20-y cohort of 2774 adults. Data are taken from CARDIA (Coronary Artery Risk Development in Young Adults) Study examination years 0 (1985-1986), 7 (1992-1993), and 20 (2005-2006). Beverage intake was averaged across years 0 and 7, and continuous and categorical (quartile) distributions were used. Incident (year 20) high waist circumference (WC), high triglycerides, high LDL cholesterol, low HDL cholesterol, hypertension, and metabolic syndrome were examined by using multivariable-adjusted Poisson regression models.\n## RESULTS\nHigher SSB consumption (across quartiles) was associated with higher risk of high WC [adjusted relative risk (aRR): 1.09; 95% CI: 1.04, 1.14; P for trend < 0.001]; high LDL cholesterol (aRR: 1.18; 95% CI: 1.02, 1.35; P for trend = 0.018), high triglycerides (aRR: 1.06; 95% CI: 1.01, 1.13; P for trend = 0.033), and hypertension (aRR: 1.06; 95% CI: 1.01, 1.12; P for trend = 0.023). Whole-fat milk consumption was associated with lower risk of high triglycerides (aRR: 0.91; 95% CI: 0.81, 1.00; P for trend = 0.046). With the use of continuous beverage intake, results were similar. Consumers of whole-fat milk and SSBs were more likely to be younger, black, and male and to have lower levels of physical activity and higher total energy intake in comparison with nonconsumers (P < 0.05).\n## CONCLUSIONS\nOur findings suggest that higher SSB consumption is associated with cardiometabolic risk. Recommendations to limit consumption of these caloric beverages may help reduce the burden of these risk factors in US adult populations.",
        "methods": "### METHODS\n\n**Study sample**\n\nThe Coronary Artery Risk Development in Young Adults (CARDIA) Study is a prospective study of cardiovascular risk factors in 5115 persons aged 18\u201330 y at baseline examination (1985\u20131986). Detailed description of recruitment procedures are provided elsewhere (10). We used data from exam years 0 (1985\u20131986, baseline), 7 (1992\u20131993), and 20 (2005\u20132006) in which dietary intake data were collected. We excluded women who were pregnant at the time of interview (n = 75). Analysis procedures were in accordance with the ethical standards of the University of North Carolina at Chapel Hill (UNC-CH), Gillings School of Global Public Health, with UNC-CH Institutional Review Board approval.\n\n**Exposure and outcome measures**\n\n**Beverage intake**\n\nDietary intake was assessed by using a semiquantitative, interviewer-administered, validated (11) diet history food-frequency questionnaire, which consisted of a short questionnaire on general dietary practices followed by a comprehensive food-frequency questionnaire, which asked about the previous month\u2019s consumption. We focused on low-fat (skim and \u22642% fat) and whole-fat (\u22653% fat) milk (consumed as a beverage, not used in recipes), fruit juice, and sugar-sweetened soda and fruit drinks (SSBs) as grouped by the CARDIA diet coordinating center. For each beverage, baseline intake was defined as the average intake from exam year 0 and 7 among persons with complete data. Because diet soda does not provide calories, we do not examine patterns of its consumption in this article. Furthermore, there is a strong bimodal distribution of alcohol intake, and high, long-term consumption represents a complex behavioral issue that is outside the purview of this study. We examined 3 consumption trends: energy per capita (estimates that account for persons who did not consume the beverage), percentage consuming and energy per consumer (estimates that apply only to those individuals who reported consuming the beverage).\n\n**Anthropometric measurements, blood lipids, and metabolic syndrome**\n\nAnthropometric measurements and blood lipid samples were taken at each exam year (0, 7, and 20). Waist circumference (WC) was measured as the average of 2 measures at the minimum abdominal girth (nearest 0.5 cm) from participants standing upright. High WC was defined as WC >88 cm (women) or >102 cm (men). Seated blood pressure [millimeters of mercury (mm Hg)] was measured 3 times; we used the average of the last 2 measurements. We excluded individuals who fasted for <8 h. Hypertension was defined as systolic blood pressure (SBP) \u2265130 mm Hg, diastolic blood pressure (DBP) \u226585 mm Hg, or use of antihypertensive medication. Fasting blood samples were collected from which assays determined lipids, glucose, and insulin according to standardized CARDIA procedures (10). Briefly, fasting insulin and glucose were obtained by venous blood draw, with glucose measured by using hexokinase coupled to glucose-6-phosphate dehydrogenase. We created the following measures (which identify adverse incident year 20 health outcomes) among persons without prevalent outcome at year 0 or 7: high fasting glucose \u22656.1 mmol/L or use of diabetic medication; low HDL cholesterol, defined as <1.3 mmol/L in women or <1.03 mmol/L in men or use of cholesterol-lowering medication; high LDL cholesterol, defined as \u22654.1 mmol/L or use of cholesterol-lowering medication; and high triglyceride concentrations, defined as \u22651.7 mmol/L or use of cholesterol-lowering medication.\n\nMetabolic syndrome was defined according to Adult Treatment Panel III guidelines (12) as being present if \u22653 of the following factors were present at year 20 among those who did not have metabolic syndrome at year 0 or 7: abdominal obesity, high fasting glucose, low HDL cholesterol, hypertension, and/or high triglycerides.\n\n**Covariates**\n\nSociodemographic and selected health behavior covariates included the following: race and sex, smoking status (classified as never, former, or current), and highest education completed (classified as less than high school, completed high school, or more than high school). Physical activity was assessed by using the CARDIA physical activity questionnaire, a validated and reliable assessment of physical activity (13). A total score [exercise units (EU) per week] was computed by multiplying the frequency of participation by the intensity of the activity. As an example, a score of 100 EU is roughly equivalent to participation in a vigorous physical activity 2 or 3 h/wk for 6 mo of the year, calculated as [6 metabolic equivalent tasks (METs) \u00b7 (3 \u00b7 6 mo of high volume activity)]. A set of food groups measuring energy per day from meats, vegetables (except fried potatoes), fruit (except fried fruit), and nonmilk dairy products was created from the CARDIA diet history. Energy from food was calculated by subtracting all beverage energy from total energy intake.\n\n**Statistical analysis**\n\nAll analyses were conducted by using STATA (version 11.0; Stata Corp, College Station, TX). We examined differences in baseline sociodemographic covariates between consumers and nonconsumers of whole-fat milk and SSBs by using 2-sided t tests and chi-square tests for continuous and categorical variables, respectively. Within each of the 4 beverage groups, 2-sided t tests (per capita and per consumer) and chi-square tests (percentage consuming) were also used to determine statistical differences between years (with Bonferroni correction for multiple comparisons) for per capita/per consumer and percentage consuming estimates, respectively. To estimate the risk of incident year 20 outcomes, we used separate multivariate Poisson regression models (n = 7 models) of incident high WC (n = 2444), high fasting glucose (n = 2160), high triglycerides (n = 2627), high LDL cholesterol (n = 2640), low HDL cholesterol (n = 1837), hypertension (n = 2639), and metabolic syndrome (n = 3596) on the 4 beverages of interest (SSBs, fruit juice, low-fat milk, and whole-fat milk). All models used beverages as continuous and divided in quartiles, which allowed for tests of trend. Categorical results are presented here, with continuous results shown under \u201cSupplemental data\u201d in the online issue. We used robust estimators that allow the variance of the error term to be dependent on our independent variables (energy from beverages). Year 20 incidence models (ie, incident high WC) were run by using individuals without prevalent outcome (ie, high WC) at exam year 0 or 7. Models were adjusted for race and sex as well as year 0 age, weight, smoking status, total physical activity, calories from food, calories from the other beverages, calories from alcohol, and CARDIA exam center. Inclusion of a set of food groups (baseline servings per day of meats, vegetables, fruit, and nonmilk dairy products) as opposed to total food calories did not affect results so we included only the latter in the interest of building the most parsimonious model possible (ie, the one with the fewest variables). Interactions with baseline overweight and race were tested in multivariable linear models but were not statistically significant (likelihood ratio test, P > 0.05). P value < 0.05 was set for determining statistical significance in our continuous models.",
        "results": "### RESULTS\nBaseline (year 0) demographics are presented in Table 1. Consumers of whole-fat milk and SSBs were significantly more likely to be younger, black, male, and current smokers and to have lower levels (EU/wk) of physical activity, higher total daily energy intake, and higher SBP compared with nonconsumers (P, 0.05). Consumers of SSBs were also more likely to have lower levels of HDL cholesterol and higher diastolic blood pressure compared with nonconsumers (Table 1). There were no differences in any other anthropometric or blood lipid measure between consumers and nonconsumers. Differences between low-fat milk and fruit Juice consumers and nonconsumers are available on request from the corresponding author. Over 20 y, we observed a consistent increase in per capita calories for SSBs, adjusted for age and sex, whereas per capita intake of whole-fat and low-fat milk and fruit juice declined (Table 2). There was a slight (7%) increase between year 0 and year 7 followed by a considerable (14%) decrease between year 7 and year 20 in the percentage of the sample who consumed low-fat milk. The percentage of the population consuming SSBs steadily increased over the full time period, although only by a small percentage, whereas the percentage consuming fruit juice and whole-fat milk continuously declined. Energy per consumer from low- and whole-fat milk experienced consistent declines, whereas energy per consumer from SSBs increased steadily over the 20 y (Table 2). The most consistent adverse associations between beverage intake and incident cardiometabolic outcomes were observed for SSBs, in which higher (moving from one quartile to the next) baseline consumption was associated with a significant increase in the risk of incident high WC [adjusted relative risk (aRR): 1.09; 95% CI: 1.04, 1.14; P for trend, 0.001], high triglycerides (aRR: 1.06; 95% CI: 1.01, 1.13; P for trend = 0.033), high LDL cholesterol (aRR: 1.18; 95% CI: 1.02,1.36; P for trend = 0.018), and hypertension (aRR: 1.06; 95% CI: 1.01, 1.12; P for trend = 0.023; Table 3). Associations for SSB intake were in the same direction (associated with higher risk) for all other outcomes, although these estimates were not statistically significant (Table 3). Moving from one quartile to the next of whole-fat milk and fruit juice consumption was associated with lower risks of incident high triglycerides (aRR: 0.91; 95% CI: 0.81, 1.00; P for trend = 0.046) and hypertension (aRR: 0.89; 95% CI: 0.82, 0.97; P for trend= 0.007), respectively. Consumption of low-fat milk was not statistically significantly related to any incident cardiometabolic risk factor or to the metabolic syndrome (Table 3). Results using continuous beverage consumption were similar with the major difference that SSB intake was significantly associated with incident low HDL cholesterol (aRR: 1.06; 95% CI: 1.02, 1.10; P, 0.05) and not hypertension or incident high triglycerides, and fruit juice intake was not significantly associated with any incident cardiometabolic outcome (see the supplemental table under \u201cSupplemental data\u201d in the online issue).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Frequent Consumption of Sugar- and Artificially Sweetened Beverages and Natural and Bottled Fruit Juices Is Associated with an Increased Risk of Metabolic Syndrome in a Mediterranean Population at High Cardiovascular Disease Risk.",
        "abstract": "## BACKGROUND\nThe relation between the consumption of sweetened beverages and metabolic syndrome (MetS) is controversial.\n## OBJECTIVE\nThis analysis evaluated the associations between intakes of sugar-sweetened beverages (SSBs), artificially sweetened beverages, and natural and bottled fruit juices and the incidence of MetS in elderly individuals at high risk of cardiovascular disease (CVD) and without MetS at baseline.\n## METHODS\nWe prospectively examined 1868 participants free of MetS at baseline from the PREDIMED (PREvenci\u00f3n con DIeta MEDiterr\u00e1nea) study. MetS was defined by using the updated harmonized criteria of the International Diabetes Federation, the American Heart Association, and National Heart, Lung, and Blood Institute. Energy and nutrient intakes were evaluated at baseline and then yearly by using a validated 137-item food-frequency questionnaire. Multivariable-adjusted HRs for MetS and its components were estimated from mean intakes during follow-up. We compared the 2 highest consumption categories (1-5 and >5 servings/wk) with the lowest category (<1 serving/wk).\n## RESULTS\nA total of 930 incident cases of MetS were documented during a median follow-up of 3.24 y. When we compared consumption of >5 servings/wk with consumption of <1 serving/wk, multivariable HRs (95% CIs) for MetS incidence were 1.43 (1.00, 2.15), 1.74 (1.26, 2.41), 1.30 (1.00, 1.69), and 1.14 (1.04, 1.65) for SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices, respectively.\n## CONCLUSIONS\nThe occasional consumption of SSBs and artificially sweetened beverages (1-5 servings/wk) was not associated with the incidence of MetS in middle-aged and elderly individuals at high risk of CVD. The consumption of >5 servings/wk of all of the types of beverages analyzed was associated with an increased risk of MetS and some of its components. However, for SSBs and bottled fruit juices these associations must be interpreted with caution because of the low frequency of consumption in this population. This trial was registered at clinicaltrials.gov as ISRCTN35739639.",
        "methods": "### Methods\n\n##### Design and study population. The present study (clinicaltrials.gov; ISRCTN35739639) was conducted within the framework of the PREDIMED trial. Full details of the study design and protocol have been published elsewhere (25, 26) and are available on the PREDIMED study\u2019s website (27). The PREDIMED study is a large, multicenter, randomized parallel-group and controlled field trial conducted in Spain for the primary prevention of cardiovascular events (a list of the PREDIMED Investigators can be found in Supplemental Text). The main results of the PREDIMED trial at the primary endpoints (a composite of myocardial infarction, stroke, and cardiovascular mortality) have been published (28). Briefly, between October 2003 and June 2009, a total of 7447 participants (men aged 55\u201380 y and women aged 60\u201380 y) were randomly assigned to 1 of the 3 intervention groups: a Mediterranean diet supplemented with extra-virgin olive oil (;50 mL/d), a Mediterranean diet supplemented with mixed nuts (15 g walnuts, 7.5 g hazelnuts, and 7.5 g almonds daily), or advice on a low-fat diet (control group). Participants had no history of CVD at baseline but were at high cardiovascular risk because of the presence of T2D or $3 of the following risk factors: current smoking (>1 cigarette/d during the past month), hypertension (systolic blood pressure $140 mm Hg, diastolic blood pressure $90 mm Hg, or taking antihypertensive medication), high LDL cholesterol ($160 mg/dL), low HDL cholesterol (<40 mg/dL for men or <50 mg/dL for women), overweight or obese [BMI (in kg/m[2]) $25], or family history of premature CVD. Participants with severe chronic illness, drug or alcohol addiction, history of allergy or intolerance to olive oil or nuts, and a low predicted likelihood of changing dietary habits according to Prochaska and DiClemente\u2019s stages-of-change model (29) were excluded from the study. The institutional review boards of each recruitment center approved the protocol, and all of the subjects provided written informed consent. The study follow-up ended in December 2010. The present data were analyzed by using an observational prospective design, and participants were selected from all of the PREDIMED recruitment centers with biochemical determinations available for a follow-up of $2 y (n = 5801). The main aim of the present report was to explore the associations between the consumption of SSBs, artificially sweetened beverages, and natural and bottled fruit juices and the incidence of MetS. For this reason, participants with MetS at baseline (n = 3707) were excluded. Those who did not complete the baseline FFQ or who reported implausible total energy intakes (#500 and $3500 kcal/d in women and #800 and $4000 kcal/d in men) were also excluded. Of a total of 2094 participants fulfilling these characteristics, 226 were excluded because of missing data that prevented MetS from being diagnosed. Of the 5801 participants, 1019, 1766, 1804, 240, and 1269 did not meet the MetS criteria of abdominal obesity, hypertriglyceridemia, low HDL cholesterol, high blood pressure, and high fasting plasma glucose concentrations at baseline. Finally, 1868 participants were included in the present analysis.\n\n MetS. MetS was defined in accordance with the updated harmonized criteria of the International Diabetes Federation, the American Heart Association, and National Heart, Lung, and Blood Institute (1). Participants were considered to have MetS if they had $3 of the following: abdominal obesity for European individuals (waist circumference $88 cm in women and $102 cm in men), hypertriglyceridemia or drug treatment for high plasma TG concentrations, low HDL cholesterol (<50 mg/dL in women and <40 mg/dL in men), high blood pressure (systolic blood pressure $130 mm Hg or diastolic blood pressure $85 mm Hg) or antihypertensive drug treatment, or high fasting glucose ($100 mg/dL) or drug treatment for T2D.\n\n Dietary assessment. Dietary intake was assessed by trained dietitians with the use of a 137-item semiquantitative FFQ validated for the PREDIMED study (30), which was administered at baseline and yearly during follow-up. The reproducibility and relative validity of the FFQ used in the present study were assessed for several nutrients and food groups (30). The reproducibility and relative validity of the FFQ were also assessed in relation to the consumption of SSBs, artificially sweetened beverages, natural fruit juices (freshly extracted juice, for which the only procedure accepted was the squeezing of the whole piece of fruit), and bottled fruit juices (natural fruit juice that has been chemically changed by using authorized methods and packed and commercialized for subsequent consumption). By using the intraclass correlation coefficients (ICCs) between the FFQ and dietary records, the following reliabilities were found: 0.67 for SSBs, 0.46 for artificially sweetened beverages, 0.81 for natural fruit juices, and 0.88 for bottled fruit juices. In the reproducibility analysis, the ICCs were 0.69 for SSBs, 0.63 for artificially sweetened beverages, 0.71 for natural fruit juices, and 0.48 for bottled fruit juices. The intake of sweetened beverages and fruit juices was assessed yearly by using 5 items (SSBs, artificially sweetened beverages, bottled fruit juices, natural orange juice, and natural juices from other fruit) from the FFQ. To assess habitual intake during the previous year, frequencies of consumption were measured in 9 categories (from never or almost never to >6 servings/d) for each FFQ item. The responses to individual items were then converted into mean daily consumption (mL/d) during the follow-up by multiplying the typical portion sizes (mL) by the consumption frequency for each food and making the appropriate division for the period assessed to obtain the daily consumption. Each serving of sweetened beverages was considered to be 200 mL. In the present analysis the categories were SSBs, artificially sweetened beverages, natural fruit juices (the result of combining natural orange fruit juice and other natural fruit juices), and bottled fruit juices. Energy and nutrient intakes were calculated by using Spanish food-composition tables (31, 32).\n\n Anthropometric, biochemical, and lifestyle measurements. At baseline and yearly, participants completed the following: 1) a questionnaire on medical history, medication use, and lifestyle variables; 2) a 14-item validated questionnaire designed to assess adherence to the Mediterranean diet (33); and 3) a validated Spanish version of the Minnesota Leisure Time Physical Activity Questionnaire (34). In addition, at baseline and yearly thereafter, trained personnel measured weight and height with the use of calibrated scales and a wall-mounted stadiometer. Participants wore light clothing and no shoes. Waist circumference was measured by using an anthropometric tape midway between the lower rib and the superior border of the iliac crest. Blood pressure was measured in triplicate after 5 min of rest by using a validated semiautomatic sphygmomanometer (HEM-705CP; Omron), and the mean of these measurements was recorded. Blood samples were collected after an overnight fast, coded, shipped to a central laboratory, and stored at 280\u00b0C until analysis. Biochemical analysis was performed in local laboratories. Plasma glucose was analyzed by glucose-oxidase methodology, serum cholesterol by esterase-oxidase-peroxidase, serum TGs by glycerol-phosphate oxidase-peroxidase, and serum HDL cholesterol mainly by direct measurement, or precipitation methodology. All local laboratories satisfied external quality-control requirements. A concordance study of 9 laboratories was conducted. From each study, a mean of 200 samples was analyzed for total cholesterol, HDL cholesterol, and TGs. The laboratory of the Medical Research Institute of the Del Mar Hospital, which used ABX-Horiba commercial kits in a PENTRA-400 autoanalyzer (ABX-Horiba), was used as a reference. One center was unable to provide samples for the concordance study. The concordance analysis of lipid measurements showed the following values: r[2] [ICC (95% CI)] values between 0.85 and 0.97 [0.85 (0.77, 0.90) and 0.97 (0.95, 0.98)] for total cholesterol, between 0.81 and 0.99 [0.81 (0.73, 0.87) and 0.99 (0.99, 0.99)] for TGs, and between 0.82 and 0.96 [0.82 (0.74, 0.88) and 0.99 (0.99, 0.99)] for glucose.\n\n Statistical analyses. Intakes reported during the baseline interview and yearly during follow-up were averaged. The participants were classified according to the frequency of servings of different beverages. To better represent the long-term consumption of the different types of beverage, we used the mean beverage consumption for all analyses on the basis of assessments of items from all FFQs, which were made at baseline and yearly during the follow-up for participants who did not develop MetS. For those who did develop MetS, and because participants could have changed their dietary pattern after developing MetS, we used only data from all of the available FFQs until the year before MetS was diagnosed. The baseline characteristics of the participants are expressed as means 6 SDs or as medians (IQRs) for continuous variables and number and percentages for categorical variables. Chi-square and 1-factor ANOVA tests were used to assess differences in the baseline characteristics of the study population. Multivariable time-dependent Cox proportional regression models were fitted to assess the HRs for the incidence of MetS and its components during the follow-up according to servings (200 mL) of SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices. Both of the highest categories (1\u20135 servings/wk and >5 servings/wk) were compared with the lowest category (<1 serving/wk) as a reference. The assumption of proportional hazards was tested with time-dependent covariates. The time variable was defined as the interval between random assignment and the date of the last follow-up or the last recorded clinical event (MetS incidence) of participants who were still alive, whichever occurred first. Participants who were free of MetS or who were lost to follow-up were censored at the date of the last visit. Three different Cox regression models were adjusted for potential confounders. Model 1 adjusted for intervention group, age in years, sex, leisure-time physical activity (metabolic equivalent tasks/d) measured by the Minnesota Leisure Time Physical Activity Questionnaire, BMI (kg/m[2]), and smoking status (never, current, or former) at baseline. Model 2 additionally adjusted for total energy intake (in kcal/d) and average consumption (g/d) during follow-up of vegetables, legumes, fruit, cereals, meat, fish, bakery products, dairy products, olive oil, nuts, and alcohol (with a quadratic form being added only in the case of alcohol consumption). Model 3 additionally adjusted for the prevalence of MetS components at baseline, including abdominal obesity (yes or no), hypertriglyceridemia or drug treatment for elevated TGs (yes or no), low HDL cholesterol (yes or no), hypertension or antihypertensive treatment (yes or no), and high fasting plasma glucose or medication for hyperglycemia (yes or no). When the association between the intake of sweetened beverages and the incidence of each component of MetS was assessed, the components were excluded from the analysis, and model 2 represented the fully adjusted model. Statistical interactions between categories of sweetened beverage intake and potential effect-modifying variables such as sex, intervention group, diabetes, and smoking status were assessed by including product terms in the models; no significant interactions were found. To assess the linear trend, the median value of each category of each type of sweetened beverage analyzed was assigned and used as a continuous variable in the Cox regression models. The level of significance for all statistical tests was set at P < 0.05 for bilateral contrast. All of the analyses were performed with SPSS software, version 22.0.",
        "results": "### Results\n\nThe present analysis was conducted in 1868 participants, of whom 930 (430 men and 500 women) without MetS at baseline developed new-onset MetS during a median follow-up time of 3.24 y (IQR: 1.91\u20135.80). The mean daily intakes during follow-up were 14.5 mL, 17.1 mL, 29.3 mL, and 16.6 mL for SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices, respectively.\n\nCompared with those in the lowest category, participants in the highest category of SSB consumption were younger and presented higher diastolic blood pressure and TG concentrations. Those who consumed more SSBs also showed lower adherence to the Mediterranean diet and consumed less fruit and more baked products, alcohol, and total energy than did participants who consumed <3 servings/mo. The baseline characteristics of the population in terms of artificially sweetened beverages, natural fruit juices, and bottled fruit juices are shown in Supplemental Tables 1\u20133, respectively. Table 2 shows the multivariable-adjusted HRs (95% CIs) for MetS incidence for mean servings of various beverages during follow-up. After adjusting for potential confounders, individuals who consumed >5 servings total SSBs/wk had a higher risk of MetS (HR: 1.43; 95% CI: 1.00, 2.15; P-trend = 0.27) than those who consumed <1 serving/wk. A positive association was also found between consumption of artificially sweetened beverages and incidence of MetS for participants who consumed >5 servings/wk compared with those who consumed <1 serving/wk (HR: 1.74; 95% CI: 1.26, 2.41; P-trend = 0.004). An average consumption of 1\u20135 servings natural fruit juices/wk during follow-up was associated with a decreased risk of MetS (HR: 0.77; 95% CI: 0.65, 0.93). However, the risk was higher when consumption was >5 servings/wk (HR: 1.30; 95% CI: 1.00, 1.69; P-trend = 0.39). An average consumption of 1\u20135 servings of bottled fruit juices/wk was inversely associated with MetS incidence (HR: 0.66; 95% CI: 0.53, 0.81). However, when intake was >5 servings/wk the association was positive (HR: 1.14; 95% CI: 1.04, 1.65; P-trend = 0.311). The association between average beverage consumption during follow-up and components of MetS is presented in Table 3. Consumption of >5 servings SSBs/wk was associated with a higher risk of low HDL cholesterol (HR: 1.18; 95% CI: 1.06, 2.11; P-trend = 0.71) and high blood pressure (HR: 1.09; 95% CI: 1.04, 2.80; P-trend = 0.39) than was consumption of <1 serving/wk. Individuals who consumed 1\u20135 servings SSBs/wk also had a higher risk of high blood pressure than did those who rarely consumed SSBs (HR: 1.89; 95% CI: 1.14, 3.13). Consumption of >5 servings artificially sweetened beverages/wk was associated with an increased risk of developing abdominal obesity (HR: 1.82; 95% CI: 1.13, 2.92; P-trend = 0.039) and the hypertriglyceridemia component of MetS (HR: 1.52; 95% CI: 1.00, 2.37; P-trend = 0.08). The results for the hypertriglyceridemia component were similar for bottled fruit juices when consumption during follow-up was >5 servings/wk (HR: 1.51; 95% CI: 1.03, 2.46; P-trend = 0.23). A positive association between the intake of >5 servings natural fruit juices/wk and abdominal obesity was also observed (HR: 1.52; 95% CI: 1.02, 2.25; P-trend = 0.08).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Frequent Consumption of Sugar- and Artificially Sweetened Beverages and Natural and Bottled Fruit Juices Is Associated with an Increased Risk of Metabolic Syndrome in a Mediterranean Population at High Cardiovascular Disease Risk.",
        "abstract": "## BACKGROUND\nThe relation between the consumption of sweetened beverages and metabolic syndrome (MetS) is controversial.\n## OBJECTIVE\nThis analysis evaluated the associations between intakes of sugar-sweetened beverages (SSBs), artificially sweetened beverages, and natural and bottled fruit juices and the incidence of MetS in elderly individuals at high risk of cardiovascular disease (CVD) and without MetS at baseline.\n## METHODS\nWe prospectively examined 1868 participants free of MetS at baseline from the PREDIMED (PREvenci\u00f3n con DIeta MEDiterr\u00e1nea) study. MetS was defined by using the updated harmonized criteria of the International Diabetes Federation, the American Heart Association, and National Heart, Lung, and Blood Institute. Energy and nutrient intakes were evaluated at baseline and then yearly by using a validated 137-item food-frequency questionnaire. Multivariable-adjusted HRs for MetS and its components were estimated from mean intakes during follow-up. We compared the 2 highest consumption categories (1-5 and >5 servings/wk) with the lowest category (<1 serving/wk).\n## RESULTS\nA total of 930 incident cases of MetS were documented during a median follow-up of 3.24 y. When we compared consumption of >5 servings/wk with consumption of <1 serving/wk, multivariable HRs (95% CIs) for MetS incidence were 1.43 (1.00, 2.15), 1.74 (1.26, 2.41), 1.30 (1.00, 1.69), and 1.14 (1.04, 1.65) for SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices, respectively.\n## CONCLUSIONS\nThe occasional consumption of SSBs and artificially sweetened beverages (1-5 servings/wk) was not associated with the incidence of MetS in middle-aged and elderly individuals at high risk of CVD. The consumption of >5 servings/wk of all of the types of beverages analyzed was associated with an increased risk of MetS and some of its components. However, for SSBs and bottled fruit juices these associations must be interpreted with caution because of the low frequency of consumption in this population. This trial was registered at clinicaltrials.gov as ISRCTN35739639.",
        "methods": "### Methods\n\n##### Design and study population. The present study (clinicaltrials.gov; ISRCTN35739639) was conducted within the framework of the PREDIMED trial. Full details of the study design and protocol have been published elsewhere (25, 26) and are available on the PREDIMED study\u2019s website (27). The PREDIMED study is a large, multicenter, randomized parallel-group and controlled field trial conducted in Spain for the primary prevention of cardiovascular events (a list of the PREDIMED Investigators can be found in Supplemental Text). The main results of the PREDIMED trial at the primary endpoints (a composite of myocardial infarction, stroke, and cardiovascular mortality) have been published (28). Briefly, between October 2003 and June 2009, a total of 7447 participants (men aged 55\u201380 y and women aged 60\u201380 y) were randomly assigned to 1 of the 3 intervention groups: a Mediterranean diet supplemented with extra-virgin olive oil (;50 mL/d), a Mediterranean diet supplemented with mixed nuts (15 g walnuts, 7.5 g hazelnuts, and 7.5 g almonds daily), or advice on a low-fat diet (control group). Participants had no history of CVD at baseline but were at high cardiovascular risk because of the presence of T2D or $3 of the following risk factors: current smoking (>1 cigarette/d during the past month), hypertension (systolic blood pressure $140 mm Hg, diastolic blood pressure $90 mm Hg, or taking antihypertensive medication), high LDL cholesterol ($160 mg/dL), low HDL cholesterol (<40 mg/dL for men or <50 mg/dL for women), overweight or obese [BMI (in kg/m[2]) $25], or family history of premature CVD. Participants with severe chronic illness, drug or alcohol addiction, history of allergy or intolerance to olive oil or nuts, and a low predicted likelihood of changing dietary habits according to Prochaska and DiClemente\u2019s stages-of-change model (29) were excluded from the study. The institutional review boards of each recruitment center approved the protocol, and all of the subjects provided written informed consent. The study follow-up ended in December 2010. The present data were analyzed by using an observational prospective design, and participants were selected from all of the PREDIMED recruitment centers with biochemical determinations available for a follow-up of $2 y (n = 5801). The main aim of the present report was to explore the associations between the consumption of SSBs, artificially sweetened beverages, and natural and bottled fruit juices and the incidence of MetS. For this reason, participants with MetS at baseline (n = 3707) were excluded. Those who did not complete the baseline FFQ or who reported implausible total energy intakes (#500 and $3500 kcal/d in women and #800 and $4000 kcal/d in men) were also excluded. Of a total of 2094 participants fulfilling these characteristics, 226 were excluded because of missing data that prevented MetS from being diagnosed. Of the 5801 participants, 1019, 1766, 1804, 240, and 1269 did not meet the MetS criteria of abdominal obesity, hypertriglyceridemia, low HDL cholesterol, high blood pressure, and high fasting plasma glucose concentrations at baseline. Finally, 1868 participants were included in the present analysis.\n\n MetS. MetS was defined in accordance with the updated harmonized criteria of the International Diabetes Federation, the American Heart Association, and National Heart, Lung, and Blood Institute (1). Participants were considered to have MetS if they had $3 of the following: abdominal obesity for European individuals (waist circumference $88 cm in women and $102 cm in men), hypertriglyceridemia or drug treatment for high plasma TG concentrations, low HDL cholesterol (<50 mg/dL in women and <40 mg/dL in men), high blood pressure (systolic blood pressure $130 mm Hg or diastolic blood pressure $85 mm Hg) or antihypertensive drug treatment, or high fasting glucose ($100 mg/dL) or drug treatment for T2D.\n\n Dietary assessment. Dietary intake was assessed by trained dietitians with the use of a 137-item semiquantitative FFQ validated for the PREDIMED study (30), which was administered at baseline and yearly during follow-up. The reproducibility and relative validity of the FFQ used in the present study were assessed for several nutrients and food groups (30). The reproducibility and relative validity of the FFQ were also assessed in relation to the consumption of SSBs, artificially sweetened beverages, natural fruit juices (freshly extracted juice, for which the only procedure accepted was the squeezing of the whole piece of fruit), and bottled fruit juices (natural fruit juice that has been chemically changed by using authorized methods and packed and commercialized for subsequent consumption). By using the intraclass correlation coefficients (ICCs) between the FFQ and dietary records, the following reliabilities were found: 0.67 for SSBs, 0.46 for artificially sweetened beverages, 0.81 for natural fruit juices, and 0.88 for bottled fruit juices. In the reproducibility analysis, the ICCs were 0.69 for SSBs, 0.63 for artificially sweetened beverages, 0.71 for natural fruit juices, and 0.48 for bottled fruit juices. The intake of sweetened beverages and fruit juices was assessed yearly by using 5 items (SSBs, artificially sweetened beverages, bottled fruit juices, natural orange juice, and natural juices from other fruit) from the FFQ. To assess habitual intake during the previous year, frequencies of consumption were measured in 9 categories (from never or almost never to >6 servings/d) for each FFQ item. The responses to individual items were then converted into mean daily consumption (mL/d) during the follow-up by multiplying the typical portion sizes (mL) by the consumption frequency for each food and making the appropriate division for the period assessed to obtain the daily consumption. Each serving of sweetened beverages was considered to be 200 mL. In the present analysis the categories were SSBs, artificially sweetened beverages, natural fruit juices (the result of combining natural orange fruit juice and other natural fruit juices), and bottled fruit juices. Energy and nutrient intakes were calculated by using Spanish food-composition tables (31, 32).\n\n Anthropometric, biochemical, and lifestyle measurements. At baseline and yearly, participants completed the following: 1) a questionnaire on medical history, medication use, and lifestyle variables; 2) a 14-item validated questionnaire designed to assess adherence to the Mediterranean diet (33); and 3) a validated Spanish version of the Minnesota Leisure Time Physical Activity Questionnaire (34). In addition, at baseline and yearly thereafter, trained personnel measured weight and height with the use of calibrated scales and a wall-mounted stadiometer. Participants wore light clothing and no shoes. Waist circumference was measured by using an anthropometric tape midway between the lower rib and the superior border of the iliac crest. Blood pressure was measured in triplicate after 5 min of rest by using a validated semiautomatic sphygmomanometer (HEM-705CP; Omron), and the mean of these measurements was recorded. Blood samples were collected after an overnight fast, coded, shipped to a central laboratory, and stored at 280\u00b0C until analysis. Biochemical analysis was performed in local laboratories. Plasma glucose was analyzed by glucose-oxidase methodology, serum cholesterol by esterase-oxidase-peroxidase, serum TGs by glycerol-phosphate oxidase-peroxidase, and serum HDL cholesterol mainly by direct measurement, or precipitation methodology. All local laboratories satisfied external quality-control requirements. A concordance study of 9 laboratories was conducted. From each study, a mean of 200 samples was analyzed for total cholesterol, HDL cholesterol, and TGs. The laboratory of the Medical Research Institute of the Del Mar Hospital, which used ABX-Horiba commercial kits in a PENTRA-400 autoanalyzer (ABX-Horiba), was used as a reference. One center was unable to provide samples for the concordance study. The concordance analysis of lipid measurements showed the following values: r[2] [ICC (95% CI)] values between 0.85 and 0.97 [0.85 (0.77, 0.90) and 0.97 (0.95, 0.98)] for total cholesterol, between 0.81 and 0.99 [0.81 (0.73, 0.87) and 0.99 (0.99, 0.99)] for TGs, and between 0.82 and 0.96 [0.82 (0.74, 0.88) and 0.99 (0.99, 0.99)] for glucose.\n\n Statistical analyses. Intakes reported during the baseline interview and yearly during follow-up were averaged. The participants were classified according to the frequency of servings of different beverages. To better represent the long-term consumption of the different types of beverage, we used the mean beverage consumption for all analyses on the basis of assessments of items from all FFQs, which were made at baseline and yearly during the follow-up for participants who did not develop MetS. For those who did develop MetS, and because participants could have changed their dietary pattern after developing MetS, we used only data from all of the available FFQs until the year before MetS was diagnosed. The baseline characteristics of the participants are expressed as means 6 SDs or as medians (IQRs) for continuous variables and number and percentages for categorical variables. Chi-square and 1-factor ANOVA tests were used to assess differences in the baseline characteristics of the study population. Multivariable time-dependent Cox proportional regression models were fitted to assess the HRs for the incidence of MetS and its components during the follow-up according to servings (200 mL) of SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices. Both of the highest categories (1\u20135 servings/wk and >5 servings/wk) were compared with the lowest category (<1 serving/wk) as a reference. The assumption of proportional hazards was tested with time-dependent covariates. The time variable was defined as the interval between random assignment and the date of the last follow-up or the last recorded clinical event (MetS incidence) of participants who were still alive, whichever occurred first. Participants who were free of MetS or who were lost to follow-up were censored at the date of the last visit. Three different Cox regression models were adjusted for potential confounders. Model 1 adjusted for intervention group, age in years, sex, leisure-time physical activity (metabolic equivalent tasks/d) measured by the Minnesota Leisure Time Physical Activity Questionnaire, BMI (kg/m[2]), and smoking status (never, current, or former) at baseline. Model 2 additionally adjusted for total energy intake (in kcal/d) and average consumption (g/d) during follow-up of vegetables, legumes, fruit, cereals, meat, fish, bakery products, dairy products, olive oil, nuts, and alcohol (with a quadratic form being added only in the case of alcohol consumption). Model 3 additionally adjusted for the prevalence of MetS components at baseline, including abdominal obesity (yes or no), hypertriglyceridemia or drug treatment for elevated TGs (yes or no), low HDL cholesterol (yes or no), hypertension or antihypertensive treatment (yes or no), and high fasting plasma glucose or medication for hyperglycemia (yes or no). When the association between the intake of sweetened beverages and the incidence of each component of MetS was assessed, the components were excluded from the analysis, and model 2 represented the fully adjusted model. Statistical interactions between categories of sweetened beverage intake and potential effect-modifying variables such as sex, intervention group, diabetes, and smoking status were assessed by including product terms in the models; no significant interactions were found. To assess the linear trend, the median value of each category of each type of sweetened beverage analyzed was assigned and used as a continuous variable in the Cox regression models. The level of significance for all statistical tests was set at P < 0.05 for bilateral contrast. All of the analyses were performed with SPSS software, version 22.0.",
        "results": "### Results\n\nThe present analysis was conducted in 1868 participants, of whom 930 (430 men and 500 women) without MetS at baseline developed new-onset MetS during a median follow-up time of 3.24 y (IQR: 1.91\u20135.80). The mean daily intakes during follow-up were 14.5 mL, 17.1 mL, 29.3 mL, and 16.6 mL for SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices, respectively.\n\nCompared with those in the lowest category, participants in the highest category of SSB consumption were younger and presented higher diastolic blood pressure and TG concentrations. Those who consumed more SSBs also showed lower adherence to the Mediterranean diet and consumed less fruit and more baked products, alcohol, and total energy than did participants who consumed <3 servings/mo. The baseline characteristics of the population in terms of artificially sweetened beverages, natural fruit juices, and bottled fruit juices are shown in Supplemental Tables 1\u20133, respectively. Table 2 shows the multivariable-adjusted HRs (95% CIs) for MetS incidence for mean servings of various beverages during follow-up. After adjusting for potential confounders, individuals who consumed >5 servings total SSBs/wk had a higher risk of MetS (HR: 1.43; 95% CI: 1.00, 2.15; P-trend = 0.27) than those who consumed <1 serving/wk. A positive association was also found between consumption of artificially sweetened beverages and incidence of MetS for participants who consumed >5 servings/wk compared with those who consumed <1 serving/wk (HR: 1.74; 95% CI: 1.26, 2.41; P-trend = 0.004). An average consumption of 1\u20135 servings natural fruit juices/wk during follow-up was associated with a decreased risk of MetS (HR: 0.77; 95% CI: 0.65, 0.93). However, the risk was higher when consumption was >5 servings/wk (HR: 1.30; 95% CI: 1.00, 1.69; P-trend = 0.39). An average consumption of 1\u20135 servings of bottled fruit juices/wk was inversely associated with MetS incidence (HR: 0.66; 95% CI: 0.53, 0.81). However, when intake was >5 servings/wk the association was positive (HR: 1.14; 95% CI: 1.04, 1.65; P-trend = 0.311). The association between average beverage consumption during follow-up and components of MetS is presented in Table 3. Consumption of >5 servings SSBs/wk was associated with a higher risk of low HDL cholesterol (HR: 1.18; 95% CI: 1.06, 2.11; P-trend = 0.71) and high blood pressure (HR: 1.09; 95% CI: 1.04, 2.80; P-trend = 0.39) than was consumption of <1 serving/wk. Individuals who consumed 1\u20135 servings SSBs/wk also had a higher risk of high blood pressure than did those who rarely consumed SSBs (HR: 1.89; 95% CI: 1.14, 3.13). Consumption of >5 servings artificially sweetened beverages/wk was associated with an increased risk of developing abdominal obesity (HR: 1.82; 95% CI: 1.13, 2.92; P-trend = 0.039) and the hypertriglyceridemia component of MetS (HR: 1.52; 95% CI: 1.00, 2.37; P-trend = 0.08). The results for the hypertriglyceridemia component were similar for bottled fruit juices when consumption during follow-up was >5 servings/wk (HR: 1.51; 95% CI: 1.03, 2.46; P-trend = 0.23). A positive association between the intake of >5 servings natural fruit juices/wk and abdominal obesity was also observed (HR: 1.52; 95% CI: 1.02, 2.25; P-trend = 0.08).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Frequent Consumption of Sugar- and Artificially Sweetened Beverages and Natural and Bottled Fruit Juices Is Associated with an Increased Risk of Metabolic Syndrome in a Mediterranean Population at High Cardiovascular Disease Risk.",
        "abstract": "## BACKGROUND\nThe relation between the consumption of sweetened beverages and metabolic syndrome (MetS) is controversial.\n## OBJECTIVE\nThis analysis evaluated the associations between intakes of sugar-sweetened beverages (SSBs), artificially sweetened beverages, and natural and bottled fruit juices and the incidence of MetS in elderly individuals at high risk of cardiovascular disease (CVD) and without MetS at baseline.\n## METHODS\nWe prospectively examined 1868 participants free of MetS at baseline from the PREDIMED (PREvenci\u00f3n con DIeta MEDiterr\u00e1nea) study. MetS was defined by using the updated harmonized criteria of the International Diabetes Federation, the American Heart Association, and National Heart, Lung, and Blood Institute. Energy and nutrient intakes were evaluated at baseline and then yearly by using a validated 137-item food-frequency questionnaire. Multivariable-adjusted HRs for MetS and its components were estimated from mean intakes during follow-up. We compared the 2 highest consumption categories (1-5 and >5 servings/wk) with the lowest category (<1 serving/wk).\n## RESULTS\nA total of 930 incident cases of MetS were documented during a median follow-up of 3.24 y. When we compared consumption of >5 servings/wk with consumption of <1 serving/wk, multivariable HRs (95% CIs) for MetS incidence were 1.43 (1.00, 2.15), 1.74 (1.26, 2.41), 1.30 (1.00, 1.69), and 1.14 (1.04, 1.65) for SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices, respectively.\n## CONCLUSIONS\nThe occasional consumption of SSBs and artificially sweetened beverages (1-5 servings/wk) was not associated with the incidence of MetS in middle-aged and elderly individuals at high risk of CVD. The consumption of >5 servings/wk of all of the types of beverages analyzed was associated with an increased risk of MetS and some of its components. However, for SSBs and bottled fruit juices these associations must be interpreted with caution because of the low frequency of consumption in this population. This trial was registered at clinicaltrials.gov as ISRCTN35739639.",
        "methods": "### Methods\n\n##### Design and study population. The present study (clinicaltrials.gov; ISRCTN35739639) was conducted within the framework of the PREDIMED trial. Full details of the study design and protocol have been published elsewhere (25, 26) and are available on the PREDIMED study\u2019s website (27). The PREDIMED study is a large, multicenter, randomized parallel-group and controlled field trial conducted in Spain for the primary prevention of cardiovascular events (a list of the PREDIMED Investigators can be found in Supplemental Text). The main results of the PREDIMED trial at the primary endpoints (a composite of myocardial infarction, stroke, and cardiovascular mortality) have been published (28). Briefly, between October 2003 and June 2009, a total of 7447 participants (men aged 55\u201380 y and women aged 60\u201380 y) were randomly assigned to 1 of the 3 intervention groups: a Mediterranean diet supplemented with extra-virgin olive oil (;50 mL/d), a Mediterranean diet supplemented with mixed nuts (15 g walnuts, 7.5 g hazelnuts, and 7.5 g almonds daily), or advice on a low-fat diet (control group). Participants had no history of CVD at baseline but were at high cardiovascular risk because of the presence of T2D or $3 of the following risk factors: current smoking (>1 cigarette/d during the past month), hypertension (systolic blood pressure $140 mm Hg, diastolic blood pressure $90 mm Hg, or taking antihypertensive medication), high LDL cholesterol ($160 mg/dL), low HDL cholesterol (<40 mg/dL for men or <50 mg/dL for women), overweight or obese [BMI (in kg/m[2]) $25], or family history of premature CVD. Participants with severe chronic illness, drug or alcohol addiction, history of allergy or intolerance to olive oil or nuts, and a low predicted likelihood of changing dietary habits according to Prochaska and DiClemente\u2019s stages-of-change model (29) were excluded from the study. The institutional review boards of each recruitment center approved the protocol, and all of the subjects provided written informed consent. The study follow-up ended in December 2010. The present data were analyzed by using an observational prospective design, and participants were selected from all of the PREDIMED recruitment centers with biochemical determinations available for a follow-up of $2 y (n = 5801). The main aim of the present report was to explore the associations between the consumption of SSBs, artificially sweetened beverages, and natural and bottled fruit juices and the incidence of MetS. For this reason, participants with MetS at baseline (n = 3707) were excluded. Those who did not complete the baseline FFQ or who reported implausible total energy intakes (#500 and $3500 kcal/d in women and #800 and $4000 kcal/d in men) were also excluded. Of a total of 2094 participants fulfilling these characteristics, 226 were excluded because of missing data that prevented MetS from being diagnosed. Of the 5801 participants, 1019, 1766, 1804, 240, and 1269 did not meet the MetS criteria of abdominal obesity, hypertriglyceridemia, low HDL cholesterol, high blood pressure, and high fasting plasma glucose concentrations at baseline. Finally, 1868 participants were included in the present analysis.\n\n MetS. MetS was defined in accordance with the updated harmonized criteria of the International Diabetes Federation, the American Heart Association, and National Heart, Lung, and Blood Institute (1). Participants were considered to have MetS if they had $3 of the following: abdominal obesity for European individuals (waist circumference $88 cm in women and $102 cm in men), hypertriglyceridemia or drug treatment for high plasma TG concentrations, low HDL cholesterol (<50 mg/dL in women and <40 mg/dL in men), high blood pressure (systolic blood pressure $130 mm Hg or diastolic blood pressure $85 mm Hg) or antihypertensive drug treatment, or high fasting glucose ($100 mg/dL) or drug treatment for T2D.\n\n Dietary assessment. Dietary intake was assessed by trained dietitians with the use of a 137-item semiquantitative FFQ validated for the PREDIMED study (30), which was administered at baseline and yearly during follow-up. The reproducibility and relative validity of the FFQ used in the present study were assessed for several nutrients and food groups (30). The reproducibility and relative validity of the FFQ were also assessed in relation to the consumption of SSBs, artificially sweetened beverages, natural fruit juices (freshly extracted juice, for which the only procedure accepted was the squeezing of the whole piece of fruit), and bottled fruit juices (natural fruit juice that has been chemically changed by using authorized methods and packed and commercialized for subsequent consumption). By using the intraclass correlation coefficients (ICCs) between the FFQ and dietary records, the following reliabilities were found: 0.67 for SSBs, 0.46 for artificially sweetened beverages, 0.81 for natural fruit juices, and 0.88 for bottled fruit juices. In the reproducibility analysis, the ICCs were 0.69 for SSBs, 0.63 for artificially sweetened beverages, 0.71 for natural fruit juices, and 0.48 for bottled fruit juices. The intake of sweetened beverages and fruit juices was assessed yearly by using 5 items (SSBs, artificially sweetened beverages, bottled fruit juices, natural orange juice, and natural juices from other fruit) from the FFQ. To assess habitual intake during the previous year, frequencies of consumption were measured in 9 categories (from never or almost never to >6 servings/d) for each FFQ item. The responses to individual items were then converted into mean daily consumption (mL/d) during the follow-up by multiplying the typical portion sizes (mL) by the consumption frequency for each food and making the appropriate division for the period assessed to obtain the daily consumption. Each serving of sweetened beverages was considered to be 200 mL. In the present analysis the categories were SSBs, artificially sweetened beverages, natural fruit juices (the result of combining natural orange fruit juice and other natural fruit juices), and bottled fruit juices. Energy and nutrient intakes were calculated by using Spanish food-composition tables (31, 32).\n\n Anthropometric, biochemical, and lifestyle measurements. At baseline and yearly, participants completed the following: 1) a questionnaire on medical history, medication use, and lifestyle variables; 2) a 14-item validated questionnaire designed to assess adherence to the Mediterranean diet (33); and 3) a validated Spanish version of the Minnesota Leisure Time Physical Activity Questionnaire (34). In addition, at baseline and yearly thereafter, trained personnel measured weight and height with the use of calibrated scales and a wall-mounted stadiometer. Participants wore light clothing and no shoes. Waist circumference was measured by using an anthropometric tape midway between the lower rib and the superior border of the iliac crest. Blood pressure was measured in triplicate after 5 min of rest by using a validated semiautomatic sphygmomanometer (HEM-705CP; Omron), and the mean of these measurements was recorded. Blood samples were collected after an overnight fast, coded, shipped to a central laboratory, and stored at 280\u00b0C until analysis. Biochemical analysis was performed in local laboratories. Plasma glucose was analyzed by glucose-oxidase methodology, serum cholesterol by esterase-oxidase-peroxidase, serum TGs by glycerol-phosphate oxidase-peroxidase, and serum HDL cholesterol mainly by direct measurement, or precipitation methodology. All local laboratories satisfied external quality-control requirements. A concordance study of 9 laboratories was conducted. From each study, a mean of 200 samples was analyzed for total cholesterol, HDL cholesterol, and TGs. The laboratory of the Medical Research Institute of the Del Mar Hospital, which used ABX-Horiba commercial kits in a PENTRA-400 autoanalyzer (ABX-Horiba), was used as a reference. One center was unable to provide samples for the concordance study. The concordance analysis of lipid measurements showed the following values: r[2] [ICC (95% CI)] values between 0.85 and 0.97 [0.85 (0.77, 0.90) and 0.97 (0.95, 0.98)] for total cholesterol, between 0.81 and 0.99 [0.81 (0.73, 0.87) and 0.99 (0.99, 0.99)] for TGs, and between 0.82 and 0.96 [0.82 (0.74, 0.88) and 0.99 (0.99, 0.99)] for glucose.\n\n Statistical analyses. Intakes reported during the baseline interview and yearly during follow-up were averaged. The participants were classified according to the frequency of servings of different beverages. To better represent the long-term consumption of the different types of beverage, we used the mean beverage consumption for all analyses on the basis of assessments of items from all FFQs, which were made at baseline and yearly during the follow-up for participants who did not develop MetS. For those who did develop MetS, and because participants could have changed their dietary pattern after developing MetS, we used only data from all of the available FFQs until the year before MetS was diagnosed. The baseline characteristics of the participants are expressed as means 6 SDs or as medians (IQRs) for continuous variables and number and percentages for categorical variables. Chi-square and 1-factor ANOVA tests were used to assess differences in the baseline characteristics of the study population. Multivariable time-dependent Cox proportional regression models were fitted to assess the HRs for the incidence of MetS and its components during the follow-up according to servings (200 mL) of SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices. Both of the highest categories (1\u20135 servings/wk and >5 servings/wk) were compared with the lowest category (<1 serving/wk) as a reference. The assumption of proportional hazards was tested with time-dependent covariates. The time variable was defined as the interval between random assignment and the date of the last follow-up or the last recorded clinical event (MetS incidence) of participants who were still alive, whichever occurred first. Participants who were free of MetS or who were lost to follow-up were censored at the date of the last visit. Three different Cox regression models were adjusted for potential confounders. Model 1 adjusted for intervention group, age in years, sex, leisure-time physical activity (metabolic equivalent tasks/d) measured by the Minnesota Leisure Time Physical Activity Questionnaire, BMI (kg/m[2]), and smoking status (never, current, or former) at baseline. Model 2 additionally adjusted for total energy intake (in kcal/d) and average consumption (g/d) during follow-up of vegetables, legumes, fruit, cereals, meat, fish, bakery products, dairy products, olive oil, nuts, and alcohol (with a quadratic form being added only in the case of alcohol consumption). Model 3 additionally adjusted for the prevalence of MetS components at baseline, including abdominal obesity (yes or no), hypertriglyceridemia or drug treatment for elevated TGs (yes or no), low HDL cholesterol (yes or no), hypertension or antihypertensive treatment (yes or no), and high fasting plasma glucose or medication for hyperglycemia (yes or no). When the association between the intake of sweetened beverages and the incidence of each component of MetS was assessed, the components were excluded from the analysis, and model 2 represented the fully adjusted model. Statistical interactions between categories of sweetened beverage intake and potential effect-modifying variables such as sex, intervention group, diabetes, and smoking status were assessed by including product terms in the models; no significant interactions were found. To assess the linear trend, the median value of each category of each type of sweetened beverage analyzed was assigned and used as a continuous variable in the Cox regression models. The level of significance for all statistical tests was set at P < 0.05 for bilateral contrast. All of the analyses were performed with SPSS software, version 22.0.",
        "results": "### Results\n\nThe present analysis was conducted in 1868 participants, of whom 930 (430 men and 500 women) without MetS at baseline developed new-onset MetS during a median follow-up time of 3.24 y (IQR: 1.91\u20135.80). The mean daily intakes during follow-up were 14.5 mL, 17.1 mL, 29.3 mL, and 16.6 mL for SSBs, artificially sweetened beverages, natural fruit juices, and bottled fruit juices, respectively.\n\nCompared with those in the lowest category, participants in the highest category of SSB consumption were younger and presented higher diastolic blood pressure and TG concentrations. Those who consumed more SSBs also showed lower adherence to the Mediterranean diet and consumed less fruit and more baked products, alcohol, and total energy than did participants who consumed <3 servings/mo. The baseline characteristics of the population in terms of artificially sweetened beverages, natural fruit juices, and bottled fruit juices are shown in Supplemental Tables 1\u20133, respectively. Table 2 shows the multivariable-adjusted HRs (95% CIs) for MetS incidence for mean servings of various beverages during follow-up. After adjusting for potential confounders, individuals who consumed >5 servings total SSBs/wk had a higher risk of MetS (HR: 1.43; 95% CI: 1.00, 2.15; P-trend = 0.27) than those who consumed <1 serving/wk. A positive association was also found between consumption of artificially sweetened beverages and incidence of MetS for participants who consumed >5 servings/wk compared with those who consumed <1 serving/wk (HR: 1.74; 95% CI: 1.26, 2.41; P-trend = 0.004). An average consumption of 1\u20135 servings natural fruit juices/wk during follow-up was associated with a decreased risk of MetS (HR: 0.77; 95% CI: 0.65, 0.93). However, the risk was higher when consumption was >5 servings/wk (HR: 1.30; 95% CI: 1.00, 1.69; P-trend = 0.39). An average consumption of 1\u20135 servings of bottled fruit juices/wk was inversely associated with MetS incidence (HR: 0.66; 95% CI: 0.53, 0.81). However, when intake was >5 servings/wk the association was positive (HR: 1.14; 95% CI: 1.04, 1.65; P-trend = 0.311). The association between average beverage consumption during follow-up and components of MetS is presented in Table 3. Consumption of >5 servings SSBs/wk was associated with a higher risk of low HDL cholesterol (HR: 1.18; 95% CI: 1.06, 2.11; P-trend = 0.71) and high blood pressure (HR: 1.09; 95% CI: 1.04, 2.80; P-trend = 0.39) than was consumption of <1 serving/wk. Individuals who consumed 1\u20135 servings SSBs/wk also had a higher risk of high blood pressure than did those who rarely consumed SSBs (HR: 1.89; 95% CI: 1.14, 3.13). Consumption of >5 servings artificially sweetened beverages/wk was associated with an increased risk of developing abdominal obesity (HR: 1.82; 95% CI: 1.13, 2.92; P-trend = 0.039) and the hypertriglyceridemia component of MetS (HR: 1.52; 95% CI: 1.00, 2.37; P-trend = 0.08). The results for the hypertriglyceridemia component were similar for bottled fruit juices when consumption during follow-up was >5 servings/wk (HR: 1.51; 95% CI: 1.03, 2.46; P-trend = 0.23). A positive association between the intake of >5 servings natural fruit juices/wk and abdominal obesity was also observed (HR: 1.52; 95% CI: 1.02, 2.25; P-trend = 0.08).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Associations between Sugar Intake from Different Food Sources and Adiposity or Cardio-Metabolic Risk in Childhood and Adolescence: The Korean Child-Adolescent Cohort Study.",
        "abstract": "The increasing prevalence of childhood obesity is a serious public health problem associated with co-morbidities in adulthood, as well as childhood. This study was conducted to identify associations between total sugar intake and sugar intake from different foods (fruit, milk, and sugar-sweetened beverages (SSBs)), and adiposity and continuous metabolic syndrome scores (cMetS) among Korean children and adolescents using cohort data. The study subjects were children (n = 770) who participated in the 4th year (2008) of the Korean Child-Adolescent Cohort Study (KoCAS). Dietary intake data were collected via three-day 24-h food records, and sugar intake was calculated for the total sugar content of foods using our database compiled from various sources. Anthropometric measurements, assessments of body composition, and blood sample analysis were performed at baseline and at follow-up four years later. The cMetS was calculated based on waist circumference, triglycerides, high-density lipoprotein cholesterol, glucose, and mean arterial blood pressure. According to multiple linear regression analysis, there were no significant associations between total sugar intake and adiposity and cMetS. However, higher intake of fruit sugar at baseline was significantly associated with lower body mass index (BMI) z-scores and body fat percentages at baseline (\u03b2 = -0.10, p = 0.02 and \u03b2 = -0.78, p < 0.01, respectively). At follow-up, sugar intake from fruit at baseline was still negatively associated with the above outcomes, but only the relationship with BMI z-scores retained statistical significance (\u03b2 = -0.08, p < 0.05). There was a significant positive relationship between consumption of sugar from SSBs and cMetS at baseline (\u03b2 = 0.04, p = 0.02), but that relationship was not observed at follow-up (p = 0.83). Differences in consumption sugars from fruit and SSBs might play an important role in the risk of adiposity and metabolic disease in children and adolescents. Our results suggest that strategies for reducing sugar intake need to target particular food groups. Consequently, this information could be of value to obesity- and metabolic disease-prevention strategies.",
        "methods": "**2. Methods**\n\n_2.1. Study Population_\n\nThis study was conducted using data from the Korean Child-Adolescent Cohort Study (KoCAS), which followed a cohort of Korean students annually at four elementary schools in Gwacheon from entry into school at the age of seven years, in 2005. The overall objective of the cohort study was to identify early risk factors for obesity and associated metabolic diseases in Korean children. Details have been published previously [21]. Informed consent was obtained from participating children and their parents based on basic ethical principles. The study protocol, including consent procedures, was approved by the institutional review boards of Seoul-Paik Hospital, Inje University, and the Korea Center for Disease Control and Prevention (KCDC) (IRB number IIT-82). In Figure 1, baseline study subjects were children (mean age = 9.9 years) who were still participating in the fourth year of the KoCAS (2008) and lived in Gwacheon (n = 811). Twenty-three subjects were excluded due to missing data for age and BMI. Eighteen subjects were eliminated due to missing data for sugar intake. No subjects were eliminated due to the following exclusion criteria: (1) daily energy intake <500 kcal or >4000 kcal; (2) current treatment for hypertension, dyslipidemia, or diabetes or for a disease that impacts weight; or (3) attempting weight loss at baseline. The final study population consisted of 770 children (398 boys, 372 girls) aged nine and 10 years at baseline; 165 children dropped out during the four-year study period. Only 437 subjects at baseline and 345 subjects at follow-up had data to evaluate metabolic disease.\n\n**Figure 1. Flow chart of the study participants included for analysis.**\n\n_2.2. Measurements_\n\nThe same procedures and equipment were used at baseline and follow-up. Before height and weight measurements were taken, it was recommended that subjects remove shoes and extra clothing. An automatic stadiometer (DS-102; Jenix, Seoul, Korea) was used to estimate height. A body composition analyzer (BC41B; Tanita, Tokyo, Japan) was used to estimate weight and body fat percentage by the bio-electric impedance analysis (BIA) method. BMI was calculated as body weight in kilograms divided by height in meters squared (kg/m[2]). Age- and gender-specific BMI percentiles and BMI z-scores were calculated using the 2007 Korean National Growth Charts [22]. Waist circumference (WC) was measured between the midpoint of the lower border of the ribcage and the iliac crest to the nearest 0.1 cm using a non-elastic tape measure. Blood pressure (BP) was measured twice in sequence on the right arm by trained nurses using a mercury sphygmomanometer while participants were in a seated position. After a 12-h overnight fast, venous blood samples were collected from a brachial vein for biochemical measurements, including fasting blood sugar (FBS), total cholesterol, triglycerides (TG), and high-density lipoprotein cholesterol (HDLC). Serum total cholesterol, HDLC, and TG were measured using enzymatic assays and an auto-analyzer (Hitachi 7180, Tokyo, Japan). FBS levels were estimated using the hexokinase method and a glucose analyzer (model 7180; Hitachi, Tokyo, Japan).\n\n_2.3. Continuous Metabolic Syndrome Score (cMetS)_\n\nThere is no specific method for calculating cMetS. We derived cMetS based on both baseline and follow-up data using principal component analysis (PCA). The PCA method reflects the relative weights of the metabolic components of cMetS using WC, TG, HDLC, glucose, and mean arterial blood pressure (MAP) [23,24]. The above metabolic components were chosen based on clinical definitions from the National Cholesterol Education Program/Adult Treatment Panel III [25]. MAP was calculated according to the formula MAP = ((systolic BP \u00b4 diastolic BP)/3) + diastolic BP. Before creating the cMetS, TG and HDLC were transformed into natural log values to satisfy normality. In general, principal components are chosen that correspond with eigenvalues \u011b1.0. A sequential number of components reflects the degree of variance in the data [26]. Thus, first principal components (PC1s) are a linear combination of variables that account for the largest proportion of variance in data. In our study, the first two components were satisfied with the above criteria (eigenvalues \u011b1.0), so we applied varimax rotation. These first two components accounted for 59% of the variance at baseline (PC1s: 37.0% and PC2s: 22.0%). Individual factor scores for the first two components were calculated, and then combined to produce a single cMetS with the variance explained as weights. A high cMetS value reflects relatively high WC, TG, FBS and MAP values and relatively low HDLC values. In fact, the average cMetS derived from PCA increased according to the number of MetS risk components included.\n\n_2.4. Dietary Assessment_\n\nA modified three-day (two weekdays, one weekend day) food record was used to measure the typical dietary intake of each child. This was created for each child with parental assistance. Meal plans for school lunches were distributed to parents, and parents confirmed food intake amounts for their children. Before beginning the study, samples for food records were distributed to children and parents. The samples included the following items: (1) portion sizes using kitchen utensils (1 cup, 1 tablespoon, etc.), nutrition facts (g), and other counts (e.g., three pieces of chocolate); (2) product names (beverages, meals, snacks, etc.); and (3) seasonings and ingredients in the foods and drinks. Nutrient intakes were determined from food intakes using Computer-Aided Nutritional Analysis for Professionals, version 3.0 (CAN-pro 3.0, Korean Nutrition Society, Seoul, Korea).\n\n_2.5. Database of Total Sugar_\n\nWe created a database (DB) for total sugar using the 1078 different kinds of foods reported in the KoCAS. The data for sugar content was collected from several sources. If the food appeared in the KHANES dataset and its sugar content was recorded as \u201c0\u201d, we also recorded the food\u2019s sugar content as \u201c0\u201d. If data from the US Department of Agriculture (USDA; e.g., provisional tables and agricultural handbooks) or other journal sources recorded the food\u2019s sugar content as \u201c0\u201d in disagreement with the KHANES, we used the former data. Otherwise, data were collected from the following sources in order of preference: the Ministry of Food and Drug Safety; food company websites and nutrition labels; the Korea Food Research Institute; and revised data from the Rural Development Administration\u2019s research analysis. For foods that were not found in any database mentioned above, we used data for foods that had similar amounts of calories and carbohydrates. Total sugar intake was classified into subgroups by food source (fruit, milk, beverage, and other). The beverages category contained fruit juice, fruit and vegetable drinks, carbonated beverages, sports drinks, coffee, sweet tea, soy milk, energy drinks, and other beverages. Other sugar was defined as total sugar minus that from fruit, milk, and beverages, and included sweets (candies, chocolate, gum, jellies, caramels), sweetened grains (cakes, rice cakes, cookies, pies, cobblers, doughnuts, sweetened cereals, waffles and pancakes), sweetened dairy products (flavored milk, flavored yogurt), sugars and syrup (white, brown, and raw sugars, red pepper paste, edible syrups and honey, etc.) and natural sugar from vegetables and grains.\n\n_2.6. Covariates_\n\nPrevious studies regarding risk factors for obesity have indicated several potential confounding variables [27]. Among those potential confounders, we chose a child\u2019s socioeconomic status (SES) (household income), physical activity, stress level, and mother\u2019s BMI at baseline (age nine or 10 years) as covariates. Parental socioeconomic information (household income) and maternal height and weight were self-reported in questionnaires at baseline. Parents and children completed the questions about physical activity, screen time, and stress level together. Household income levels were divided into tertiles depending on the statistical distribution requirements of the \u201chousehold monthly income\u201d item of this study and then defined as low (\u010f300 \u02c6 10[4] Korean won/month), middle (301\u2013500 \u02c6 10[4] Korean won/month), or high (>500 \u02c6 10[4] Korean won/month). Screen-time factors included time spent watching TV, playing computer or video games, and using the Internet [28]. Weekday and weekend hours were examined separately. Average daily screen time was calculated as (weekday hours \u02c6 5 + weekend hours \u02c6 2)/7. The screen-time variable was dichotomized (<2 h/day or \u011b2 h/day) based upon American Academy of Pediatrics guidelines [29]. Physical activity was estimated via a self-report form, estimating participation in moderately intense physical activity for 30 min or more during the most recent week on a scale ranging from never to 7 days per week. \u201cModerately intense physical activity\u201d was defined as any activity that induced a slight increase in breathing frequency and heart rate and may cause light sweating, such as walking briskly, playing soccer (non-competitive), swimming slowly, and cycling. We divided physical activity level into two groups based on weekly frequency: \u201cfewer than three times/week\u201d and \u201cat least three times/week\u201d [30]. Stress level was measured on a scale from 1 (very much) to 4 (almost none) during the most recent week by self-reported questionnaires. Stress levels were grouped as: a lot (1 and 2); little (3); and almost none (4). Maternal BMI was calculated using questionnaire data about the mother\u2019s body weight and height and grouped into underweight (BMI < 18.5), normal weight (18.5 \u010f BMI < 25), and obese (BMI \u011b 25), according to the Asia\u2013Pacific guideline for obesity [31]. Since fruit fiber has a protective effect against obesity [32], fiber was considered as an additional confounder in the fruit sugar analysis.\n\n_2.7. Statistical Analyses_\n\nBaseline characteristics of study subjects were expressed as percentages for categorical variables and as means \u02d8 standard deviations (SDs) or medians (interquartile ranges) for continuous variables. Sex differences in baseline characteristics were analyzed using the chi-square test, Student\u2019s t-test, and Wilcoxon rank sum test. As the quantities of sugar consumed, including total energy, were skewed, the raw data were log-transformed. To identify potential confounders, we examined mean differences in daily total energy, total sugar intake and percent energy from total sugar by household income, physical activity, screen time, stress level, and maternal body weight status using one-way analysis of variance (ANOVA). The results were presented as back-transformed values. In addition, the magnitudes of the associations between the baseline predictors, including intake of sugar from specific food sources at baseline (total energy, total sugar, total sugar as a percentage of energy, milk sugar, fruit sugar, beverage sugar, and other sugars) and baseline and follow-up outcomes that were potential cardiovascular risk factors (BMI z-score, cMetS value, and body fat percentage) were examined using linear regression analysis. The multiple linear regression analysis adjusted for sex, age, total energy, and household income at baseline. Fiber was considered a confounder in the additional analysis regarding fruit sugar. Sub-analyses of cMetS components as outcome variables were to be performed if any associations between the intake of sugar from specific sources and cMetS were observed. SAS version 9.3 (SAS Institute, Cary, NC, USA) was used for all analyses. In all analyses, _p values < 0.05 in two-tailed tests were considered statistically significant._",
        "results": "**3. Results**\n\nThe baseline study population consisted of 770 children (398 boys and 372 girls). Mean participant age at baseline was 9.9 \u02d8 0.3 years. Baseline characteristics of the study population are shown in Table 1 for the overall sample and by sex. No statistically significant differences were observed between the sexes in terms of any of the confounders (SES, screen time, stress level, and maternal body weight status) except physical activity level. Relative to girls, boys had higher BMIs, WCs, FBS levels, HDLC levels, systolic BP, diastolic BP, MAP and cMetS values but lower body fat percent, total cholesterol, and TG levels. Daily sugar intake was higher for girls than boys (32.4 g/day vs. 35.6 g/day), and this was true not only for total sugar but also for sugar from the various sources.\n\nIn total, 165 of the baseline subjects were lost to follow-up. A comparison of baseline characteristics according to follow-up status is shown in Table S1. No statistically significant differences were observed between the two groups in terms of any of the confounders (SES, screen time, physical activity level, or maternal body weight status) except stress level. The rates of feeling stressed often and not feeling stressed were relatively high in the group lost to follow-up. There were no significant differences in obesity or metabolic index. Moreover, total and other sugar intake were high in the group lost to follow-up. Among the potential confounding factors taken into account, only household income level was significantly related to total energy intake and total sugar intake (p = 0.01 and p = 0.05, respectively). Children with low household income reported the lowest daily intake of sugar and of total calories (Table 2).\n\nTable 3 shows the results from the unadjusted models using linear regression analysis. The association between percent of total energy from sugar and body fat percent at follow-up had marginal significance (p = 0.05). Higher intake of fruit sugar at baseline was significantly associated with lower BMI z-scores and body fat percentages at baseline. At follow-up, sugar intake from fruit was still negatively associated with the above outcomes, but only the relationship with BMI z-scores retained statistical significance (\u03b2 = \u00b40.09, p = 0.02). There was a significant positive relationship between beverage sugar intake and baseline cMetS values.\n\nThe multiple linear regression analysis adjusted for sex, age, total energy, and household income at baseline is presented in Table 4. The negative associations between daily fruit sugar intake and BMI z-score and body fat percent at baseline retained statistical significance (\u03b2 = \u00b40.10, p = 0.02, \u03b2 = \u00b40.78, p < 0.01, respectively). At follow-up, the association between sugar intake from fruit and BMI z-scores remained significant (\u03b2 = \u00b40.08, p < 0.05), and the association between sugar intake from fruit and fat percent had marginal significance (\u03b2 = \u00b40.60, p = 0.05). Children who consumed more beverage sugar showed a higher cMetS at baseline (\u03b2 = 0.04, p = 0.02), but that relationship was not observed at follow-up even after adjustment. Milk sugar had a negative association with cMetS at baseline, but this difference did not reach statistical significance. Although positive trends were observed for sugar intake from other sources, those associations failed to reach statistical significance. After stratification by sex, we found no significant associations in girls while significant associations remained in boys. Fruit sugar had a negative association with adiposity and cMetS, and the results for fat percent and cMetS were significant in boys. There was a significant positive association between beverage sugar intake and cMetS at baseline in boys. However, in girls, milk sugar predicted negative cMetS and other sugars predicted positive BMI z-scores and fat percent up to marginal significance. An additional analysis considering fiber as a confounder found significant negative associations between daily fruit sugar intake at baseline and BMI z-scores at baseline and follow-up, and fat percent at baseline. In the sub-analysis of cMetS components, beverage sugar had a positive association with TG (\u03b2 = 0.04, p = 0.049) and a negative association with HDL (\u03b2 = \u00b40.02, p = 0.01).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "4"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Associations between Sugar Intake from Different Food Sources and Adiposity or Cardio-Metabolic Risk in Childhood and Adolescence: The Korean Child-Adolescent Cohort Study.",
        "abstract": "The increasing prevalence of childhood obesity is a serious public health problem associated with co-morbidities in adulthood, as well as childhood. This study was conducted to identify associations between total sugar intake and sugar intake from different foods (fruit, milk, and sugar-sweetened beverages (SSBs)), and adiposity and continuous metabolic syndrome scores (cMetS) among Korean children and adolescents using cohort data. The study subjects were children (n = 770) who participated in the 4th year (2008) of the Korean Child-Adolescent Cohort Study (KoCAS). Dietary intake data were collected via three-day 24-h food records, and sugar intake was calculated for the total sugar content of foods using our database compiled from various sources. Anthropometric measurements, assessments of body composition, and blood sample analysis were performed at baseline and at follow-up four years later. The cMetS was calculated based on waist circumference, triglycerides, high-density lipoprotein cholesterol, glucose, and mean arterial blood pressure. According to multiple linear regression analysis, there were no significant associations between total sugar intake and adiposity and cMetS. However, higher intake of fruit sugar at baseline was significantly associated with lower body mass index (BMI) z-scores and body fat percentages at baseline (\u03b2 = -0.10, p = 0.02 and \u03b2 = -0.78, p < 0.01, respectively). At follow-up, sugar intake from fruit at baseline was still negatively associated with the above outcomes, but only the relationship with BMI z-scores retained statistical significance (\u03b2 = -0.08, p < 0.05). There was a significant positive relationship between consumption of sugar from SSBs and cMetS at baseline (\u03b2 = 0.04, p = 0.02), but that relationship was not observed at follow-up (p = 0.83). Differences in consumption sugars from fruit and SSBs might play an important role in the risk of adiposity and metabolic disease in children and adolescents. Our results suggest that strategies for reducing sugar intake need to target particular food groups. Consequently, this information could be of value to obesity- and metabolic disease-prevention strategies.",
        "methods": "**2. Methods**\n\n_2.1. Study Population_\n\nThis study was conducted using data from the Korean Child-Adolescent Cohort Study (KoCAS), which followed a cohort of Korean students annually at four elementary schools in Gwacheon from entry into school at the age of seven years, in 2005. The overall objective of the cohort study was to identify early risk factors for obesity and associated metabolic diseases in Korean children. Details have been published previously [21]. Informed consent was obtained from participating children and their parents based on basic ethical principles. The study protocol, including consent procedures, was approved by the institutional review boards of Seoul-Paik Hospital, Inje University, and the Korea Center for Disease Control and Prevention (KCDC) (IRB number IIT-82). In Figure 1, baseline study subjects were children (mean age = 9.9 years) who were still participating in the fourth year of the KoCAS (2008) and lived in Gwacheon (n = 811). Twenty-three subjects were excluded due to missing data for age and BMI. Eighteen subjects were eliminated due to missing data for sugar intake. No subjects were eliminated due to the following exclusion criteria: (1) daily energy intake <500 kcal or >4000 kcal; (2) current treatment for hypertension, dyslipidemia, or diabetes or for a disease that impacts weight; or (3) attempting weight loss at baseline. The final study population consisted of 770 children (398 boys, 372 girls) aged nine and 10 years at baseline; 165 children dropped out during the four-year study period. Only 437 subjects at baseline and 345 subjects at follow-up had data to evaluate metabolic disease.\n\n**Figure 1. Flow chart of the study participants included for analysis.**\n\n_2.2. Measurements_\n\nThe same procedures and equipment were used at baseline and follow-up. Before height and weight measurements were taken, it was recommended that subjects remove shoes and extra clothing. An automatic stadiometer (DS-102; Jenix, Seoul, Korea) was used to estimate height. A body composition analyzer (BC41B; Tanita, Tokyo, Japan) was used to estimate weight and body fat percentage by the bio-electric impedance analysis (BIA) method. BMI was calculated as body weight in kilograms divided by height in meters squared (kg/m[2]). Age- and gender-specific BMI percentiles and BMI z-scores were calculated using the 2007 Korean National Growth Charts [22]. Waist circumference (WC) was measured between the midpoint of the lower border of the ribcage and the iliac crest to the nearest 0.1 cm using a non-elastic tape measure. Blood pressure (BP) was measured twice in sequence on the right arm by trained nurses using a mercury sphygmomanometer while participants were in a seated position. After a 12-h overnight fast, venous blood samples were collected from a brachial vein for biochemical measurements, including fasting blood sugar (FBS), total cholesterol, triglycerides (TG), and high-density lipoprotein cholesterol (HDLC). Serum total cholesterol, HDLC, and TG were measured using enzymatic assays and an auto-analyzer (Hitachi 7180, Tokyo, Japan). FBS levels were estimated using the hexokinase method and a glucose analyzer (model 7180; Hitachi, Tokyo, Japan).\n\n_2.3. Continuous Metabolic Syndrome Score (cMetS)_\n\nThere is no specific method for calculating cMetS. We derived cMetS based on both baseline and follow-up data using principal component analysis (PCA). The PCA method reflects the relative weights of the metabolic components of cMetS using WC, TG, HDLC, glucose, and mean arterial blood pressure (MAP) [23,24]. The above metabolic components were chosen based on clinical definitions from the National Cholesterol Education Program/Adult Treatment Panel III [25]. MAP was calculated according to the formula MAP = ((systolic BP \u00b4 diastolic BP)/3) + diastolic BP. Before creating the cMetS, TG and HDLC were transformed into natural log values to satisfy normality. In general, principal components are chosen that correspond with eigenvalues \u011b1.0. A sequential number of components reflects the degree of variance in the data [26]. Thus, first principal components (PC1s) are a linear combination of variables that account for the largest proportion of variance in data. In our study, the first two components were satisfied with the above criteria (eigenvalues \u011b1.0), so we applied varimax rotation. These first two components accounted for 59% of the variance at baseline (PC1s: 37.0% and PC2s: 22.0%). Individual factor scores for the first two components were calculated, and then combined to produce a single cMetS with the variance explained as weights. A high cMetS value reflects relatively high WC, TG, FBS and MAP values and relatively low HDLC values. In fact, the average cMetS derived from PCA increased according to the number of MetS risk components included.\n\n_2.4. Dietary Assessment_\n\nA modified three-day (two weekdays, one weekend day) food record was used to measure the typical dietary intake of each child. This was created for each child with parental assistance. Meal plans for school lunches were distributed to parents, and parents confirmed food intake amounts for their children. Before beginning the study, samples for food records were distributed to children and parents. The samples included the following items: (1) portion sizes using kitchen utensils (1 cup, 1 tablespoon, etc.), nutrition facts (g), and other counts (e.g., three pieces of chocolate); (2) product names (beverages, meals, snacks, etc.); and (3) seasonings and ingredients in the foods and drinks. Nutrient intakes were determined from food intakes using Computer-Aided Nutritional Analysis for Professionals, version 3.0 (CAN-pro 3.0, Korean Nutrition Society, Seoul, Korea).\n\n_2.5. Database of Total Sugar_\n\nWe created a database (DB) for total sugar using the 1078 different kinds of foods reported in the KoCAS. The data for sugar content was collected from several sources. If the food appeared in the KHANES dataset and its sugar content was recorded as \u201c0\u201d, we also recorded the food\u2019s sugar content as \u201c0\u201d. If data from the US Department of Agriculture (USDA; e.g., provisional tables and agricultural handbooks) or other journal sources recorded the food\u2019s sugar content as \u201c0\u201d in disagreement with the KHANES, we used the former data. Otherwise, data were collected from the following sources in order of preference: the Ministry of Food and Drug Safety; food company websites and nutrition labels; the Korea Food Research Institute; and revised data from the Rural Development Administration\u2019s research analysis. For foods that were not found in any database mentioned above, we used data for foods that had similar amounts of calories and carbohydrates. Total sugar intake was classified into subgroups by food source (fruit, milk, beverage, and other). The beverages category contained fruit juice, fruit and vegetable drinks, carbonated beverages, sports drinks, coffee, sweet tea, soy milk, energy drinks, and other beverages. Other sugar was defined as total sugar minus that from fruit, milk, and beverages, and included sweets (candies, chocolate, gum, jellies, caramels), sweetened grains (cakes, rice cakes, cookies, pies, cobblers, doughnuts, sweetened cereals, waffles and pancakes), sweetened dairy products (flavored milk, flavored yogurt), sugars and syrup (white, brown, and raw sugars, red pepper paste, edible syrups and honey, etc.) and natural sugar from vegetables and grains.\n\n_2.6. Covariates_\n\nPrevious studies regarding risk factors for obesity have indicated several potential confounding variables [27]. Among those potential confounders, we chose a child\u2019s socioeconomic status (SES) (household income), physical activity, stress level, and mother\u2019s BMI at baseline (age nine or 10 years) as covariates. Parental socioeconomic information (household income) and maternal height and weight were self-reported in questionnaires at baseline. Parents and children completed the questions about physical activity, screen time, and stress level together. Household income levels were divided into tertiles depending on the statistical distribution requirements of the \u201chousehold monthly income\u201d item of this study and then defined as low (\u010f300 \u02c6 10[4] Korean won/month), middle (301\u2013500 \u02c6 10[4] Korean won/month), or high (>500 \u02c6 10[4] Korean won/month). Screen-time factors included time spent watching TV, playing computer or video games, and using the Internet [28]. Weekday and weekend hours were examined separately. Average daily screen time was calculated as (weekday hours \u02c6 5 + weekend hours \u02c6 2)/7. The screen-time variable was dichotomized (<2 h/day or \u011b2 h/day) based upon American Academy of Pediatrics guidelines [29]. Physical activity was estimated via a self-report form, estimating participation in moderately intense physical activity for 30 min or more during the most recent week on a scale ranging from never to 7 days per week. \u201cModerately intense physical activity\u201d was defined as any activity that induced a slight increase in breathing frequency and heart rate and may cause light sweating, such as walking briskly, playing soccer (non-competitive), swimming slowly, and cycling. We divided physical activity level into two groups based on weekly frequency: \u201cfewer than three times/week\u201d and \u201cat least three times/week\u201d [30]. Stress level was measured on a scale from 1 (very much) to 4 (almost none) during the most recent week by self-reported questionnaires. Stress levels were grouped as: a lot (1 and 2); little (3); and almost none (4). Maternal BMI was calculated using questionnaire data about the mother\u2019s body weight and height and grouped into underweight (BMI < 18.5), normal weight (18.5 \u010f BMI < 25), and obese (BMI \u011b 25), according to the Asia\u2013Pacific guideline for obesity [31]. Since fruit fiber has a protective effect against obesity [32], fiber was considered as an additional confounder in the fruit sugar analysis.\n\n_2.7. Statistical Analyses_\n\nBaseline characteristics of study subjects were expressed as percentages for categorical variables and as means \u02d8 standard deviations (SDs) or medians (interquartile ranges) for continuous variables. Sex differences in baseline characteristics were analyzed using the chi-square test, Student\u2019s t-test, and Wilcoxon rank sum test. As the quantities of sugar consumed, including total energy, were skewed, the raw data were log-transformed. To identify potential confounders, we examined mean differences in daily total energy, total sugar intake and percent energy from total sugar by household income, physical activity, screen time, stress level, and maternal body weight status using one-way analysis of variance (ANOVA). The results were presented as back-transformed values. In addition, the magnitudes of the associations between the baseline predictors, including intake of sugar from specific food sources at baseline (total energy, total sugar, total sugar as a percentage of energy, milk sugar, fruit sugar, beverage sugar, and other sugars) and baseline and follow-up outcomes that were potential cardiovascular risk factors (BMI z-score, cMetS value, and body fat percentage) were examined using linear regression analysis. The multiple linear regression analysis adjusted for sex, age, total energy, and household income at baseline. Fiber was considered a confounder in the additional analysis regarding fruit sugar. Sub-analyses of cMetS components as outcome variables were to be performed if any associations between the intake of sugar from specific sources and cMetS were observed. SAS version 9.3 (SAS Institute, Cary, NC, USA) was used for all analyses. In all analyses, _p values < 0.05 in two-tailed tests were considered statistically significant._",
        "results": "**3. Results**\n\nThe baseline study population consisted of 770 children (398 boys and 372 girls). Mean participant age at baseline was 9.9 \u02d8 0.3 years. Baseline characteristics of the study population are shown in Table 1 for the overall sample and by sex. No statistically significant differences were observed between the sexes in terms of any of the confounders (SES, screen time, stress level, and maternal body weight status) except physical activity level. Relative to girls, boys had higher BMIs, WCs, FBS levels, HDLC levels, systolic BP, diastolic BP, MAP and cMetS values but lower body fat percent, total cholesterol, and TG levels. Daily sugar intake was higher for girls than boys (32.4 g/day vs. 35.6 g/day), and this was true not only for total sugar but also for sugar from the various sources.\n\nIn total, 165 of the baseline subjects were lost to follow-up. A comparison of baseline characteristics according to follow-up status is shown in Table S1. No statistically significant differences were observed between the two groups in terms of any of the confounders (SES, screen time, physical activity level, or maternal body weight status) except stress level. The rates of feeling stressed often and not feeling stressed were relatively high in the group lost to follow-up. There were no significant differences in obesity or metabolic index. Moreover, total and other sugar intake were high in the group lost to follow-up. Among the potential confounding factors taken into account, only household income level was significantly related to total energy intake and total sugar intake (p = 0.01 and p = 0.05, respectively). Children with low household income reported the lowest daily intake of sugar and of total calories (Table 2).\n\nTable 3 shows the results from the unadjusted models using linear regression analysis. The association between percent of total energy from sugar and body fat percent at follow-up had marginal significance (p = 0.05). Higher intake of fruit sugar at baseline was significantly associated with lower BMI z-scores and body fat percentages at baseline. At follow-up, sugar intake from fruit was still negatively associated with the above outcomes, but only the relationship with BMI z-scores retained statistical significance (\u03b2 = \u00b40.09, p = 0.02). There was a significant positive relationship between beverage sugar intake and baseline cMetS values.\n\nThe multiple linear regression analysis adjusted for sex, age, total energy, and household income at baseline is presented in Table 4. The negative associations between daily fruit sugar intake and BMI z-score and body fat percent at baseline retained statistical significance (\u03b2 = \u00b40.10, p = 0.02, \u03b2 = \u00b40.78, p < 0.01, respectively). At follow-up, the association between sugar intake from fruit and BMI z-scores remained significant (\u03b2 = \u00b40.08, p < 0.05), and the association between sugar intake from fruit and fat percent had marginal significance (\u03b2 = \u00b40.60, p = 0.05). Children who consumed more beverage sugar showed a higher cMetS at baseline (\u03b2 = 0.04, p = 0.02), but that relationship was not observed at follow-up even after adjustment. Milk sugar had a negative association with cMetS at baseline, but this difference did not reach statistical significance. Although positive trends were observed for sugar intake from other sources, those associations failed to reach statistical significance. After stratification by sex, we found no significant associations in girls while significant associations remained in boys. Fruit sugar had a negative association with adiposity and cMetS, and the results for fat percent and cMetS were significant in boys. There was a significant positive association between beverage sugar intake and cMetS at baseline in boys. However, in girls, milk sugar predicted negative cMetS and other sugars predicted positive BMI z-scores and fat percent up to marginal significance. An additional analysis considering fiber as a confounder found significant negative associations between daily fruit sugar intake at baseline and BMI z-scores at baseline and follow-up, and fat percent at baseline. In the sub-analysis of cMetS components, beverage sugar had a positive association with TG (\u03b2 = 0.04, p = 0.049) and a negative association with HDL (\u03b2 = \u00b40.02, p = 0.01).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Associations between Sugar Intake from Different Food Sources and Adiposity or Cardio-Metabolic Risk in Childhood and Adolescence: The Korean Child-Adolescent Cohort Study.",
        "abstract": "The increasing prevalence of childhood obesity is a serious public health problem associated with co-morbidities in adulthood, as well as childhood. This study was conducted to identify associations between total sugar intake and sugar intake from different foods (fruit, milk, and sugar-sweetened beverages (SSBs)), and adiposity and continuous metabolic syndrome scores (cMetS) among Korean children and adolescents using cohort data. The study subjects were children (n = 770) who participated in the 4th year (2008) of the Korean Child-Adolescent Cohort Study (KoCAS). Dietary intake data were collected via three-day 24-h food records, and sugar intake was calculated for the total sugar content of foods using our database compiled from various sources. Anthropometric measurements, assessments of body composition, and blood sample analysis were performed at baseline and at follow-up four years later. The cMetS was calculated based on waist circumference, triglycerides, high-density lipoprotein cholesterol, glucose, and mean arterial blood pressure. According to multiple linear regression analysis, there were no significant associations between total sugar intake and adiposity and cMetS. However, higher intake of fruit sugar at baseline was significantly associated with lower body mass index (BMI) z-scores and body fat percentages at baseline (\u03b2 = -0.10, p = 0.02 and \u03b2 = -0.78, p < 0.01, respectively). At follow-up, sugar intake from fruit at baseline was still negatively associated with the above outcomes, but only the relationship with BMI z-scores retained statistical significance (\u03b2 = -0.08, p < 0.05). There was a significant positive relationship between consumption of sugar from SSBs and cMetS at baseline (\u03b2 = 0.04, p = 0.02), but that relationship was not observed at follow-up (p = 0.83). Differences in consumption sugars from fruit and SSBs might play an important role in the risk of adiposity and metabolic disease in children and adolescents. Our results suggest that strategies for reducing sugar intake need to target particular food groups. Consequently, this information could be of value to obesity- and metabolic disease-prevention strategies.",
        "methods": "**2. Methods**\n\n_2.1. Study Population_\n\nThis study was conducted using data from the Korean Child-Adolescent Cohort Study (KoCAS), which followed a cohort of Korean students annually at four elementary schools in Gwacheon from entry into school at the age of seven years, in 2005. The overall objective of the cohort study was to identify early risk factors for obesity and associated metabolic diseases in Korean children. Details have been published previously [21]. Informed consent was obtained from participating children and their parents based on basic ethical principles. The study protocol, including consent procedures, was approved by the institutional review boards of Seoul-Paik Hospital, Inje University, and the Korea Center for Disease Control and Prevention (KCDC) (IRB number IIT-82). In Figure 1, baseline study subjects were children (mean age = 9.9 years) who were still participating in the fourth year of the KoCAS (2008) and lived in Gwacheon (n = 811). Twenty-three subjects were excluded due to missing data for age and BMI. Eighteen subjects were eliminated due to missing data for sugar intake. No subjects were eliminated due to the following exclusion criteria: (1) daily energy intake <500 kcal or >4000 kcal; (2) current treatment for hypertension, dyslipidemia, or diabetes or for a disease that impacts weight; or (3) attempting weight loss at baseline. The final study population consisted of 770 children (398 boys, 372 girls) aged nine and 10 years at baseline; 165 children dropped out during the four-year study period. Only 437 subjects at baseline and 345 subjects at follow-up had data to evaluate metabolic disease.\n\n**Figure 1. Flow chart of the study participants included for analysis.**\n\n_2.2. Measurements_\n\nThe same procedures and equipment were used at baseline and follow-up. Before height and weight measurements were taken, it was recommended that subjects remove shoes and extra clothing. An automatic stadiometer (DS-102; Jenix, Seoul, Korea) was used to estimate height. A body composition analyzer (BC41B; Tanita, Tokyo, Japan) was used to estimate weight and body fat percentage by the bio-electric impedance analysis (BIA) method. BMI was calculated as body weight in kilograms divided by height in meters squared (kg/m[2]). Age- and gender-specific BMI percentiles and BMI z-scores were calculated using the 2007 Korean National Growth Charts [22]. Waist circumference (WC) was measured between the midpoint of the lower border of the ribcage and the iliac crest to the nearest 0.1 cm using a non-elastic tape measure. Blood pressure (BP) was measured twice in sequence on the right arm by trained nurses using a mercury sphygmomanometer while participants were in a seated position. After a 12-h overnight fast, venous blood samples were collected from a brachial vein for biochemical measurements, including fasting blood sugar (FBS), total cholesterol, triglycerides (TG), and high-density lipoprotein cholesterol (HDLC). Serum total cholesterol, HDLC, and TG were measured using enzymatic assays and an auto-analyzer (Hitachi 7180, Tokyo, Japan). FBS levels were estimated using the hexokinase method and a glucose analyzer (model 7180; Hitachi, Tokyo, Japan).\n\n_2.3. Continuous Metabolic Syndrome Score (cMetS)_\n\nThere is no specific method for calculating cMetS. We derived cMetS based on both baseline and follow-up data using principal component analysis (PCA). The PCA method reflects the relative weights of the metabolic components of cMetS using WC, TG, HDLC, glucose, and mean arterial blood pressure (MAP) [23,24]. The above metabolic components were chosen based on clinical definitions from the National Cholesterol Education Program/Adult Treatment Panel III [25]. MAP was calculated according to the formula MAP = ((systolic BP \u00b4 diastolic BP)/3) + diastolic BP. Before creating the cMetS, TG and HDLC were transformed into natural log values to satisfy normality. In general, principal components are chosen that correspond with eigenvalues \u011b1.0. A sequential number of components reflects the degree of variance in the data [26]. Thus, first principal components (PC1s) are a linear combination of variables that account for the largest proportion of variance in data. In our study, the first two components were satisfied with the above criteria (eigenvalues \u011b1.0), so we applied varimax rotation. These first two components accounted for 59% of the variance at baseline (PC1s: 37.0% and PC2s: 22.0%). Individual factor scores for the first two components were calculated, and then combined to produce a single cMetS with the variance explained as weights. A high cMetS value reflects relatively high WC, TG, FBS and MAP values and relatively low HDLC values. In fact, the average cMetS derived from PCA increased according to the number of MetS risk components included.\n\n_2.4. Dietary Assessment_\n\nA modified three-day (two weekdays, one weekend day) food record was used to measure the typical dietary intake of each child. This was created for each child with parental assistance. Meal plans for school lunches were distributed to parents, and parents confirmed food intake amounts for their children. Before beginning the study, samples for food records were distributed to children and parents. The samples included the following items: (1) portion sizes using kitchen utensils (1 cup, 1 tablespoon, etc.), nutrition facts (g), and other counts (e.g., three pieces of chocolate); (2) product names (beverages, meals, snacks, etc.); and (3) seasonings and ingredients in the foods and drinks. Nutrient intakes were determined from food intakes using Computer-Aided Nutritional Analysis for Professionals, version 3.0 (CAN-pro 3.0, Korean Nutrition Society, Seoul, Korea).\n\n_2.5. Database of Total Sugar_\n\nWe created a database (DB) for total sugar using the 1078 different kinds of foods reported in the KoCAS. The data for sugar content was collected from several sources. If the food appeared in the KHANES dataset and its sugar content was recorded as \u201c0\u201d, we also recorded the food\u2019s sugar content as \u201c0\u201d. If data from the US Department of Agriculture (USDA; e.g., provisional tables and agricultural handbooks) or other journal sources recorded the food\u2019s sugar content as \u201c0\u201d in disagreement with the KHANES, we used the former data. Otherwise, data were collected from the following sources in order of preference: the Ministry of Food and Drug Safety; food company websites and nutrition labels; the Korea Food Research Institute; and revised data from the Rural Development Administration\u2019s research analysis. For foods that were not found in any database mentioned above, we used data for foods that had similar amounts of calories and carbohydrates. Total sugar intake was classified into subgroups by food source (fruit, milk, beverage, and other). The beverages category contained fruit juice, fruit and vegetable drinks, carbonated beverages, sports drinks, coffee, sweet tea, soy milk, energy drinks, and other beverages. Other sugar was defined as total sugar minus that from fruit, milk, and beverages, and included sweets (candies, chocolate, gum, jellies, caramels), sweetened grains (cakes, rice cakes, cookies, pies, cobblers, doughnuts, sweetened cereals, waffles and pancakes), sweetened dairy products (flavored milk, flavored yogurt), sugars and syrup (white, brown, and raw sugars, red pepper paste, edible syrups and honey, etc.) and natural sugar from vegetables and grains.\n\n_2.6. Covariates_\n\nPrevious studies regarding risk factors for obesity have indicated several potential confounding variables [27]. Among those potential confounders, we chose a child\u2019s socioeconomic status (SES) (household income), physical activity, stress level, and mother\u2019s BMI at baseline (age nine or 10 years) as covariates. Parental socioeconomic information (household income) and maternal height and weight were self-reported in questionnaires at baseline. Parents and children completed the questions about physical activity, screen time, and stress level together. Household income levels were divided into tertiles depending on the statistical distribution requirements of the \u201chousehold monthly income\u201d item of this study and then defined as low (\u010f300 \u02c6 10[4] Korean won/month), middle (301\u2013500 \u02c6 10[4] Korean won/month), or high (>500 \u02c6 10[4] Korean won/month). Screen-time factors included time spent watching TV, playing computer or video games, and using the Internet [28]. Weekday and weekend hours were examined separately. Average daily screen time was calculated as (weekday hours \u02c6 5 + weekend hours \u02c6 2)/7. The screen-time variable was dichotomized (<2 h/day or \u011b2 h/day) based upon American Academy of Pediatrics guidelines [29]. Physical activity was estimated via a self-report form, estimating participation in moderately intense physical activity for 30 min or more during the most recent week on a scale ranging from never to 7 days per week. \u201cModerately intense physical activity\u201d was defined as any activity that induced a slight increase in breathing frequency and heart rate and may cause light sweating, such as walking briskly, playing soccer (non-competitive), swimming slowly, and cycling. We divided physical activity level into two groups based on weekly frequency: \u201cfewer than three times/week\u201d and \u201cat least three times/week\u201d [30]. Stress level was measured on a scale from 1 (very much) to 4 (almost none) during the most recent week by self-reported questionnaires. Stress levels were grouped as: a lot (1 and 2); little (3); and almost none (4). Maternal BMI was calculated using questionnaire data about the mother\u2019s body weight and height and grouped into underweight (BMI < 18.5), normal weight (18.5 \u010f BMI < 25), and obese (BMI \u011b 25), according to the Asia\u2013Pacific guideline for obesity [31]. Since fruit fiber has a protective effect against obesity [32], fiber was considered as an additional confounder in the fruit sugar analysis.\n\n_2.7. Statistical Analyses_\n\nBaseline characteristics of study subjects were expressed as percentages for categorical variables and as means \u02d8 standard deviations (SDs) or medians (interquartile ranges) for continuous variables. Sex differences in baseline characteristics were analyzed using the chi-square test, Student\u2019s t-test, and Wilcoxon rank sum test. As the quantities of sugar consumed, including total energy, were skewed, the raw data were log-transformed. To identify potential confounders, we examined mean differences in daily total energy, total sugar intake and percent energy from total sugar by household income, physical activity, screen time, stress level, and maternal body weight status using one-way analysis of variance (ANOVA). The results were presented as back-transformed values. In addition, the magnitudes of the associations between the baseline predictors, including intake of sugar from specific food sources at baseline (total energy, total sugar, total sugar as a percentage of energy, milk sugar, fruit sugar, beverage sugar, and other sugars) and baseline and follow-up outcomes that were potential cardiovascular risk factors (BMI z-score, cMetS value, and body fat percentage) were examined using linear regression analysis. The multiple linear regression analysis adjusted for sex, age, total energy, and household income at baseline. Fiber was considered a confounder in the additional analysis regarding fruit sugar. Sub-analyses of cMetS components as outcome variables were to be performed if any associations between the intake of sugar from specific sources and cMetS were observed. SAS version 9.3 (SAS Institute, Cary, NC, USA) was used for all analyses. In all analyses, _p values < 0.05 in two-tailed tests were considered statistically significant._",
        "results": "**3. Results**\n\nThe baseline study population consisted of 770 children (398 boys and 372 girls). Mean participant age at baseline was 9.9 \u02d8 0.3 years. Baseline characteristics of the study population are shown in Table 1 for the overall sample and by sex. No statistically significant differences were observed between the sexes in terms of any of the confounders (SES, screen time, stress level, and maternal body weight status) except physical activity level. Relative to girls, boys had higher BMIs, WCs, FBS levels, HDLC levels, systolic BP, diastolic BP, MAP and cMetS values but lower body fat percent, total cholesterol, and TG levels. Daily sugar intake was higher for girls than boys (32.4 g/day vs. 35.6 g/day), and this was true not only for total sugar but also for sugar from the various sources.\n\nIn total, 165 of the baseline subjects were lost to follow-up. A comparison of baseline characteristics according to follow-up status is shown in Table S1. No statistically significant differences were observed between the two groups in terms of any of the confounders (SES, screen time, physical activity level, or maternal body weight status) except stress level. The rates of feeling stressed often and not feeling stressed were relatively high in the group lost to follow-up. There were no significant differences in obesity or metabolic index. Moreover, total and other sugar intake were high in the group lost to follow-up. Among the potential confounding factors taken into account, only household income level was significantly related to total energy intake and total sugar intake (p = 0.01 and p = 0.05, respectively). Children with low household income reported the lowest daily intake of sugar and of total calories (Table 2).\n\nTable 3 shows the results from the unadjusted models using linear regression analysis. The association between percent of total energy from sugar and body fat percent at follow-up had marginal significance (p = 0.05). Higher intake of fruit sugar at baseline was significantly associated with lower BMI z-scores and body fat percentages at baseline. At follow-up, sugar intake from fruit was still negatively associated with the above outcomes, but only the relationship with BMI z-scores retained statistical significance (\u03b2 = \u00b40.09, p = 0.02). There was a significant positive relationship between beverage sugar intake and baseline cMetS values.\n\nThe multiple linear regression analysis adjusted for sex, age, total energy, and household income at baseline is presented in Table 4. The negative associations between daily fruit sugar intake and BMI z-score and body fat percent at baseline retained statistical significance (\u03b2 = \u00b40.10, p = 0.02, \u03b2 = \u00b40.78, p < 0.01, respectively). At follow-up, the association between sugar intake from fruit and BMI z-scores remained significant (\u03b2 = \u00b40.08, p < 0.05), and the association between sugar intake from fruit and fat percent had marginal significance (\u03b2 = \u00b40.60, p = 0.05). Children who consumed more beverage sugar showed a higher cMetS at baseline (\u03b2 = 0.04, p = 0.02), but that relationship was not observed at follow-up even after adjustment. Milk sugar had a negative association with cMetS at baseline, but this difference did not reach statistical significance. Although positive trends were observed for sugar intake from other sources, those associations failed to reach statistical significance. After stratification by sex, we found no significant associations in girls while significant associations remained in boys. Fruit sugar had a negative association with adiposity and cMetS, and the results for fat percent and cMetS were significant in boys. There was a significant positive association between beverage sugar intake and cMetS at baseline in boys. However, in girls, milk sugar predicted negative cMetS and other sugars predicted positive BMI z-scores and fat percent up to marginal significance. An additional analysis considering fiber as a confounder found significant negative associations between daily fruit sugar intake at baseline and BMI z-scores at baseline and follow-up, and fat percent at baseline. In the sub-analysis of cMetS components, beverage sugar had a positive association with TG (\u03b2 = 0.04, p = 0.049) and a negative association with HDL (\u03b2 = \u00b40.02, p = 0.01).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "1"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Soft drink consumption is associated with increased incidence of the metabolic syndrome only in women.",
        "abstract": "Prospective studies on the association between soft drink consumption and incident risk of the metabolic syndrome (MetS) have not been carried out in Asians. We explored the sex-specific association between soft drink consumption and incident risk of the MetS in Korean adults during 10 years of follow-up. A total of 5797 subjects who were free of the MetS at baseline were studied. Soft drink consumption was assessed using a semi-quantitative FFQ. Time-dependent Cox proportional hazard model was used to examine hazard ratios (HR) of incidence of the MetS and its components in relation to soft drink consumption. In women, the multivariable-adjusted HR for developing the MetS was 1\u00b78-fold higher in frequent consumers of soft drinks (\u22654 servings/week) compared with rare consumers (95 % CI 1\u00b723, 2\u00b764). The adjusted HR for elevated blood pressure increased by 2-fold (95 % CI 1\u00b724, 3\u00b714) and for hypertriacylglycerolaemia by 1\u00b79-fold (95 % CI 1\u00b719, 2\u00b788) in frequent consumers of soft drinks compared with rare consumers. However, in men, there was no association between soft drink consumption and incident risk of the MetS or its components. Frequent soft drink consumption was associated with increased risk of developing the MetS and its components only in middle-aged Korean women, suggesting sex differences for the risk of the MetS related to diet.",
        "methods": "## Methods\n\n### Study subjects\n\nThe KoGES, a large-scale, community-based cohort study, was initiated in 2001. Initially, the target population consisted of 10 030 Korean adults aged 40\u201369 years and living in Ansan (urban) and Ansung (rural) areas. All participants responded to a baseline examination (2001\u20132002) with questionnaires including demographic information, socio-economic status, lifestyle, dietary intake, medical history, health examinations and biochemical measurements. Follow-up examinations were performed every 2 years over a 10-year period (2009\u20132010). The present study was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention (KCDC). Informed written consent was obtained from all participants. Of the original 10030 participants, participants without the MetS (n 7053) at baseline were included. We excluded participants who refused to participate in follow-up examinations (n 747), those who provided insufficient information (n 313), those who did not respond to dietary examination (n 59) and those who had CVD or cancer (n 137). After exclusion, a total of 5797 participants (3027 men and 2770 women) were included in the analysis during the 10-year period. A follow-up rate of 63\u00b73 % was achieved, resulting in 33269 person-years accrued. The average follow-up period was 5\u00b77 years (68\u00b78 months).\n\n### Dietary assessment\n\nTrained dietitians examined dietary intake both at baseline and at the second follow-up examination (2005\u20132006) using a 103-item, semi-quantitative FFQ. The FFQ was developed and validated by the KCDC. In the FFQ, participants reported the frequency and portion sizes of soft drink consumption during the past year. Soft drink consumption was estimated by the questions \u2018How often do you consume soft drinks (carbonated beverages, e.g., Cola and Sprite)?\u2019 for frequency and \u2018How much soft drink do you consume at once?\u2019 for portion size. The original responses for frequency included none or rarely, once/month, two to three times per month, one to two times per week, three to four times per week, five to six times per week, one to two times per d, three to four times per d and \u22655 times/d. The response options for portion size were as follows: 1/2 cup (100ml), one cup (200ml \u2013 one serving) and two or more cups (\u2265400ml). For analysis, the nine response items were converted to frequency per week, and the serving per week was calculated on the basis of frequency and portion size (mean frequency/week \u00d7 0\u00b75 for 1/2 cup, \u00d71\u00b70 for one cup, \u00d72\u00b70 for two or more cups) and then categorised into four groups: none or rarely, <1 serving/week, \u22651 serving/week to <4 servings/week and \u22654 servings/week. Nutrient intakes were measured using a database developed by the Rural Development and Administration.\n\n### Health examination\n\nTrained professionals conducted a comprehensive health examination under the KCDC protocol. Height and body weight were measured to the nearest 0\u00b71 cm and 0\u00b71kg, respectively, with no shoes and while wearing light clothing. BMI was defined as weight (kg)/height squared (m[2]). Waist circumference (WC) measurements were repeated three times, and then averaged after measuring to the nearest 0\u00b71cm at the narrowest point between the lowest rib and the right iliac crest. Blood pressure (BP) was measured after participants had rested for more than 5min in a sitting position (Baumanometer-Standby; W.A. Baum Co. Inc.). Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were calculated at Korotkoff phase I and Korotkoff phase V, respectively. Averaged values of left and right arms were measured a few times at 30-s intervals. Blood samples were collected to measure fasting blood glucose (FBG), TAG and HDL-cholesterol. The blood samples were collected after at least 8 h of fasting at baseline and during every follow-up examination. The concentrations of glucose, TAG and HDL-cholesterol in plasma were enzymatically measured using an autoanalyzer (ADVIA 1650; Bayer HealthCare). Incidence of the MetS was diagnosed at the follow-up examination by a physician.\n\n### Definition of the metabolic syndrome\n\nThe MetS was diagnosed using criteria based on the National Cholesterol Education Program Adult Treatment Panel III definition. The MetS is defined as the presence of more than three of the following indicators: (1) abdominal obesity (WC \u2265 90cm for men or \u226580cm for women), (2) elevated BP (SBP \u2265 130mmHg or DBP \u2265 85mmHg), use of antihypertensive medication, hypertension diagnosis by a physician, (3) high blood glucose (FBG \u2265 5.6 mmol/l), current use of insulin or oral hypoglycaemic medication, diabetes diagnosis by a physician, (4) hypertriacylglycerolaemia (TAG \u22651.7mmol/l) and (5) low HDL-cholesterol (HDL-cholesterol <1.0mmol/l in men or <1.3 mmol/l in women).\n\n### Other measurements\n\nDemographic characteristics, socio-economic status and lifestyle factors were examined from the baseline questionnaires. Income level was divided into four groups: <1 million KRW, 1\u20132 million KRW, 2\u20133 million KRW and \u22653 million KRW. Education level was categorised into three groups: \u22646 years (elementary school level), 7\u201312 years (middle/high school level) and >12 years (college level) of education. Smoking status was classified as follows: non-smoker, former smoker or current smoker. Alcohol intake was classified as follows: non-drinker, former drinker or current drinker. Physical activity was self-reported by a questionnaire. The questionnaire asked how many hours per day was spent performing physical activities according to its intensity (sedentary, very light, light, moderate, vigorous). Physical activities are presented as metabolic equivalents.\n\n### Statistical analysis\n\nAll the data were analysed using SAS software version 9.3 (SAS Institute). The results are expressed as percentages (categorical) or as means with their standard deviations (continuous). Differences in baseline characteristics by sex were examined by the \u03c7[2] test for categorical variables or Student\u2019s t test for continuous variables. Differences in characteristics across soft drink consumption were evaluated by the \u03c7[2] tests or a generalised linear model with post hoc Tukey\u2019s honestly significant difference test. Time-dependent Cox proportional hazard models were used as a method of survival analysis to examine the hazard ratios (HR) and 95 % CI for the incident risk of the MetS and its individual components according to soft drink consumption. Survival analysis was performed separately in men and women. In multivariable adjusted models, model 1 was adjusted for age and model 2 was adjusted for age, income level, education level, alcohol consumption, smoking status, physical activity, BMI, energy intake, percentage of fat from energy, fibre intake and the presence of diseases (diabetes and hypertension). For the selection of variables for adjustment in the multivariable model, potential confounders from the previously published scientific literature were taken into account with the statistical approach, such as stepwise procedures or comparing adjusted and unadjusted effect estimates. Individuals with hypertension or diabetes at baseline were excluded from the survival analysis to examine the risk of incident elevated BP and high blood glucose, respectively. The proportional hazards assumption was assessed graphically using log\u2013log plots and statistically using Schoenfeld\u2019s residuals. No violation of the proportional hazard assumption was detected. Tests for linear trends were performed on the basis of the median value of each category. All P values < 0\u00b705 were considered statistically significant.",
        "results": "## Results\n\nTable 1. Characteristics of the study subjects at baseline (Numbers and percentages (categorised variables); mean values and standard deviations (continuous variables))\n\nMetS Non-MetS  \n(n 2129) (n 3668)  \n\nCharacteristics of subjects at baseline\n\nAmong 5797 subjects, a total of 2129 (1046 men and 1083 women) developed the MetS. Table 1 shows the baseline characteristics of subjects by the presence of the MetS. Subjects with the MetS were older, more likely to be women, to live in the rural areas, to be exercisers, to have lower income and less likely to be educated compared with subjects without the MetS. In addition, subjects with the MetS had higher intakes of energy from carbohydrate as well as lower intakes of energy from fat compared with subjects without the MetS.\n\nCharacteristics of subjects according to soft drink consumption\n\nTable 2 shows the characteristics of subjects according to soft drink consumption. Men consuming \u22654 servings/week of soft drinks were younger, more likely to have higher income, more likely to be educated and current smokers compared with rare consumers (none or rarely). Women consuming \u22654 servings/week of soft drinks were younger and more likely to have higher income, more likely to be educated and current drinkers compared with rare consumers. Regardless of sex, frequent consumers had higher intakes of energy, mostly energy from fat, along with lower intakes of energy from carbohydrates compared with rare consumers.\n\nAssociation between soft drink consumption and incidence of the metabolic syndrome and its individual components\n\nThe relative risks for the MetS and its components according to soft drink consumption by sex are shown in Table 3. There was no association between soft drink consumption and risk of the MetS or its components in men. The risk of incident MetS increased significantly in women consuming \u22654 servings/week of soft drinks compared with rare consumers after adjustment for potential confounding factors such as age, income, education, smoking status, alcohol intake, physical activity, BMI, intakes of energy, percentage of fat, fibre intake and the presence of diseases (HR 1\u00b782; 95 % CI 1\u00b724, 2\u00b768, Pfor trend = 0\u00b70005). Women consuming \u22654 servings/week of soft drinks also had greater risks of elevated BP (HR 1\u00b797; 95 % CI 1\u00b723, 3\u00b714, Pfor trend = 0\u00b70242) and hypertriacylglycerolaemia (HR 1\u00b790; 95 % CI 1\u00b722, 2\u00b796, Pfor trend = 0\u00b70030) compared with rare consumers after adjustment for confounders.\n\nThe relative risks for the MetS and its components according to soft drink consumption and area of residence are shown in Table 4. There was no association between soft drink consumption and risk of the MetS or its components either in the rural area or in the urban area in men (data not shown). In addition, in women living in the rural area, there was no association between soft drink consumption and risk of the MetS and its components. However, in women living in the urban area, the risk of incident MetS increased significantly with consuming \u22654 servings/week of soft drinks compared with rare consumers after adjustment for potential confounding factors such as age, income, education, smoking status, alcohol intake, physical activity, BMI, intakes of energy, percentage of fat, fibre intake and the presence of diseases (HR 1\u00b771; 95 % CI 0\u00b784, 3\u00b747, Pfor trend = 0\u00b70021). Women living in the urban area and consuming \u22654 servings/week of soft drinks also had greater risks of hypertriacylglycerolaemia (HR 1\u00b776; 95 % CI 0\u00b783, 3\u00b774, Pfor trend = 0\u00b70004) compared with rare consumers after adjustment for confounders.",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "4"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Soft drink consumption is associated with increased incidence of the metabolic syndrome only in women.",
        "abstract": "Prospective studies on the association between soft drink consumption and incident risk of the metabolic syndrome (MetS) have not been carried out in Asians. We explored the sex-specific association between soft drink consumption and incident risk of the MetS in Korean adults during 10 years of follow-up. A total of 5797 subjects who were free of the MetS at baseline were studied. Soft drink consumption was assessed using a semi-quantitative FFQ. Time-dependent Cox proportional hazard model was used to examine hazard ratios (HR) of incidence of the MetS and its components in relation to soft drink consumption. In women, the multivariable-adjusted HR for developing the MetS was 1\u00b78-fold higher in frequent consumers of soft drinks (\u22654 servings/week) compared with rare consumers (95 % CI 1\u00b723, 2\u00b764). The adjusted HR for elevated blood pressure increased by 2-fold (95 % CI 1\u00b724, 3\u00b714) and for hypertriacylglycerolaemia by 1\u00b79-fold (95 % CI 1\u00b719, 2\u00b788) in frequent consumers of soft drinks compared with rare consumers. However, in men, there was no association between soft drink consumption and incident risk of the MetS or its components. Frequent soft drink consumption was associated with increased risk of developing the MetS and its components only in middle-aged Korean women, suggesting sex differences for the risk of the MetS related to diet.",
        "methods": "## Methods\n\n### Study subjects\n\nThe KoGES, a large-scale, community-based cohort study, was initiated in 2001. Initially, the target population consisted of 10 030 Korean adults aged 40\u201369 years and living in Ansan (urban) and Ansung (rural) areas. All participants responded to a baseline examination (2001\u20132002) with questionnaires including demographic information, socio-economic status, lifestyle, dietary intake, medical history, health examinations and biochemical measurements. Follow-up examinations were performed every 2 years over a 10-year period (2009\u20132010). The present study was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention (KCDC). Informed written consent was obtained from all participants. Of the original 10030 participants, participants without the MetS (n 7053) at baseline were included. We excluded participants who refused to participate in follow-up examinations (n 747), those who provided insufficient information (n 313), those who did not respond to dietary examination (n 59) and those who had CVD or cancer (n 137). After exclusion, a total of 5797 participants (3027 men and 2770 women) were included in the analysis during the 10-year period. A follow-up rate of 63\u00b73 % was achieved, resulting in 33269 person-years accrued. The average follow-up period was 5\u00b77 years (68\u00b78 months).\n\n### Dietary assessment\n\nTrained dietitians examined dietary intake both at baseline and at the second follow-up examination (2005\u20132006) using a 103-item, semi-quantitative FFQ. The FFQ was developed and validated by the KCDC. In the FFQ, participants reported the frequency and portion sizes of soft drink consumption during the past year. Soft drink consumption was estimated by the questions \u2018How often do you consume soft drinks (carbonated beverages, e.g., Cola and Sprite)?\u2019 for frequency and \u2018How much soft drink do you consume at once?\u2019 for portion size. The original responses for frequency included none or rarely, once/month, two to three times per month, one to two times per week, three to four times per week, five to six times per week, one to two times per d, three to four times per d and \u22655 times/d. The response options for portion size were as follows: 1/2 cup (100ml), one cup (200ml \u2013 one serving) and two or more cups (\u2265400ml). For analysis, the nine response items were converted to frequency per week, and the serving per week was calculated on the basis of frequency and portion size (mean frequency/week \u00d7 0\u00b75 for 1/2 cup, \u00d71\u00b70 for one cup, \u00d72\u00b70 for two or more cups) and then categorised into four groups: none or rarely, <1 serving/week, \u22651 serving/week to <4 servings/week and \u22654 servings/week. Nutrient intakes were measured using a database developed by the Rural Development and Administration.\n\n### Health examination\n\nTrained professionals conducted a comprehensive health examination under the KCDC protocol. Height and body weight were measured to the nearest 0\u00b71 cm and 0\u00b71kg, respectively, with no shoes and while wearing light clothing. BMI was defined as weight (kg)/height squared (m[2]). Waist circumference (WC) measurements were repeated three times, and then averaged after measuring to the nearest 0\u00b71cm at the narrowest point between the lowest rib and the right iliac crest. Blood pressure (BP) was measured after participants had rested for more than 5min in a sitting position (Baumanometer-Standby; W.A. Baum Co. Inc.). Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were calculated at Korotkoff phase I and Korotkoff phase V, respectively. Averaged values of left and right arms were measured a few times at 30-s intervals. Blood samples were collected to measure fasting blood glucose (FBG), TAG and HDL-cholesterol. The blood samples were collected after at least 8 h of fasting at baseline and during every follow-up examination. The concentrations of glucose, TAG and HDL-cholesterol in plasma were enzymatically measured using an autoanalyzer (ADVIA 1650; Bayer HealthCare). Incidence of the MetS was diagnosed at the follow-up examination by a physician.\n\n### Definition of the metabolic syndrome\n\nThe MetS was diagnosed using criteria based on the National Cholesterol Education Program Adult Treatment Panel III definition. The MetS is defined as the presence of more than three of the following indicators: (1) abdominal obesity (WC \u2265 90cm for men or \u226580cm for women), (2) elevated BP (SBP \u2265 130mmHg or DBP \u2265 85mmHg), use of antihypertensive medication, hypertension diagnosis by a physician, (3) high blood glucose (FBG \u2265 5.6 mmol/l), current use of insulin or oral hypoglycaemic medication, diabetes diagnosis by a physician, (4) hypertriacylglycerolaemia (TAG \u22651.7mmol/l) and (5) low HDL-cholesterol (HDL-cholesterol <1.0mmol/l in men or <1.3 mmol/l in women).\n\n### Other measurements\n\nDemographic characteristics, socio-economic status and lifestyle factors were examined from the baseline questionnaires. Income level was divided into four groups: <1 million KRW, 1\u20132 million KRW, 2\u20133 million KRW and \u22653 million KRW. Education level was categorised into three groups: \u22646 years (elementary school level), 7\u201312 years (middle/high school level) and >12 years (college level) of education. Smoking status was classified as follows: non-smoker, former smoker or current smoker. Alcohol intake was classified as follows: non-drinker, former drinker or current drinker. Physical activity was self-reported by a questionnaire. The questionnaire asked how many hours per day was spent performing physical activities according to its intensity (sedentary, very light, light, moderate, vigorous). Physical activities are presented as metabolic equivalents.\n\n### Statistical analysis\n\nAll the data were analysed using SAS software version 9.3 (SAS Institute). The results are expressed as percentages (categorical) or as means with their standard deviations (continuous). Differences in baseline characteristics by sex were examined by the \u03c7[2] test for categorical variables or Student\u2019s t test for continuous variables. Differences in characteristics across soft drink consumption were evaluated by the \u03c7[2] tests or a generalised linear model with post hoc Tukey\u2019s honestly significant difference test. Time-dependent Cox proportional hazard models were used as a method of survival analysis to examine the hazard ratios (HR) and 95 % CI for the incident risk of the MetS and its individual components according to soft drink consumption. Survival analysis was performed separately in men and women. In multivariable adjusted models, model 1 was adjusted for age and model 2 was adjusted for age, income level, education level, alcohol consumption, smoking status, physical activity, BMI, energy intake, percentage of fat from energy, fibre intake and the presence of diseases (diabetes and hypertension). For the selection of variables for adjustment in the multivariable model, potential confounders from the previously published scientific literature were taken into account with the statistical approach, such as stepwise procedures or comparing adjusted and unadjusted effect estimates. Individuals with hypertension or diabetes at baseline were excluded from the survival analysis to examine the risk of incident elevated BP and high blood glucose, respectively. The proportional hazards assumption was assessed graphically using log\u2013log plots and statistically using Schoenfeld\u2019s residuals. No violation of the proportional hazard assumption was detected. Tests for linear trends were performed on the basis of the median value of each category. All P values < 0\u00b705 were considered statistically significant.",
        "results": "## Results\n\nTable 1. Characteristics of the study subjects at baseline (Numbers and percentages (categorised variables); mean values and standard deviations (continuous variables))\n\nMetS Non-MetS  \n(n 2129) (n 3668)  \n\nCharacteristics of subjects at baseline\n\nAmong 5797 subjects, a total of 2129 (1046 men and 1083 women) developed the MetS. Table 1 shows the baseline characteristics of subjects by the presence of the MetS. Subjects with the MetS were older, more likely to be women, to live in the rural areas, to be exercisers, to have lower income and less likely to be educated compared with subjects without the MetS. In addition, subjects with the MetS had higher intakes of energy from carbohydrate as well as lower intakes of energy from fat compared with subjects without the MetS.\n\nCharacteristics of subjects according to soft drink consumption\n\nTable 2 shows the characteristics of subjects according to soft drink consumption. Men consuming \u22654 servings/week of soft drinks were younger, more likely to have higher income, more likely to be educated and current smokers compared with rare consumers (none or rarely). Women consuming \u22654 servings/week of soft drinks were younger and more likely to have higher income, more likely to be educated and current drinkers compared with rare consumers. Regardless of sex, frequent consumers had higher intakes of energy, mostly energy from fat, along with lower intakes of energy from carbohydrates compared with rare consumers.\n\nAssociation between soft drink consumption and incidence of the metabolic syndrome and its individual components\n\nThe relative risks for the MetS and its components according to soft drink consumption by sex are shown in Table 3. There was no association between soft drink consumption and risk of the MetS or its components in men. The risk of incident MetS increased significantly in women consuming \u22654 servings/week of soft drinks compared with rare consumers after adjustment for potential confounding factors such as age, income, education, smoking status, alcohol intake, physical activity, BMI, intakes of energy, percentage of fat, fibre intake and the presence of diseases (HR 1\u00b782; 95 % CI 1\u00b724, 2\u00b768, Pfor trend = 0\u00b70005). Women consuming \u22654 servings/week of soft drinks also had greater risks of elevated BP (HR 1\u00b797; 95 % CI 1\u00b723, 3\u00b714, Pfor trend = 0\u00b70242) and hypertriacylglycerolaemia (HR 1\u00b790; 95 % CI 1\u00b722, 2\u00b796, Pfor trend = 0\u00b70030) compared with rare consumers after adjustment for confounders.\n\nThe relative risks for the MetS and its components according to soft drink consumption and area of residence are shown in Table 4. There was no association between soft drink consumption and risk of the MetS or its components either in the rural area or in the urban area in men (data not shown). In addition, in women living in the rural area, there was no association between soft drink consumption and risk of the MetS and its components. However, in women living in the urban area, the risk of incident MetS increased significantly with consuming \u22654 servings/week of soft drinks compared with rare consumers after adjustment for potential confounding factors such as age, income, education, smoking status, alcohol intake, physical activity, BMI, intakes of energy, percentage of fat, fibre intake and the presence of diseases (HR 1\u00b771; 95 % CI 0\u00b784, 3\u00b747, Pfor trend = 0\u00b70021). Women living in the urban area and consuming \u22654 servings/week of soft drinks also had greater risks of hypertriacylglycerolaemia (HR 1\u00b776; 95 % CI 0\u00b783, 3\u00b774, Pfor trend = 0\u00b70004) compared with rare consumers after adjustment for confounders.",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Soft drink consumption is associated with increased incidence of the metabolic syndrome only in women.",
        "abstract": "Prospective studies on the association between soft drink consumption and incident risk of the metabolic syndrome (MetS) have not been carried out in Asians. We explored the sex-specific association between soft drink consumption and incident risk of the MetS in Korean adults during 10 years of follow-up. A total of 5797 subjects who were free of the MetS at baseline were studied. Soft drink consumption was assessed using a semi-quantitative FFQ. Time-dependent Cox proportional hazard model was used to examine hazard ratios (HR) of incidence of the MetS and its components in relation to soft drink consumption. In women, the multivariable-adjusted HR for developing the MetS was 1\u00b78-fold higher in frequent consumers of soft drinks (\u22654 servings/week) compared with rare consumers (95 % CI 1\u00b723, 2\u00b764). The adjusted HR for elevated blood pressure increased by 2-fold (95 % CI 1\u00b724, 3\u00b714) and for hypertriacylglycerolaemia by 1\u00b79-fold (95 % CI 1\u00b719, 2\u00b788) in frequent consumers of soft drinks compared with rare consumers. However, in men, there was no association between soft drink consumption and incident risk of the MetS or its components. Frequent soft drink consumption was associated with increased risk of developing the MetS and its components only in middle-aged Korean women, suggesting sex differences for the risk of the MetS related to diet.",
        "methods": "## Methods\n\n### Study subjects\n\nThe KoGES, a large-scale, community-based cohort study, was initiated in 2001. Initially, the target population consisted of 10 030 Korean adults aged 40\u201369 years and living in Ansan (urban) and Ansung (rural) areas. All participants responded to a baseline examination (2001\u20132002) with questionnaires including demographic information, socio-economic status, lifestyle, dietary intake, medical history, health examinations and biochemical measurements. Follow-up examinations were performed every 2 years over a 10-year period (2009\u20132010). The present study was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention (KCDC). Informed written consent was obtained from all participants. Of the original 10030 participants, participants without the MetS (n 7053) at baseline were included. We excluded participants who refused to participate in follow-up examinations (n 747), those who provided insufficient information (n 313), those who did not respond to dietary examination (n 59) and those who had CVD or cancer (n 137). After exclusion, a total of 5797 participants (3027 men and 2770 women) were included in the analysis during the 10-year period. A follow-up rate of 63\u00b73 % was achieved, resulting in 33269 person-years accrued. The average follow-up period was 5\u00b77 years (68\u00b78 months).\n\n### Dietary assessment\n\nTrained dietitians examined dietary intake both at baseline and at the second follow-up examination (2005\u20132006) using a 103-item, semi-quantitative FFQ. The FFQ was developed and validated by the KCDC. In the FFQ, participants reported the frequency and portion sizes of soft drink consumption during the past year. Soft drink consumption was estimated by the questions \u2018How often do you consume soft drinks (carbonated beverages, e.g., Cola and Sprite)?\u2019 for frequency and \u2018How much soft drink do you consume at once?\u2019 for portion size. The original responses for frequency included none or rarely, once/month, two to three times per month, one to two times per week, three to four times per week, five to six times per week, one to two times per d, three to four times per d and \u22655 times/d. The response options for portion size were as follows: 1/2 cup (100ml), one cup (200ml \u2013 one serving) and two or more cups (\u2265400ml). For analysis, the nine response items were converted to frequency per week, and the serving per week was calculated on the basis of frequency and portion size (mean frequency/week \u00d7 0\u00b75 for 1/2 cup, \u00d71\u00b70 for one cup, \u00d72\u00b70 for two or more cups) and then categorised into four groups: none or rarely, <1 serving/week, \u22651 serving/week to <4 servings/week and \u22654 servings/week. Nutrient intakes were measured using a database developed by the Rural Development and Administration.\n\n### Health examination\n\nTrained professionals conducted a comprehensive health examination under the KCDC protocol. Height and body weight were measured to the nearest 0\u00b71 cm and 0\u00b71kg, respectively, with no shoes and while wearing light clothing. BMI was defined as weight (kg)/height squared (m[2]). Waist circumference (WC) measurements were repeated three times, and then averaged after measuring to the nearest 0\u00b71cm at the narrowest point between the lowest rib and the right iliac crest. Blood pressure (BP) was measured after participants had rested for more than 5min in a sitting position (Baumanometer-Standby; W.A. Baum Co. Inc.). Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were calculated at Korotkoff phase I and Korotkoff phase V, respectively. Averaged values of left and right arms were measured a few times at 30-s intervals. Blood samples were collected to measure fasting blood glucose (FBG), TAG and HDL-cholesterol. The blood samples were collected after at least 8 h of fasting at baseline and during every follow-up examination. The concentrations of glucose, TAG and HDL-cholesterol in plasma were enzymatically measured using an autoanalyzer (ADVIA 1650; Bayer HealthCare). Incidence of the MetS was diagnosed at the follow-up examination by a physician.\n\n### Definition of the metabolic syndrome\n\nThe MetS was diagnosed using criteria based on the National Cholesterol Education Program Adult Treatment Panel III definition. The MetS is defined as the presence of more than three of the following indicators: (1) abdominal obesity (WC \u2265 90cm for men or \u226580cm for women), (2) elevated BP (SBP \u2265 130mmHg or DBP \u2265 85mmHg), use of antihypertensive medication, hypertension diagnosis by a physician, (3) high blood glucose (FBG \u2265 5.6 mmol/l), current use of insulin or oral hypoglycaemic medication, diabetes diagnosis by a physician, (4) hypertriacylglycerolaemia (TAG \u22651.7mmol/l) and (5) low HDL-cholesterol (HDL-cholesterol <1.0mmol/l in men or <1.3 mmol/l in women).\n\n### Other measurements\n\nDemographic characteristics, socio-economic status and lifestyle factors were examined from the baseline questionnaires. Income level was divided into four groups: <1 million KRW, 1\u20132 million KRW, 2\u20133 million KRW and \u22653 million KRW. Education level was categorised into three groups: \u22646 years (elementary school level), 7\u201312 years (middle/high school level) and >12 years (college level) of education. Smoking status was classified as follows: non-smoker, former smoker or current smoker. Alcohol intake was classified as follows: non-drinker, former drinker or current drinker. Physical activity was self-reported by a questionnaire. The questionnaire asked how many hours per day was spent performing physical activities according to its intensity (sedentary, very light, light, moderate, vigorous). Physical activities are presented as metabolic equivalents.\n\n### Statistical analysis\n\nAll the data were analysed using SAS software version 9.3 (SAS Institute). The results are expressed as percentages (categorical) or as means with their standard deviations (continuous). Differences in baseline characteristics by sex were examined by the \u03c7[2] test for categorical variables or Student\u2019s t test for continuous variables. Differences in characteristics across soft drink consumption were evaluated by the \u03c7[2] tests or a generalised linear model with post hoc Tukey\u2019s honestly significant difference test. Time-dependent Cox proportional hazard models were used as a method of survival analysis to examine the hazard ratios (HR) and 95 % CI for the incident risk of the MetS and its individual components according to soft drink consumption. Survival analysis was performed separately in men and women. In multivariable adjusted models, model 1 was adjusted for age and model 2 was adjusted for age, income level, education level, alcohol consumption, smoking status, physical activity, BMI, energy intake, percentage of fat from energy, fibre intake and the presence of diseases (diabetes and hypertension). For the selection of variables for adjustment in the multivariable model, potential confounders from the previously published scientific literature were taken into account with the statistical approach, such as stepwise procedures or comparing adjusted and unadjusted effect estimates. Individuals with hypertension or diabetes at baseline were excluded from the survival analysis to examine the risk of incident elevated BP and high blood glucose, respectively. The proportional hazards assumption was assessed graphically using log\u2013log plots and statistically using Schoenfeld\u2019s residuals. No violation of the proportional hazard assumption was detected. Tests for linear trends were performed on the basis of the median value of each category. All P values < 0\u00b705 were considered statistically significant.",
        "results": "## Results\n\nTable 1. Characteristics of the study subjects at baseline (Numbers and percentages (categorised variables); mean values and standard deviations (continuous variables))\n\nMetS Non-MetS  \n(n 2129) (n 3668)  \n\nCharacteristics of subjects at baseline\n\nAmong 5797 subjects, a total of 2129 (1046 men and 1083 women) developed the MetS. Table 1 shows the baseline characteristics of subjects by the presence of the MetS. Subjects with the MetS were older, more likely to be women, to live in the rural areas, to be exercisers, to have lower income and less likely to be educated compared with subjects without the MetS. In addition, subjects with the MetS had higher intakes of energy from carbohydrate as well as lower intakes of energy from fat compared with subjects without the MetS.\n\nCharacteristics of subjects according to soft drink consumption\n\nTable 2 shows the characteristics of subjects according to soft drink consumption. Men consuming \u22654 servings/week of soft drinks were younger, more likely to have higher income, more likely to be educated and current smokers compared with rare consumers (none or rarely). Women consuming \u22654 servings/week of soft drinks were younger and more likely to have higher income, more likely to be educated and current drinkers compared with rare consumers. Regardless of sex, frequent consumers had higher intakes of energy, mostly energy from fat, along with lower intakes of energy from carbohydrates compared with rare consumers.\n\nAssociation between soft drink consumption and incidence of the metabolic syndrome and its individual components\n\nThe relative risks for the MetS and its components according to soft drink consumption by sex are shown in Table 3. There was no association between soft drink consumption and risk of the MetS or its components in men. The risk of incident MetS increased significantly in women consuming \u22654 servings/week of soft drinks compared with rare consumers after adjustment for potential confounding factors such as age, income, education, smoking status, alcohol intake, physical activity, BMI, intakes of energy, percentage of fat, fibre intake and the presence of diseases (HR 1\u00b782; 95 % CI 1\u00b724, 2\u00b768, Pfor trend = 0\u00b70005). Women consuming \u22654 servings/week of soft drinks also had greater risks of elevated BP (HR 1\u00b797; 95 % CI 1\u00b723, 3\u00b714, Pfor trend = 0\u00b70242) and hypertriacylglycerolaemia (HR 1\u00b790; 95 % CI 1\u00b722, 2\u00b796, Pfor trend = 0\u00b70030) compared with rare consumers after adjustment for confounders.\n\nThe relative risks for the MetS and its components according to soft drink consumption and area of residence are shown in Table 4. There was no association between soft drink consumption and risk of the MetS or its components either in the rural area or in the urban area in men (data not shown). In addition, in women living in the rural area, there was no association between soft drink consumption and risk of the MetS and its components. However, in women living in the urban area, the risk of incident MetS increased significantly with consuming \u22654 servings/week of soft drinks compared with rare consumers after adjustment for potential confounding factors such as age, income, education, smoking status, alcohol intake, physical activity, BMI, intakes of energy, percentage of fat, fibre intake and the presence of diseases (HR 1\u00b771; 95 % CI 0\u00b784, 3\u00b747, Pfor trend = 0\u00b70021). Women living in the urban area and consuming \u22654 servings/week of soft drinks also had greater risks of hypertriacylglycerolaemia (HR 1\u00b776; 95 % CI 0\u00b783, 3\u00b774, Pfor trend = 0\u00b70004) compared with rare consumers after adjustment for confounders.",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Dairy consumption is associated with a lower incidence of the metabolic syndrome in middle-aged and older Korean adults: the Korean Genome and Epidemiology Study (KoGES).",
        "abstract": "This cohort study examined the association between total and individual dairy products and the risk of developing the metabolic syndrome (MetS) and its components in Korean adults from the Korean Genome and Epidemiology Study. We prospectively analysed 5510 participants aged 40-69 years without the MetS at baseline during a 10-year follow-up period. Dairy consumption was assessed with a semi-quantitative FFQ at baseline and after 4 years. The MetS was defined according to the criteria by the National Cholesterol Education Program Adult Treatment Panel III. The Cox's proportional hazard model was used to examine the association between consumption of total dairy products, milk and yogurt in servings per week and the risk of incident MetS or individual components. A total of 2103 subjects developed the MetS (38\u00b72 %) during an average follow-up of 67\u00b74 months (range 17-104 months). Frequent dairy consumption (>7 servings of total dairy and milk/week, \u22654 servings of yogurt/week) was associated with a reduced risk of incident MetS and its components. In the multivariable adjusted model, hazard ratios for the MetS were 0\u00b751 (95 % CI 0\u00b743, 0\u00b761) for total dairy products, 0\u00b750 (95 % CI 0\u00b738, 0\u00b766) for milk and 0\u00b767 (95 % CI 0\u00b757, 0\u00b778) for yogurt in frequent consumers compared with non-consumers. An inverse association between milk/yogurt and low HDL-cholesterol was shown only in women. In conclusion, high consumption of individual dairy products including milk and yogurt as well as total dairy were associated with a reduced risk of incident MetS and individual components in Korean adults.",
        "methods": "## Methods\n\n### Study population\n\nThis study is a prospective, community-based cohort study with data collected from the KoGES. KoGES began in 2001 and sought to assess the effects of dietary, environmental and lifestyle determinants on the incidence of chronic diseases such as diabetes mellitus, hypertension and the MetS in the general Korean population. Detailed information on the study procedure was described in a previous report[(20)]. In brief, 10 030 Korean adults aged 40\u201369 years who lived in Ansan (urban) and Ansung (rural) were recruited at baseline between 2001 and 2002. All participants completed interviewer-administered questionnaires on demographic information, lifestyle including dietary habits, their health condition, and medical history, and also anthropometric measurements were acquired and biochemical tests were conducted biennially. The study protocol was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention, and informed written consent was obtained from all participants. The present study was based on the data from baseline through the fourth follow-up examination between 2009 and 2010. Of the original 10 030 participants, 2977 participants had the MetS at baseline, 747 participants refused to participate in the follow-up examinations, 137 participants had CVD or cancer and 659 participants did not complete the baseline or follow-up FFQ; these participants were excluded from the study. In our data, the ninety participants had very low or high energy intake (<2092 or 20 920 kJ/d (<500 or >5000 kcal/d)). We included the participants with unusually low or high energy intake in our analysis because the exclusion of ninety subjects did not affect the results. Thus, 5510 participants (2859 men and 2651 women) were entered into the analysis during the 10-year period from 2001 to 2010, and a follow-up rate of 66 % was achieved, resulting in 30 947 person-years accrued. The average follow-up period was 67\u00b74 months.\n\n### Dietary assessment\n\nAt baseline and the second follow-up examination, dietary intake was assessed by trained dietitians with a 103-item semi-quantitative FFQ, which has been validated[(21)]. Total dairy products included milk, yogurt and cheese. Participants were asked to report usual frequency and portion size of dairy consumption during the past year. The answer for frequency had nine options for each food (never or almost never, once per month, two to three times per month, one to two times per week, three to four times per week, five to six times per week, once per day, two times per day, three times per day). The answer for portion size had three options for each food (1/2 serving, 1 serving (standard), \u22652 servings). To enhance the accuracy of serving size, pictures of dairy products and cup models marked with 50, 100, 150 and 200 ml were used as reference materials for serving sizes. One serving was equal to 200 ml of milk, 130 ml of liquid yogurt, 150 ml of solid yogurt and 20 g of cheese[(12)]. For analysis, dairy consumption was converted to weekly frequencies and then multiplied by the reported portion sizes of each food. To estimate whether consuming more than a recommended level of dairy food group (>7 servings/week) affect the incidence of the MetS compared with consuming non-dairy food. Finally, dairy consumption was categorised into five groups (never, <1, 1 \u2264 to <4, 4 \u2264 to \u22647, >7 servings/week) for total dairy and milk and four groups (never, <1, 1 \u2264 to <4, \u22654 servings/week) for yogurt because the number of subjects consuming yogurt frequently was very small. The effect of cheese consumption on the incidence of the MetS was not examined because most of the participants never or rarely consumed cheese. For subjects who acquired the MetS or were censored between baseline and the second follow-up examination, intake of dairy products was evaluated based on the FFQ at baseline; for those who acquired the MetS or were censored after the second follow-up examination, intake of dairy products was evaluated based on the average of FFQ at baseline and FFQ at the second follow-up.\n\n### Measurements\n\nHeight (cm) and body weight (kg) were measured to the nearest 0\u00b71 cm or 0\u00b71 kg in light clothes without shoes. BMI was calculated as weight (kg) divided by height squared (m[2]). Waist circumference was measured to the nearest 0\u00b71 cm at the narrowest point between the lowest rib and iliac crest, and the average of three repeated measurements was used in the analysis. Blood pressure was measured with a Baumanometer mercury sphygmomanometer (W.A. Baum) after subjects had rested for 5 min in a sitting position. Systolic and diastolic blood pressures were measured at phases I and V Korotkoff sound, respectively, and were determined as the average of the both arm readings. The blood samples were collected after at least 8 h of fasting state and plasma was separated for biochemical measurements. The concentrations of glucose, TAG and HDL-cholesterol in plasma were measured enzymatically using an autoanalyzer (ADIVA 1650; Bayer HealthCare).\n\n### Definition of the metabolic syndrome\n\nWe used the diagnostic criteria for the MetS based on the National Cholesterol Education Program Adult Treatment Panel III definition modified by the American Heart Association and the National Heart Lung and Blood Institute[(22)]. According to the criteria, the MetS was defined as the presence of three or more of the following: abdominal obesity (waist circumference \u226590 cm in men or \u226580 cm in women); elevated blood pressure (systolic blood pressure \u2265130 mmHg or diastolic blood pressure \u226585 mmHg or on antihypertensive medication or physician\u2019s diagnosis of hypertension); hyperglycaemia (fasting blood glucose \u22655\u00b76 mmol/l, current use of insulin, oral hypoglycaemic medication or physician\u2019s diagnosis of diabetes mellitus); hypertriacylglycerolaemia (TAG \u2265 1\u00b77 mmol/l); low HDL-cholesterol (HDL-cholesterol < 1\u00b70 mmol/l in men or <1\u00b73 mmol/l in women).\n\n### Covariates\n\nEducational levels were categorised into three groups: \u22646 years (elementary school graduates), 7\u2264 to \u226412 years (high school graduates) and >12 years (college graduates or more). Average monthly household income was categorised into four groups: <1 million KRW (approximately <850 USD at 2016), 1\u2264 to <2 million KRW (approximately 850\u2264 to <1700 USD), 2 \u2264 to <3 million KRW (approximately 1700\u2264 to <2500 USD) and \u22653 million KRW (approximately \u22652500 USD). Smoking status was classified as non-smoker, former smoker or current smoker. Alcohol consumption was classified into three groups: nondrinker, former drinker and current drinker. Physical activity was assessed using metabolic equivalent of task (MET) hours per day. To obtain MET, participants reported hours spent on sleep and five types of activities classified according to activity intensity including sedentary, very light, light, moderate and heavy. Total MET-hours per day were calculated by multiplying the reported hours spent by the MET values that were determined based on each type of activity[(23)]. Daily nutrient intakes such as energy, Ca and fibre were obtained from the semi-quantitative FFQ, and calculated on the basis of the fifth edition of Food Composition Table developed by Rural Development Administration of Korea[(24)].\n\n### Statistical analysis\n\nSex-stratified analyses were conducted to investigate the association between dairy consumption and the risks of the MetS and individual components of the MetS. To assess the statistical significance for trends of the MetS risk across categories, the median consumption values for each category were used as a continuous variable in the Cox\u2019s proportional hazard models. Statistical analysis were performed with SAS version 9.3 (SAS Institute). Baseline characteristics are expressed as percentage (categorical) or as means and standard deviation (continuous). Comparisons of variables at baseline across total dairy product consumption were analysed by the \u03c7[2] test for categorical data and by the generalised linear models with post hoc Tukey\u2019s HSD (honestly significant difference) test for continuous data. Multivariate adjusted Cox\u2019s proportional hazard models were conducted as a method for survival analysis to assess the hazard ratios (HR) and 95 % CI for incident MetS and its components during follow-up according to the consumption of total dairy products, milk and yogurt either in categorical form or in continuous form (dairy consumption as servings/week). We used an average of dairy products intake calculated from the two dietary measures instead of single measurement to minimise within-subject variation in diet and well represent long-term diet during follow-up[(25)]. The proportional hazard assumption that the ratio of hazards is proportional and constant over time was confirmed graphically using log-log plots[(26)] and statistically based on Schoenfeld\u2019s residuals[(27)] with no major violation of the assumption. In multivariable adjusted model, model 1 was adjusted for age, sex, BMI, residential location, educational level, household income, smoking status, alcohol intake and physical activity as covariates and model 2 was adjusted for covariates included in model 1 plus nutrient intakes such as energy and energy-adjusted Ca and fibre using the residual method because dietary Ca and fibre have been suggested to be associated with the risk of the MetS[(28,29)].",
        "results": "## Results\n\nAmong the 5510 participants, 2103 subjects developed the MetS during the 10-year follow-up. The baseline characteristics of the study participants according to total dairy consumption is shown in Table 1. Compared with non-dairy consumers, frequent dairy consumers (>7 servings/week) were more likely to be younger male residents of Ansan (urban area), be educated, to have higher income and were less likely to be current smokers and be physically active. Frequent dairy consumers had a lower waist circumference, systolic and diastolic blood pressures, and TAG level and a higher concentration of HDL-cholesterol compared with non-dairy consumers. In addition, frequent dairy consumers had a higher intake of non-dairy food group such as vegetables, fruits and meat as well as nutrient such as energy, energy from fat, Ca and fibre, and a lower intake of energy from carbohydrates.\n\nThe associations between total dairy consumption and incidence of the MetS and its components are summarised in Table 2. Frequent dairy consumers had a 39 % lower risk of the MetS than non-dairy consumers after adjusting for potential confounders such as age, sex, BMI, residential location, educational level, income, smoking status, alcohol intake and physical activity (HR 0\u00b761; 95 % CI 0\u00b752, 0\u00b771, Pfor trend < 0\u00b70001). The inverse association became even stronger after further adjustment for nutrient intake such as energy, Ca and fibre (HR 0\u00b751; 95 % CI 0\u00b743, 0\u00b761, Pfor trend < 0\u00b70001). Frequent total dairy consumption was associated with a reduced risk of individual MetS components (HR 0\u00b751; 95 % CI 0\u00b742, 0\u00b763, Pfor trend < 0\u00b70001 for elevated blood pressure; HR 0\u00b754; 95 % CI 0\u00b745, 0\u00b766, Pfor trend < 0\u00b70001 for hypertriacylglycerolaemia; HR 0\u00b751; 95 % CI 0\u00b743, 0\u00b761, Pfor trend < 0\u00b70001 for hyperglycaemia) compared with that of non-dairy consumers after adjusting for all confounding factors. We analysed sex-stratified association of total dairy consumption with the risk of MetS components. Frequent total dairy consumption was associated with lower risk of incident abdominal obesity (HR 0\u00b749; 95 % CI 0\u00b736, 0\u00b765, Pfor trend < 0\u00b70001 for men; HR 0\u00b745; 95 % CI 0\u00b733, 0\u00b760, Pfor trend < 0\u00b70001 for women) and low HDL-cholesterol (HR 0\u00b781; 95 % CI 0\u00b765, 1\u00b700, Pfor trend = 0\u00b70438 for men; HR 0\u00b744; 95 % CI 0\u00b735, 0\u00b754, Pfor trend < 0\u00b70001 for women) in both sexes. Except for abdominal obesity and low HDL-cholesterol, no sex difference between total dairy consumption and the risk of other MetS components was found (range of P values < 0\u00b70001\u20130\u00b70054).\n\nWe also observed associations between individual dairy products such as milk or yogurt and the risk of incident MetS and its components (Tables 3 and 4). Frequent milk consumers (>7 servings/week) had a 50 % reduction in the risk of developing the MetS (HR 0\u00b750; 95 % CI 0\u00b738, 0\u00b766, Pfor trend < 0\u00b70001) and individual components of the MetS such as elevated blood pressure (HR 0\u00b745; 95 % CI 0\u00b732, 0\u00b762, Pfor trend < 0\u00b70001) hypertriacylglycerolaemia (HR 0\u00b751; 95 % CI 0\u00b738, 0\u00b769, Pfor trend <0\u00b70001) and hyperglycaemia (HR 0\u00b744; 95 % CI 0\u00b733, 0\u00b759, Pfor trend < 0\u00b70001) after adjustment for confounding factors compared with non-milk consumers. In sex-stratified analysis, frequent milk consumption was associated with a reduced risk of abdominal obesity only in men (HR 0\u00b764; 95 % CI 0\u00b740, 1\u00b702, Pfor trend = 0\u00b70115) and of low HDL-cholesterol only in women (HR 0\u00b756; 95 % CI 0\u00b742, 0\u00b776, Pfor trend < 0\u00b70001) compared with that of non-milk consumers.\n\nSimilarly, frequent yogurt consumption (\u22654 servings/week) was associated with a 33 % reduction in the risk of developing the MetS (HR 0\u00b767; 95 % CI 0\u00b757, 0\u00b778, Pfor trend < 0\u00b70001), and individual components such as abdominal obesity (HR 0\u00b764; 95 % CI 0\u00b750, 0\u00b781, Pfor trend < 0\u00b70001), elevated blood pressure (HR 0\u00b770; 95 % CI 0\u00b759, 0\u00b784, Pfor trend < 0\u00b70001), hypertriacylglycerolaemia (HR 0\u00b768; 95 % CI 0\u00b758, 0\u00b780, Pfor trend < 0\u00b70001) and hyperglycaemia (HR 0\u00b772; 95 % CI 0\u00b762, 0\u00b784, Pfor trend < 0\u00b70001) compared with non-yogurt consumers. In sex-stratified analysis, the risk of low HDL-cholesterol decreased significantly according to increasing yogurt consumption only in women (HR 0\u00b761; 95 % CI 0\u00b751, 0\u00b773, Pfor trend < 0\u00b70001).\n\nWhen dairy consumption was included in the analysis as a continuous form, total dairy consumption was associated with a reduced risk of incident MetS (HR 0\u00b795; 95 % CI 0\u00b790, 0\u00b799, Pfor trend = 0\u00b70231), but it was not significantly associated with individual components of the MetS (range of P values 0\u00b70667\u20130\u00b79674, data not shown).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "4"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Dairy consumption is associated with a lower incidence of the metabolic syndrome in middle-aged and older Korean adults: the Korean Genome and Epidemiology Study (KoGES).",
        "abstract": "This cohort study examined the association between total and individual dairy products and the risk of developing the metabolic syndrome (MetS) and its components in Korean adults from the Korean Genome and Epidemiology Study. We prospectively analysed 5510 participants aged 40-69 years without the MetS at baseline during a 10-year follow-up period. Dairy consumption was assessed with a semi-quantitative FFQ at baseline and after 4 years. The MetS was defined according to the criteria by the National Cholesterol Education Program Adult Treatment Panel III. The Cox's proportional hazard model was used to examine the association between consumption of total dairy products, milk and yogurt in servings per week and the risk of incident MetS or individual components. A total of 2103 subjects developed the MetS (38\u00b72 %) during an average follow-up of 67\u00b74 months (range 17-104 months). Frequent dairy consumption (>7 servings of total dairy and milk/week, \u22654 servings of yogurt/week) was associated with a reduced risk of incident MetS and its components. In the multivariable adjusted model, hazard ratios for the MetS were 0\u00b751 (95 % CI 0\u00b743, 0\u00b761) for total dairy products, 0\u00b750 (95 % CI 0\u00b738, 0\u00b766) for milk and 0\u00b767 (95 % CI 0\u00b757, 0\u00b778) for yogurt in frequent consumers compared with non-consumers. An inverse association between milk/yogurt and low HDL-cholesterol was shown only in women. In conclusion, high consumption of individual dairy products including milk and yogurt as well as total dairy were associated with a reduced risk of incident MetS and individual components in Korean adults.",
        "methods": "## Methods\n\n### Study population\n\nThis study is a prospective, community-based cohort study with data collected from the KoGES. KoGES began in 2001 and sought to assess the effects of dietary, environmental and lifestyle determinants on the incidence of chronic diseases such as diabetes mellitus, hypertension and the MetS in the general Korean population. Detailed information on the study procedure was described in a previous report[(20)]. In brief, 10 030 Korean adults aged 40\u201369 years who lived in Ansan (urban) and Ansung (rural) were recruited at baseline between 2001 and 2002. All participants completed interviewer-administered questionnaires on demographic information, lifestyle including dietary habits, their health condition, and medical history, and also anthropometric measurements were acquired and biochemical tests were conducted biennially. The study protocol was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention, and informed written consent was obtained from all participants. The present study was based on the data from baseline through the fourth follow-up examination between 2009 and 2010. Of the original 10 030 participants, 2977 participants had the MetS at baseline, 747 participants refused to participate in the follow-up examinations, 137 participants had CVD or cancer and 659 participants did not complete the baseline or follow-up FFQ; these participants were excluded from the study. In our data, the ninety participants had very low or high energy intake (<2092 or 20 920 kJ/d (<500 or >5000 kcal/d)). We included the participants with unusually low or high energy intake in our analysis because the exclusion of ninety subjects did not affect the results. Thus, 5510 participants (2859 men and 2651 women) were entered into the analysis during the 10-year period from 2001 to 2010, and a follow-up rate of 66 % was achieved, resulting in 30 947 person-years accrued. The average follow-up period was 67\u00b74 months.\n\n### Dietary assessment\n\nAt baseline and the second follow-up examination, dietary intake was assessed by trained dietitians with a 103-item semi-quantitative FFQ, which has been validated[(21)]. Total dairy products included milk, yogurt and cheese. Participants were asked to report usual frequency and portion size of dairy consumption during the past year. The answer for frequency had nine options for each food (never or almost never, once per month, two to three times per month, one to two times per week, three to four times per week, five to six times per week, once per day, two times per day, three times per day). The answer for portion size had three options for each food (1/2 serving, 1 serving (standard), \u22652 servings). To enhance the accuracy of serving size, pictures of dairy products and cup models marked with 50, 100, 150 and 200 ml were used as reference materials for serving sizes. One serving was equal to 200 ml of milk, 130 ml of liquid yogurt, 150 ml of solid yogurt and 20 g of cheese[(12)]. For analysis, dairy consumption was converted to weekly frequencies and then multiplied by the reported portion sizes of each food. To estimate whether consuming more than a recommended level of dairy food group (>7 servings/week) affect the incidence of the MetS compared with consuming non-dairy food. Finally, dairy consumption was categorised into five groups (never, <1, 1 \u2264 to <4, 4 \u2264 to \u22647, >7 servings/week) for total dairy and milk and four groups (never, <1, 1 \u2264 to <4, \u22654 servings/week) for yogurt because the number of subjects consuming yogurt frequently was very small. The effect of cheese consumption on the incidence of the MetS was not examined because most of the participants never or rarely consumed cheese. For subjects who acquired the MetS or were censored between baseline and the second follow-up examination, intake of dairy products was evaluated based on the FFQ at baseline; for those who acquired the MetS or were censored after the second follow-up examination, intake of dairy products was evaluated based on the average of FFQ at baseline and FFQ at the second follow-up.\n\n### Measurements\n\nHeight (cm) and body weight (kg) were measured to the nearest 0\u00b71 cm or 0\u00b71 kg in light clothes without shoes. BMI was calculated as weight (kg) divided by height squared (m[2]). Waist circumference was measured to the nearest 0\u00b71 cm at the narrowest point between the lowest rib and iliac crest, and the average of three repeated measurements was used in the analysis. Blood pressure was measured with a Baumanometer mercury sphygmomanometer (W.A. Baum) after subjects had rested for 5 min in a sitting position. Systolic and diastolic blood pressures were measured at phases I and V Korotkoff sound, respectively, and were determined as the average of the both arm readings. The blood samples were collected after at least 8 h of fasting state and plasma was separated for biochemical measurements. The concentrations of glucose, TAG and HDL-cholesterol in plasma were measured enzymatically using an autoanalyzer (ADIVA 1650; Bayer HealthCare).\n\n### Definition of the metabolic syndrome\n\nWe used the diagnostic criteria for the MetS based on the National Cholesterol Education Program Adult Treatment Panel III definition modified by the American Heart Association and the National Heart Lung and Blood Institute[(22)]. According to the criteria, the MetS was defined as the presence of three or more of the following: abdominal obesity (waist circumference \u226590 cm in men or \u226580 cm in women); elevated blood pressure (systolic blood pressure \u2265130 mmHg or diastolic blood pressure \u226585 mmHg or on antihypertensive medication or physician\u2019s diagnosis of hypertension); hyperglycaemia (fasting blood glucose \u22655\u00b76 mmol/l, current use of insulin, oral hypoglycaemic medication or physician\u2019s diagnosis of diabetes mellitus); hypertriacylglycerolaemia (TAG \u2265 1\u00b77 mmol/l); low HDL-cholesterol (HDL-cholesterol < 1\u00b70 mmol/l in men or <1\u00b73 mmol/l in women).\n\n### Covariates\n\nEducational levels were categorised into three groups: \u22646 years (elementary school graduates), 7\u2264 to \u226412 years (high school graduates) and >12 years (college graduates or more). Average monthly household income was categorised into four groups: <1 million KRW (approximately <850 USD at 2016), 1\u2264 to <2 million KRW (approximately 850\u2264 to <1700 USD), 2 \u2264 to <3 million KRW (approximately 1700\u2264 to <2500 USD) and \u22653 million KRW (approximately \u22652500 USD). Smoking status was classified as non-smoker, former smoker or current smoker. Alcohol consumption was classified into three groups: nondrinker, former drinker and current drinker. Physical activity was assessed using metabolic equivalent of task (MET) hours per day. To obtain MET, participants reported hours spent on sleep and five types of activities classified according to activity intensity including sedentary, very light, light, moderate and heavy. Total MET-hours per day were calculated by multiplying the reported hours spent by the MET values that were determined based on each type of activity[(23)]. Daily nutrient intakes such as energy, Ca and fibre were obtained from the semi-quantitative FFQ, and calculated on the basis of the fifth edition of Food Composition Table developed by Rural Development Administration of Korea[(24)].\n\n### Statistical analysis\n\nSex-stratified analyses were conducted to investigate the association between dairy consumption and the risks of the MetS and individual components of the MetS. To assess the statistical significance for trends of the MetS risk across categories, the median consumption values for each category were used as a continuous variable in the Cox\u2019s proportional hazard models. Statistical analysis were performed with SAS version 9.3 (SAS Institute). Baseline characteristics are expressed as percentage (categorical) or as means and standard deviation (continuous). Comparisons of variables at baseline across total dairy product consumption were analysed by the \u03c7[2] test for categorical data and by the generalised linear models with post hoc Tukey\u2019s HSD (honestly significant difference) test for continuous data. Multivariate adjusted Cox\u2019s proportional hazard models were conducted as a method for survival analysis to assess the hazard ratios (HR) and 95 % CI for incident MetS and its components during follow-up according to the consumption of total dairy products, milk and yogurt either in categorical form or in continuous form (dairy consumption as servings/week). We used an average of dairy products intake calculated from the two dietary measures instead of single measurement to minimise within-subject variation in diet and well represent long-term diet during follow-up[(25)]. The proportional hazard assumption that the ratio of hazards is proportional and constant over time was confirmed graphically using log-log plots[(26)] and statistically based on Schoenfeld\u2019s residuals[(27)] with no major violation of the assumption. In multivariable adjusted model, model 1 was adjusted for age, sex, BMI, residential location, educational level, household income, smoking status, alcohol intake and physical activity as covariates and model 2 was adjusted for covariates included in model 1 plus nutrient intakes such as energy and energy-adjusted Ca and fibre using the residual method because dietary Ca and fibre have been suggested to be associated with the risk of the MetS[(28,29)].",
        "results": "## Results\n\nAmong the 5510 participants, 2103 subjects developed the MetS during the 10-year follow-up. The baseline characteristics of the study participants according to total dairy consumption is shown in Table 1. Compared with non-dairy consumers, frequent dairy consumers (>7 servings/week) were more likely to be younger male residents of Ansan (urban area), be educated, to have higher income and were less likely to be current smokers and be physically active. Frequent dairy consumers had a lower waist circumference, systolic and diastolic blood pressures, and TAG level and a higher concentration of HDL-cholesterol compared with non-dairy consumers. In addition, frequent dairy consumers had a higher intake of non-dairy food group such as vegetables, fruits and meat as well as nutrient such as energy, energy from fat, Ca and fibre, and a lower intake of energy from carbohydrates.\n\nThe associations between total dairy consumption and incidence of the MetS and its components are summarised in Table 2. Frequent dairy consumers had a 39 % lower risk of the MetS than non-dairy consumers after adjusting for potential confounders such as age, sex, BMI, residential location, educational level, income, smoking status, alcohol intake and physical activity (HR 0\u00b761; 95 % CI 0\u00b752, 0\u00b771, Pfor trend < 0\u00b70001). The inverse association became even stronger after further adjustment for nutrient intake such as energy, Ca and fibre (HR 0\u00b751; 95 % CI 0\u00b743, 0\u00b761, Pfor trend < 0\u00b70001). Frequent total dairy consumption was associated with a reduced risk of individual MetS components (HR 0\u00b751; 95 % CI 0\u00b742, 0\u00b763, Pfor trend < 0\u00b70001 for elevated blood pressure; HR 0\u00b754; 95 % CI 0\u00b745, 0\u00b766, Pfor trend < 0\u00b70001 for hypertriacylglycerolaemia; HR 0\u00b751; 95 % CI 0\u00b743, 0\u00b761, Pfor trend < 0\u00b70001 for hyperglycaemia) compared with that of non-dairy consumers after adjusting for all confounding factors. We analysed sex-stratified association of total dairy consumption with the risk of MetS components. Frequent total dairy consumption was associated with lower risk of incident abdominal obesity (HR 0\u00b749; 95 % CI 0\u00b736, 0\u00b765, Pfor trend < 0\u00b70001 for men; HR 0\u00b745; 95 % CI 0\u00b733, 0\u00b760, Pfor trend < 0\u00b70001 for women) and low HDL-cholesterol (HR 0\u00b781; 95 % CI 0\u00b765, 1\u00b700, Pfor trend = 0\u00b70438 for men; HR 0\u00b744; 95 % CI 0\u00b735, 0\u00b754, Pfor trend < 0\u00b70001 for women) in both sexes. Except for abdominal obesity and low HDL-cholesterol, no sex difference between total dairy consumption and the risk of other MetS components was found (range of P values < 0\u00b70001\u20130\u00b70054).\n\nWe also observed associations between individual dairy products such as milk or yogurt and the risk of incident MetS and its components (Tables 3 and 4). Frequent milk consumers (>7 servings/week) had a 50 % reduction in the risk of developing the MetS (HR 0\u00b750; 95 % CI 0\u00b738, 0\u00b766, Pfor trend < 0\u00b70001) and individual components of the MetS such as elevated blood pressure (HR 0\u00b745; 95 % CI 0\u00b732, 0\u00b762, Pfor trend < 0\u00b70001) hypertriacylglycerolaemia (HR 0\u00b751; 95 % CI 0\u00b738, 0\u00b769, Pfor trend <0\u00b70001) and hyperglycaemia (HR 0\u00b744; 95 % CI 0\u00b733, 0\u00b759, Pfor trend < 0\u00b70001) after adjustment for confounding factors compared with non-milk consumers. In sex-stratified analysis, frequent milk consumption was associated with a reduced risk of abdominal obesity only in men (HR 0\u00b764; 95 % CI 0\u00b740, 1\u00b702, Pfor trend = 0\u00b70115) and of low HDL-cholesterol only in women (HR 0\u00b756; 95 % CI 0\u00b742, 0\u00b776, Pfor trend < 0\u00b70001) compared with that of non-milk consumers.\n\nSimilarly, frequent yogurt consumption (\u22654 servings/week) was associated with a 33 % reduction in the risk of developing the MetS (HR 0\u00b767; 95 % CI 0\u00b757, 0\u00b778, Pfor trend < 0\u00b70001), and individual components such as abdominal obesity (HR 0\u00b764; 95 % CI 0\u00b750, 0\u00b781, Pfor trend < 0\u00b70001), elevated blood pressure (HR 0\u00b770; 95 % CI 0\u00b759, 0\u00b784, Pfor trend < 0\u00b70001), hypertriacylglycerolaemia (HR 0\u00b768; 95 % CI 0\u00b758, 0\u00b780, Pfor trend < 0\u00b70001) and hyperglycaemia (HR 0\u00b772; 95 % CI 0\u00b762, 0\u00b784, Pfor trend < 0\u00b70001) compared with non-yogurt consumers. In sex-stratified analysis, the risk of low HDL-cholesterol decreased significantly according to increasing yogurt consumption only in women (HR 0\u00b761; 95 % CI 0\u00b751, 0\u00b773, Pfor trend < 0\u00b70001).\n\nWhen dairy consumption was included in the analysis as a continuous form, total dairy consumption was associated with a reduced risk of incident MetS (HR 0\u00b795; 95 % CI 0\u00b790, 0\u00b799, Pfor trend = 0\u00b70231), but it was not significantly associated with individual components of the MetS (range of P values 0\u00b70667\u20130\u00b79674, data not shown).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Dairy consumption is associated with a lower incidence of the metabolic syndrome in middle-aged and older Korean adults: the Korean Genome and Epidemiology Study (KoGES).",
        "abstract": "This cohort study examined the association between total and individual dairy products and the risk of developing the metabolic syndrome (MetS) and its components in Korean adults from the Korean Genome and Epidemiology Study. We prospectively analysed 5510 participants aged 40-69 years without the MetS at baseline during a 10-year follow-up period. Dairy consumption was assessed with a semi-quantitative FFQ at baseline and after 4 years. The MetS was defined according to the criteria by the National Cholesterol Education Program Adult Treatment Panel III. The Cox's proportional hazard model was used to examine the association between consumption of total dairy products, milk and yogurt in servings per week and the risk of incident MetS or individual components. A total of 2103 subjects developed the MetS (38\u00b72 %) during an average follow-up of 67\u00b74 months (range 17-104 months). Frequent dairy consumption (>7 servings of total dairy and milk/week, \u22654 servings of yogurt/week) was associated with a reduced risk of incident MetS and its components. In the multivariable adjusted model, hazard ratios for the MetS were 0\u00b751 (95 % CI 0\u00b743, 0\u00b761) for total dairy products, 0\u00b750 (95 % CI 0\u00b738, 0\u00b766) for milk and 0\u00b767 (95 % CI 0\u00b757, 0\u00b778) for yogurt in frequent consumers compared with non-consumers. An inverse association between milk/yogurt and low HDL-cholesterol was shown only in women. In conclusion, high consumption of individual dairy products including milk and yogurt as well as total dairy were associated with a reduced risk of incident MetS and individual components in Korean adults.",
        "methods": "## Methods\n\n### Study population\n\nThis study is a prospective, community-based cohort study with data collected from the KoGES. KoGES began in 2001 and sought to assess the effects of dietary, environmental and lifestyle determinants on the incidence of chronic diseases such as diabetes mellitus, hypertension and the MetS in the general Korean population. Detailed information on the study procedure was described in a previous report[(20)]. In brief, 10 030 Korean adults aged 40\u201369 years who lived in Ansan (urban) and Ansung (rural) were recruited at baseline between 2001 and 2002. All participants completed interviewer-administered questionnaires on demographic information, lifestyle including dietary habits, their health condition, and medical history, and also anthropometric measurements were acquired and biochemical tests were conducted biennially. The study protocol was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention, and informed written consent was obtained from all participants. The present study was based on the data from baseline through the fourth follow-up examination between 2009 and 2010. Of the original 10 030 participants, 2977 participants had the MetS at baseline, 747 participants refused to participate in the follow-up examinations, 137 participants had CVD or cancer and 659 participants did not complete the baseline or follow-up FFQ; these participants were excluded from the study. In our data, the ninety participants had very low or high energy intake (<2092 or 20 920 kJ/d (<500 or >5000 kcal/d)). We included the participants with unusually low or high energy intake in our analysis because the exclusion of ninety subjects did not affect the results. Thus, 5510 participants (2859 men and 2651 women) were entered into the analysis during the 10-year period from 2001 to 2010, and a follow-up rate of 66 % was achieved, resulting in 30 947 person-years accrued. The average follow-up period was 67\u00b74 months.\n\n### Dietary assessment\n\nAt baseline and the second follow-up examination, dietary intake was assessed by trained dietitians with a 103-item semi-quantitative FFQ, which has been validated[(21)]. Total dairy products included milk, yogurt and cheese. Participants were asked to report usual frequency and portion size of dairy consumption during the past year. The answer for frequency had nine options for each food (never or almost never, once per month, two to three times per month, one to two times per week, three to four times per week, five to six times per week, once per day, two times per day, three times per day). The answer for portion size had three options for each food (1/2 serving, 1 serving (standard), \u22652 servings). To enhance the accuracy of serving size, pictures of dairy products and cup models marked with 50, 100, 150 and 200 ml were used as reference materials for serving sizes. One serving was equal to 200 ml of milk, 130 ml of liquid yogurt, 150 ml of solid yogurt and 20 g of cheese[(12)]. For analysis, dairy consumption was converted to weekly frequencies and then multiplied by the reported portion sizes of each food. To estimate whether consuming more than a recommended level of dairy food group (>7 servings/week) affect the incidence of the MetS compared with consuming non-dairy food. Finally, dairy consumption was categorised into five groups (never, <1, 1 \u2264 to <4, 4 \u2264 to \u22647, >7 servings/week) for total dairy and milk and four groups (never, <1, 1 \u2264 to <4, \u22654 servings/week) for yogurt because the number of subjects consuming yogurt frequently was very small. The effect of cheese consumption on the incidence of the MetS was not examined because most of the participants never or rarely consumed cheese. For subjects who acquired the MetS or were censored between baseline and the second follow-up examination, intake of dairy products was evaluated based on the FFQ at baseline; for those who acquired the MetS or were censored after the second follow-up examination, intake of dairy products was evaluated based on the average of FFQ at baseline and FFQ at the second follow-up.\n\n### Measurements\n\nHeight (cm) and body weight (kg) were measured to the nearest 0\u00b71 cm or 0\u00b71 kg in light clothes without shoes. BMI was calculated as weight (kg) divided by height squared (m[2]). Waist circumference was measured to the nearest 0\u00b71 cm at the narrowest point between the lowest rib and iliac crest, and the average of three repeated measurements was used in the analysis. Blood pressure was measured with a Baumanometer mercury sphygmomanometer (W.A. Baum) after subjects had rested for 5 min in a sitting position. Systolic and diastolic blood pressures were measured at phases I and V Korotkoff sound, respectively, and were determined as the average of the both arm readings. The blood samples were collected after at least 8 h of fasting state and plasma was separated for biochemical measurements. The concentrations of glucose, TAG and HDL-cholesterol in plasma were measured enzymatically using an autoanalyzer (ADIVA 1650; Bayer HealthCare).\n\n### Definition of the metabolic syndrome\n\nWe used the diagnostic criteria for the MetS based on the National Cholesterol Education Program Adult Treatment Panel III definition modified by the American Heart Association and the National Heart Lung and Blood Institute[(22)]. According to the criteria, the MetS was defined as the presence of three or more of the following: abdominal obesity (waist circumference \u226590 cm in men or \u226580 cm in women); elevated blood pressure (systolic blood pressure \u2265130 mmHg or diastolic blood pressure \u226585 mmHg or on antihypertensive medication or physician\u2019s diagnosis of hypertension); hyperglycaemia (fasting blood glucose \u22655\u00b76 mmol/l, current use of insulin, oral hypoglycaemic medication or physician\u2019s diagnosis of diabetes mellitus); hypertriacylglycerolaemia (TAG \u2265 1\u00b77 mmol/l); low HDL-cholesterol (HDL-cholesterol < 1\u00b70 mmol/l in men or <1\u00b73 mmol/l in women).\n\n### Covariates\n\nEducational levels were categorised into three groups: \u22646 years (elementary school graduates), 7\u2264 to \u226412 years (high school graduates) and >12 years (college graduates or more). Average monthly household income was categorised into four groups: <1 million KRW (approximately <850 USD at 2016), 1\u2264 to <2 million KRW (approximately 850\u2264 to <1700 USD), 2 \u2264 to <3 million KRW (approximately 1700\u2264 to <2500 USD) and \u22653 million KRW (approximately \u22652500 USD). Smoking status was classified as non-smoker, former smoker or current smoker. Alcohol consumption was classified into three groups: nondrinker, former drinker and current drinker. Physical activity was assessed using metabolic equivalent of task (MET) hours per day. To obtain MET, participants reported hours spent on sleep and five types of activities classified according to activity intensity including sedentary, very light, light, moderate and heavy. Total MET-hours per day were calculated by multiplying the reported hours spent by the MET values that were determined based on each type of activity[(23)]. Daily nutrient intakes such as energy, Ca and fibre were obtained from the semi-quantitative FFQ, and calculated on the basis of the fifth edition of Food Composition Table developed by Rural Development Administration of Korea[(24)].\n\n### Statistical analysis\n\nSex-stratified analyses were conducted to investigate the association between dairy consumption and the risks of the MetS and individual components of the MetS. To assess the statistical significance for trends of the MetS risk across categories, the median consumption values for each category were used as a continuous variable in the Cox\u2019s proportional hazard models. Statistical analysis were performed with SAS version 9.3 (SAS Institute). Baseline characteristics are expressed as percentage (categorical) or as means and standard deviation (continuous). Comparisons of variables at baseline across total dairy product consumption were analysed by the \u03c7[2] test for categorical data and by the generalised linear models with post hoc Tukey\u2019s HSD (honestly significant difference) test for continuous data. Multivariate adjusted Cox\u2019s proportional hazard models were conducted as a method for survival analysis to assess the hazard ratios (HR) and 95 % CI for incident MetS and its components during follow-up according to the consumption of total dairy products, milk and yogurt either in categorical form or in continuous form (dairy consumption as servings/week). We used an average of dairy products intake calculated from the two dietary measures instead of single measurement to minimise within-subject variation in diet and well represent long-term diet during follow-up[(25)]. The proportional hazard assumption that the ratio of hazards is proportional and constant over time was confirmed graphically using log-log plots[(26)] and statistically based on Schoenfeld\u2019s residuals[(27)] with no major violation of the assumption. In multivariable adjusted model, model 1 was adjusted for age, sex, BMI, residential location, educational level, household income, smoking status, alcohol intake and physical activity as covariates and model 2 was adjusted for covariates included in model 1 plus nutrient intakes such as energy and energy-adjusted Ca and fibre using the residual method because dietary Ca and fibre have been suggested to be associated with the risk of the MetS[(28,29)].",
        "results": "## Results\n\nAmong the 5510 participants, 2103 subjects developed the MetS during the 10-year follow-up. The baseline characteristics of the study participants according to total dairy consumption is shown in Table 1. Compared with non-dairy consumers, frequent dairy consumers (>7 servings/week) were more likely to be younger male residents of Ansan (urban area), be educated, to have higher income and were less likely to be current smokers and be physically active. Frequent dairy consumers had a lower waist circumference, systolic and diastolic blood pressures, and TAG level and a higher concentration of HDL-cholesterol compared with non-dairy consumers. In addition, frequent dairy consumers had a higher intake of non-dairy food group such as vegetables, fruits and meat as well as nutrient such as energy, energy from fat, Ca and fibre, and a lower intake of energy from carbohydrates.\n\nThe associations between total dairy consumption and incidence of the MetS and its components are summarised in Table 2. Frequent dairy consumers had a 39 % lower risk of the MetS than non-dairy consumers after adjusting for potential confounders such as age, sex, BMI, residential location, educational level, income, smoking status, alcohol intake and physical activity (HR 0\u00b761; 95 % CI 0\u00b752, 0\u00b771, Pfor trend < 0\u00b70001). The inverse association became even stronger after further adjustment for nutrient intake such as energy, Ca and fibre (HR 0\u00b751; 95 % CI 0\u00b743, 0\u00b761, Pfor trend < 0\u00b70001). Frequent total dairy consumption was associated with a reduced risk of individual MetS components (HR 0\u00b751; 95 % CI 0\u00b742, 0\u00b763, Pfor trend < 0\u00b70001 for elevated blood pressure; HR 0\u00b754; 95 % CI 0\u00b745, 0\u00b766, Pfor trend < 0\u00b70001 for hypertriacylglycerolaemia; HR 0\u00b751; 95 % CI 0\u00b743, 0\u00b761, Pfor trend < 0\u00b70001 for hyperglycaemia) compared with that of non-dairy consumers after adjusting for all confounding factors. We analysed sex-stratified association of total dairy consumption with the risk of MetS components. Frequent total dairy consumption was associated with lower risk of incident abdominal obesity (HR 0\u00b749; 95 % CI 0\u00b736, 0\u00b765, Pfor trend < 0\u00b70001 for men; HR 0\u00b745; 95 % CI 0\u00b733, 0\u00b760, Pfor trend < 0\u00b70001 for women) and low HDL-cholesterol (HR 0\u00b781; 95 % CI 0\u00b765, 1\u00b700, Pfor trend = 0\u00b70438 for men; HR 0\u00b744; 95 % CI 0\u00b735, 0\u00b754, Pfor trend < 0\u00b70001 for women) in both sexes. Except for abdominal obesity and low HDL-cholesterol, no sex difference between total dairy consumption and the risk of other MetS components was found (range of P values < 0\u00b70001\u20130\u00b70054).\n\nWe also observed associations between individual dairy products such as milk or yogurt and the risk of incident MetS and its components (Tables 3 and 4). Frequent milk consumers (>7 servings/week) had a 50 % reduction in the risk of developing the MetS (HR 0\u00b750; 95 % CI 0\u00b738, 0\u00b766, Pfor trend < 0\u00b70001) and individual components of the MetS such as elevated blood pressure (HR 0\u00b745; 95 % CI 0\u00b732, 0\u00b762, Pfor trend < 0\u00b70001) hypertriacylglycerolaemia (HR 0\u00b751; 95 % CI 0\u00b738, 0\u00b769, Pfor trend <0\u00b70001) and hyperglycaemia (HR 0\u00b744; 95 % CI 0\u00b733, 0\u00b759, Pfor trend < 0\u00b70001) after adjustment for confounding factors compared with non-milk consumers. In sex-stratified analysis, frequent milk consumption was associated with a reduced risk of abdominal obesity only in men (HR 0\u00b764; 95 % CI 0\u00b740, 1\u00b702, Pfor trend = 0\u00b70115) and of low HDL-cholesterol only in women (HR 0\u00b756; 95 % CI 0\u00b742, 0\u00b776, Pfor trend < 0\u00b70001) compared with that of non-milk consumers.\n\nSimilarly, frequent yogurt consumption (\u22654 servings/week) was associated with a 33 % reduction in the risk of developing the MetS (HR 0\u00b767; 95 % CI 0\u00b757, 0\u00b778, Pfor trend < 0\u00b70001), and individual components such as abdominal obesity (HR 0\u00b764; 95 % CI 0\u00b750, 0\u00b781, Pfor trend < 0\u00b70001), elevated blood pressure (HR 0\u00b770; 95 % CI 0\u00b759, 0\u00b784, Pfor trend < 0\u00b70001), hypertriacylglycerolaemia (HR 0\u00b768; 95 % CI 0\u00b758, 0\u00b780, Pfor trend < 0\u00b70001) and hyperglycaemia (HR 0\u00b772; 95 % CI 0\u00b762, 0\u00b784, Pfor trend < 0\u00b70001) compared with non-yogurt consumers. In sex-stratified analysis, the risk of low HDL-cholesterol decreased significantly according to increasing yogurt consumption only in women (HR 0\u00b761; 95 % CI 0\u00b751, 0\u00b773, Pfor trend < 0\u00b70001).\n\nWhen dairy consumption was included in the analysis as a continuous form, total dairy consumption was associated with a reduced risk of incident MetS (HR 0\u00b795; 95 % CI 0\u00b790, 0\u00b799, Pfor trend = 0\u00b70231), but it was not significantly associated with individual components of the MetS (range of P values 0\u00b70667\u20130\u00b79674, data not shown).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Association between fruit and vegetable consumption and risk of metabolic syndrome determined using the Korean Genome and Epidemiology Study (KoGES).",
        "abstract": "## PURPOSE\nThis study examined the association between fruit and vegetable consumption and the risk of metabolic syndrome (MetS) and its components in middle-aged and older adults using data from the Korean Genome and Epidemiology Study.\n## METHODS\nA total of 5688 participants aged 40-69\u00a0years without MetS at baseline were recruited and followed for 8\u00a0years. Fruit and vegetable consumption was assessed using a semi-quantitative food-frequency questionnaire at baseline and after 4\u00a0years. A multivariate Cox proportional hazards model was used to examine the risk of incident MetS and its components according to fruit and vegetable consumption.\n## RESULTS\nA total of 2067 participants (1020 men and 1047 women) developed MetS during the 8-year follow-up period. Frequent fruit consumers (\u2265\u20094 servings/day) had a lower risk of incident MetS than did rare consumers (<\u20091 serving/day) in both men and women (hazard ratio (HR)\u2009=\u20090.55, 95% CI 0.44-0.67, p\u2009<\u20090.0001 for men; HR\u2009=\u20090.57, 95% CI 0.47-0.70, p for trend\u2009<\u20090.0001 for women) after adjusting for confounders. Frequent fruit consumption was inversely associated with incidence of abdominal obesity, hypertriglyceridemia, and elevated blood pressure. Frequent vegetable consumption was inversely associated with a risk of hyperglycemia only in men (HR\u2009=\u20090.65, 95% CI 0.44-0.96, p for trend\u2009=\u20090.0275), but no association was observed with risk of incident MetS.\n## CONCLUSIONS\nHigh fruit consumption was associated with a reduced risk of incident MetS and its components, whereas high vegetable consumption was associated with a reduced risk of incident hyperglycemia, but not with MetS in middle-aged and older adults.",
        "methods": "### Methods\n\n#### Study population\n\nThe KoGES is a prospective study that explores dietary, environmental, and lifestyle factors and the incidence of chronic diseases in the general Korean population. The study consisted of 10,030 Korean adults aged 40\u201369 years who lived in one of two areas near central city in South Korea. Participants were recruited between 2001 and 2002, which is baseline, and follow-up examinations were conducted every 2 years over an 8-year period (2009\u20132010). All participants completed questionnaires regarding demographic information, lifestyle, medical history, and health conditions at baseline, and anthropometric measurements and biochemical tests were performed biennially. The study protocol was approved by the Institutional Review Boards of the Korea Centers for Disease Control and Prevention and Kyung Hee University (KHSIRB-16-022). Written informed consent was obtained from all participants. For the analysis of incident MetS, of the 10,030 participants, 2977 had MetS at baseline, 747 refused to participate in the follow-up examinations, 95 did not complete a food-frequency questionnaire (FFQ) at both baseline and the follow-up, 133 had cardiovascular disease (CVD) or cancer at baseline, 163 had extraordinary energy intake (< 500 kcal/day or > 6000 kcal/day) at baseline, and 227 had missing information on covariates or MetS parameters at baseline. All of the above participants were excluded from this study. Ultimately, a total of 5688 participants (2973 men and 2715 women) were included in the analysis of MetS. Likewise, for abdominal obesity analysis, 3962 who had abdominal obesity at baseline, 679 refused to participate in the follow-up examinations, 70 did not complete FFQ, 114 had CVD or cancer, 128 had extraordinary energy intake, and 206 had missing information were excluded. For hyperglycemia analysis, 1950 who had hyperglycemia at baseline, 820 refused to participate in the follow-up examinations, 83 did not complete FFQ, 165 had CVD or cancer, 202 had extraordinary energy intake, and 242 had missing information were excluded. For the hypertriglyceridemia, 3706 who had hypertriglyceridemia at baseline, 680 refused to participate in the follow-up examinations, 80 did not complete FFQ, 137 had CVD or cancer, 155 had extraordinary energy intake, and 266 had missing information were excluded. For low HDL-C analysis, 3781 who had low HDL-C at baseline, 649 refused to participate in the follow-up examinations, 123 did not complete FFQ, 143 had CVD or cancer, 139 had extraordinary energy intake, and 251 had missing information were excluded. For elevated blood pressure analysis, 4479 who had elevated blood pressure at baseline, 579 refused to participate in the follow-up examinations, 81 did not complete FFQ, 84 had CVD or cancer, 123 had extraordinary energy intake, and 205 had missing information were excluded. Finally, 4871 participants for abdominal obesity, 6568 participants for hyperglycemia, 5006 participants for hypertriglyceridemia, 4944 participants for low HDL-C, and 4479 participants for elevated blood pressure were included for the analysis of MetS components. A follow-up rate of 66.5% was achieved, accounting for 32,801 person-years. The average follow-up period was 70 months (range 17\u2013104 months).\n\n#### Dietary assessment\n\nAt baseline and second follow-up examinations (2005\u20132006), fruit and vegetable intakes were assessed with a semi-quantitative FFQ, which has been validated. For participants who acquired the MetS or were censored between baseline and the second follow-up examination, intake of fruit and vegetable was evaluated based on the FFQ at baseline; for those who acquired the MetS or were censored after the second follow-up examination, intake of fruit and vegetable was evaluated based on the average of FFQ at baseline and FFQ at the follow-up. The 12 types of fruits (tangerine, orange, persimmon or dried persimmon, watermelon, strawberry, grape, pear, oriental melon or melon, peach or prune, apple, banana, and tomato) and 20 types of vegetables (Korean cabbage, green pepper, red pepper leaf, spinach, lettuce, perilla leaf, chives or water parsley, leaf beet or Chinese mallow, radish or salted radish, deodeok or balloon flower, onion, cucumber, bean sprouts, carrot, pumpkin, zucchini, cabbage, fern or sweet potato stems, oyster mushroom, and black mushroom or button mushroom) were included in the FFQ. To examine the effect of fruits by subtype, fruits were classified into subgroups of citrus fruits (tangerines and oranges), non-citrus fruits (Persimmon/dried persimmon, watermelon, strawberry, grape, pear, oriental melon/melon, peach/prune, apple, and banana), and carotene-rich fruits (Tangerines, oranges, watermelon, persimmon, and tomato). In addition, to examine the effect of vegetables by subtype, vegetables were classified into subgroups of cruciferous vegetables (Korean cabbage, cabbage, radish, lettuce, and broccoli), green leafy vegetables (spinach, perilla leaf, chive/water parsley, and red pepper leaf), carotene-rich vegetables (carrots and pumpkin), and other vegetables (bean sprouts, fern/sweet potato stems, cucumber, onion, green chili, zucchini, deodeok/balloonflower, and mushrooms). In the FFQ, participants reported the usual frequency and portion size of fruit and vegetable consumption during the past year. The answer for frequency had nine options for each food (never or rarely, once/month, 2\u20133 times/month, 1\u20132 times/week, 3\u20134 times/week, 5\u20136 times/week, once/day, 2 times/day, and 3 times or more/day). The answer for portion size had three options for each food [\u00bd serving, 1 serving (standard), \u2265 2 servings]. One serving of fruit or vegetable was equal to 100 g or 70 g, respectively. For analysis, fruit and vegetable consumptions were converted to weekly frequencies and then multiplied by the reported portion size of each food. Finally, fruit and vegetable consumptions were classified into four groups (never or rarely, 1 to < 2 servings/day, 2 to < 4 servings/day, and \u2265 4 servings/day). Individual food intake was expressed as food consumption per week, and the total intake of each food was summed. In addition, seven meats (including pork, beef, chicken, and processed meat), refined grains (including white rice, noodles, and breads), whole grains (including barley, multigrain, and mixed grains), and dairy foods (including milk and yogurt) were evaluated. Nutrient intake was calculated using a database developed by the Rural Development Administration.\n\n#### Measurements\n\nHeight and body weight were measured in light clothes without shoes, and BMI was calculated as body weight (kg) divided by square of height (m\u00b2). Waist circumference (WC) was measured at the midpoint between the lowest rib and the iliac crest, and the average of three repeated measurements was used in the study. Blood pressure was measured using a mercury sphygmomanometer (W.A Baum Co. Inc., NY, USA) after at least 5 min rest in a seated position. Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured at phase I and V Korotkoff sound, respectively, and were recorded as the average of both arm readings. Blood samples were collected after at least 8 h of fasting, and plasma was separated for biochemical measurements. The concentrations of glucose, triglycerides, and HDL-C in the plasma were enzymatically measured using an autoanalyzer (Bayer HealthCare, Berlin, Germany).\n\n#### Definition of metabolic syndrome\n\nThe diagnostic criteria of MetS were defined based on the National Cholesterol Education Program Adult Treatment Panel III definition as modified by the American Heart Association and the National Heart Lung and Blood Institute. According to the criteria, MetS was defined as the presence of three or more of the following: (1) abdominal obesity, WC \u2265 90 cm in men or \u2265 80 cm in women; (2) hyperglycemia, fasting blood glucose \u2265 5.6 mmol/L, current use of insulin or oral hypoglycemic medication, or a physician\u2019s diagnosis; (3) hypertriglyceridemia, serum triglyceride concentration \u2265 1.7 mmol/L; (4) low HDL-C, serum HDL-C concentration < 1.0 mmol/L in men or < 1.3 mmol/L in women; and (5) elevated blood pressure, SBP \u2265 130 mmHg, DBP \u2265 85 mmHg, use of antihypertensive medication, or a physician\u2019s diagnosis of hypertension.\n\n#### Covariates\n\nDemographic characteristics, socioeconomic status, and lifestyle factors of the participants were recorded at baseline. Monthly household income was categorized into four groups: < 890 USD (approximately, 1 million Korean won in 2018), 890 to < 1780 USD, 1780 to < 2670 USD, and 2670 USD or higher. Education level was categorized into three groups: \u2264 6 years (elementary school graduates), 7 to \u2264 12 years (middle school or high school graduates), and > 12 years (college graduates or higher). Smoking status was classified into three groups: non-smoker, former smoker, and current smoker. Alcohol consumption was classified into three groups: non-drinker, former drinker, and current drinker. Physical activity was estimated using the metabolic equivalent of task hours per day.\n\n#### Statistical analysis\n\nContinuous variables are expressed as mean and standard deviation, and categorical variables are expressed as number and percentage. The differences in basic characteristics were examined using the Chi-square test for categorical variables and Student\u2019s t test for continuous variables. Comparisons of the variables at baseline across fruit or vegetable consumptions were made using either Chi-square tests or generalized linear models. Sex-stratified analysis was conducted because of the interaction between fruit and vegetable intake and sex. The hazard ratios (HRs) and 95% confidence intervals (CIs) for the risk of MetS and its components according to fruit or vegetable consumption were calculated using multivariate Cox proportional hazards models. Model 1 was unadjusted, and Model 2 was adjusted for age, residential location, household income, education level, smoking status, alcohol intake, physical activity, and BMI. Model 3 was adjusted for all variables in model 2 plus energy and food intakes. For the selection of variables for adjustment in the multivariable model, potential confounders from the previously published scientific literature were taken into account with the statistical approaches such as stepwise procedures or comparing the adjusted and unadjusted effect estimates of potential confounders. The proportional hazard assumption was proved graphically using log\u2013log plots and statistically using Schoenfeld\u2019s residuals. There was no violation of the proportional hazard assumption. A p value < 0.05 was considered statistically significant, and all data were analyzed using the SAS software, version 9.4 (SAS Institute, Cary, NC, USA).",
        "results": "### Results\n\nThe baseline characteristics of the participants according to development of MetS are shown in Table 1. Participants who developed MetS were older, more likely to be women, more likely to live in rural areas, had lower income, were less educated, were more likely to be current smokers, were more physically active, and had higher BMI compared with those who did not develop MetS. In addition, participants who developed MetS had a higher intake of energy from carbohydrate, vegetables, and refined grains and a lower intake of fruit and whole grains than those who did not develop MetS. The baseline characteristics of the participants according to fruit consumption, divided by sex, are shown in supplemental Table 1. Regardless of sex, frequent fruit consumers (\u2265 4 servings/day) were younger, more likely to live in urban areas, had higher income, were more educated, and were more likely non-smokers compared to rare fruit consumers (< 1 serving/day). Frequent fruit consumers also had higher intake of total energy, protein, and fat; more daily servings of vegetables, meat, whole grains, and dairy foods; and lower carbohydrate intake compared to rare fruit consumers. The characteristics of the participants according to vegetable consumption, divided by sex, are shown in supplemental Table 2. Frequent vegetable consumers (\u2265 4 servings/day) were younger, more likely to live in urban areas, had higher income, were more educated, were more likely current drinkers, and had higher BMI compared to rare vegetable consumers (< 1 serving/day). In addition, they had higher intake of total energy, protein, fat, fruits, meat, whole grains, and dairy foods and lower carbohydrate intake compared to rare vegetable consumers.\n\nThe relative risks for MetS and its components according to total fruit consumption, divided by sex, are shown in Table 2. For men, frequent fruit consumers (\u2265 4 servings/day) had a 31\u201345% reduction in the risk of developing MetS (HR = 0.55, 95% CI 0.44\u20130.67, p for trend < 0.0001) or individual components of MetS, such as abdominal obesity (HR = 0.61, 95% CI 0.47\u20130.78, _p < 0.0001),_ hypertriglyceridemia (HR = 0.60, 95% CI 0.48\u20130.76, p for trend < 0.0001), and elevated blood pressure (HR = 0.69, 95% CI 0.55\u20130.88, p for trend < 0.0001), compared with rare fruit consumers after adjusting for confounding factors of age, residence area, income, education level, smoking status, alcohol consumption, physical activity, BMI, total energy intake, and intake of vegetables, meat, refined grains, whole grains, and dairy foods. In women, the risk of developing MetS was reduced by 43% (HR = 0.57, 95% CI 0.47\u20130.70, p for trend < 0.0001) in frequent fruit consumers compared with rare consumers after adjustment for confounding factors. Frequent fruit consumers also had lower risk of abdominal obesity (HR = 0.74, 95% CI 0.58\u20130.94, _p for trend = 0.0125),_ hypertriglyceridemia (HR = 0.49, 95% CI 0.40\u20130.60, p < 0.0001), and elevated blood pressure (HR = 0.72, 95% CI 0.57\u20130.91, _p for trend = 0.0020)_ than rare consumers after adjustment for confounding factors.\n\nThe risks for MetS and its components according to total vegetable consumption by sex are shown in Table 3. There was no association between vegetable consumption and the risk of incident MetS in either men or women. Frequent vegetable consumption was associated with a lower risk of hyperglycemia after adjustment for confounding factors in men only (HR = 0.65, 95% CI 0.44\u20130.96, p for trend = 0.0275).\n\nThe associations between the consumption of fruits and vegetables by subtype and the risk of MetS were further examined. Frequent non-citrus fruits consumers (\u2265 4 servings/day) in men and women had a 31% (HR = 0.69, 95% CI 0.54\u20130.90, p for trend < 0.0001) and 20% (HR = 0.80, 95% CI 0.65\u20130.99, p for trend = 0.0005) reduced risk of developing MetS, respectively (Supplemental Table 3). In addition, frequent consumption of cruciferous vegetables, green leafy vegetables, and carotene-rich vegetables was not associated with risk of MetS, while other vegetables (\u2265 3 servings/day) had a 51% (HR = 1.51, 95% CI 1.22\u20131.86, p for trend = 0.0005) increased risk of developing MetS in men only (Supplemental Table 4).\n\nIn addition, to exclude residual confounding by smoking or alcohol consumption, the associations between the consumption of fruit or vegetable and MetS and its components in never smokers and never drinkers were examined. An inverse association between fruit consumption and MetS and its components were observed in never smokers regardless of sex. However, fruit consumption was not associated with MetS and its components in men who did never drink, although the inverse association was still found in women who did never drink. No association was observed between vegetable consumption and MetS and its components in never smokers and never drinkers regardless of sex (data not shown).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "4"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Association between fruit and vegetable consumption and risk of metabolic syndrome determined using the Korean Genome and Epidemiology Study (KoGES).",
        "abstract": "## PURPOSE\nThis study examined the association between fruit and vegetable consumption and the risk of metabolic syndrome (MetS) and its components in middle-aged and older adults using data from the Korean Genome and Epidemiology Study.\n## METHODS\nA total of 5688 participants aged 40-69\u00a0years without MetS at baseline were recruited and followed for 8\u00a0years. Fruit and vegetable consumption was assessed using a semi-quantitative food-frequency questionnaire at baseline and after 4\u00a0years. A multivariate Cox proportional hazards model was used to examine the risk of incident MetS and its components according to fruit and vegetable consumption.\n## RESULTS\nA total of 2067 participants (1020 men and 1047 women) developed MetS during the 8-year follow-up period. Frequent fruit consumers (\u2265\u20094 servings/day) had a lower risk of incident MetS than did rare consumers (<\u20091 serving/day) in both men and women (hazard ratio (HR)\u2009=\u20090.55, 95% CI 0.44-0.67, p\u2009<\u20090.0001 for men; HR\u2009=\u20090.57, 95% CI 0.47-0.70, p for trend\u2009<\u20090.0001 for women) after adjusting for confounders. Frequent fruit consumption was inversely associated with incidence of abdominal obesity, hypertriglyceridemia, and elevated blood pressure. Frequent vegetable consumption was inversely associated with a risk of hyperglycemia only in men (HR\u2009=\u20090.65, 95% CI 0.44-0.96, p for trend\u2009=\u20090.0275), but no association was observed with risk of incident MetS.\n## CONCLUSIONS\nHigh fruit consumption was associated with a reduced risk of incident MetS and its components, whereas high vegetable consumption was associated with a reduced risk of incident hyperglycemia, but not with MetS in middle-aged and older adults.",
        "methods": "### Methods\n\n#### Study population\n\nThe KoGES is a prospective study that explores dietary, environmental, and lifestyle factors and the incidence of chronic diseases in the general Korean population. The study consisted of 10,030 Korean adults aged 40\u201369 years who lived in one of two areas near central city in South Korea. Participants were recruited between 2001 and 2002, which is baseline, and follow-up examinations were conducted every 2 years over an 8-year period (2009\u20132010). All participants completed questionnaires regarding demographic information, lifestyle, medical history, and health conditions at baseline, and anthropometric measurements and biochemical tests were performed biennially. The study protocol was approved by the Institutional Review Boards of the Korea Centers for Disease Control and Prevention and Kyung Hee University (KHSIRB-16-022). Written informed consent was obtained from all participants. For the analysis of incident MetS, of the 10,030 participants, 2977 had MetS at baseline, 747 refused to participate in the follow-up examinations, 95 did not complete a food-frequency questionnaire (FFQ) at both baseline and the follow-up, 133 had cardiovascular disease (CVD) or cancer at baseline, 163 had extraordinary energy intake (< 500 kcal/day or > 6000 kcal/day) at baseline, and 227 had missing information on covariates or MetS parameters at baseline. All of the above participants were excluded from this study. Ultimately, a total of 5688 participants (2973 men and 2715 women) were included in the analysis of MetS. Likewise, for abdominal obesity analysis, 3962 who had abdominal obesity at baseline, 679 refused to participate in the follow-up examinations, 70 did not complete FFQ, 114 had CVD or cancer, 128 had extraordinary energy intake, and 206 had missing information were excluded. For hyperglycemia analysis, 1950 who had hyperglycemia at baseline, 820 refused to participate in the follow-up examinations, 83 did not complete FFQ, 165 had CVD or cancer, 202 had extraordinary energy intake, and 242 had missing information were excluded. For the hypertriglyceridemia, 3706 who had hypertriglyceridemia at baseline, 680 refused to participate in the follow-up examinations, 80 did not complete FFQ, 137 had CVD or cancer, 155 had extraordinary energy intake, and 266 had missing information were excluded. For low HDL-C analysis, 3781 who had low HDL-C at baseline, 649 refused to participate in the follow-up examinations, 123 did not complete FFQ, 143 had CVD or cancer, 139 had extraordinary energy intake, and 251 had missing information were excluded. For elevated blood pressure analysis, 4479 who had elevated blood pressure at baseline, 579 refused to participate in the follow-up examinations, 81 did not complete FFQ, 84 had CVD or cancer, 123 had extraordinary energy intake, and 205 had missing information were excluded. Finally, 4871 participants for abdominal obesity, 6568 participants for hyperglycemia, 5006 participants for hypertriglyceridemia, 4944 participants for low HDL-C, and 4479 participants for elevated blood pressure were included for the analysis of MetS components. A follow-up rate of 66.5% was achieved, accounting for 32,801 person-years. The average follow-up period was 70 months (range 17\u2013104 months).\n\n#### Dietary assessment\n\nAt baseline and second follow-up examinations (2005\u20132006), fruit and vegetable intakes were assessed with a semi-quantitative FFQ, which has been validated. For participants who acquired the MetS or were censored between baseline and the second follow-up examination, intake of fruit and vegetable was evaluated based on the FFQ at baseline; for those who acquired the MetS or were censored after the second follow-up examination, intake of fruit and vegetable was evaluated based on the average of FFQ at baseline and FFQ at the follow-up. The 12 types of fruits (tangerine, orange, persimmon or dried persimmon, watermelon, strawberry, grape, pear, oriental melon or melon, peach or prune, apple, banana, and tomato) and 20 types of vegetables (Korean cabbage, green pepper, red pepper leaf, spinach, lettuce, perilla leaf, chives or water parsley, leaf beet or Chinese mallow, radish or salted radish, deodeok or balloon flower, onion, cucumber, bean sprouts, carrot, pumpkin, zucchini, cabbage, fern or sweet potato stems, oyster mushroom, and black mushroom or button mushroom) were included in the FFQ. To examine the effect of fruits by subtype, fruits were classified into subgroups of citrus fruits (tangerines and oranges), non-citrus fruits (Persimmon/dried persimmon, watermelon, strawberry, grape, pear, oriental melon/melon, peach/prune, apple, and banana), and carotene-rich fruits (Tangerines, oranges, watermelon, persimmon, and tomato). In addition, to examine the effect of vegetables by subtype, vegetables were classified into subgroups of cruciferous vegetables (Korean cabbage, cabbage, radish, lettuce, and broccoli), green leafy vegetables (spinach, perilla leaf, chive/water parsley, and red pepper leaf), carotene-rich vegetables (carrots and pumpkin), and other vegetables (bean sprouts, fern/sweet potato stems, cucumber, onion, green chili, zucchini, deodeok/balloonflower, and mushrooms). In the FFQ, participants reported the usual frequency and portion size of fruit and vegetable consumption during the past year. The answer for frequency had nine options for each food (never or rarely, once/month, 2\u20133 times/month, 1\u20132 times/week, 3\u20134 times/week, 5\u20136 times/week, once/day, 2 times/day, and 3 times or more/day). The answer for portion size had three options for each food [\u00bd serving, 1 serving (standard), \u2265 2 servings]. One serving of fruit or vegetable was equal to 100 g or 70 g, respectively. For analysis, fruit and vegetable consumptions were converted to weekly frequencies and then multiplied by the reported portion size of each food. Finally, fruit and vegetable consumptions were classified into four groups (never or rarely, 1 to < 2 servings/day, 2 to < 4 servings/day, and \u2265 4 servings/day). Individual food intake was expressed as food consumption per week, and the total intake of each food was summed. In addition, seven meats (including pork, beef, chicken, and processed meat), refined grains (including white rice, noodles, and breads), whole grains (including barley, multigrain, and mixed grains), and dairy foods (including milk and yogurt) were evaluated. Nutrient intake was calculated using a database developed by the Rural Development Administration.\n\n#### Measurements\n\nHeight and body weight were measured in light clothes without shoes, and BMI was calculated as body weight (kg) divided by square of height (m\u00b2). Waist circumference (WC) was measured at the midpoint between the lowest rib and the iliac crest, and the average of three repeated measurements was used in the study. Blood pressure was measured using a mercury sphygmomanometer (W.A Baum Co. Inc., NY, USA) after at least 5 min rest in a seated position. Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured at phase I and V Korotkoff sound, respectively, and were recorded as the average of both arm readings. Blood samples were collected after at least 8 h of fasting, and plasma was separated for biochemical measurements. The concentrations of glucose, triglycerides, and HDL-C in the plasma were enzymatically measured using an autoanalyzer (Bayer HealthCare, Berlin, Germany).\n\n#### Definition of metabolic syndrome\n\nThe diagnostic criteria of MetS were defined based on the National Cholesterol Education Program Adult Treatment Panel III definition as modified by the American Heart Association and the National Heart Lung and Blood Institute. According to the criteria, MetS was defined as the presence of three or more of the following: (1) abdominal obesity, WC \u2265 90 cm in men or \u2265 80 cm in women; (2) hyperglycemia, fasting blood glucose \u2265 5.6 mmol/L, current use of insulin or oral hypoglycemic medication, or a physician\u2019s diagnosis; (3) hypertriglyceridemia, serum triglyceride concentration \u2265 1.7 mmol/L; (4) low HDL-C, serum HDL-C concentration < 1.0 mmol/L in men or < 1.3 mmol/L in women; and (5) elevated blood pressure, SBP \u2265 130 mmHg, DBP \u2265 85 mmHg, use of antihypertensive medication, or a physician\u2019s diagnosis of hypertension.\n\n#### Covariates\n\nDemographic characteristics, socioeconomic status, and lifestyle factors of the participants were recorded at baseline. Monthly household income was categorized into four groups: < 890 USD (approximately, 1 million Korean won in 2018), 890 to < 1780 USD, 1780 to < 2670 USD, and 2670 USD or higher. Education level was categorized into three groups: \u2264 6 years (elementary school graduates), 7 to \u2264 12 years (middle school or high school graduates), and > 12 years (college graduates or higher). Smoking status was classified into three groups: non-smoker, former smoker, and current smoker. Alcohol consumption was classified into three groups: non-drinker, former drinker, and current drinker. Physical activity was estimated using the metabolic equivalent of task hours per day.\n\n#### Statistical analysis\n\nContinuous variables are expressed as mean and standard deviation, and categorical variables are expressed as number and percentage. The differences in basic characteristics were examined using the Chi-square test for categorical variables and Student\u2019s t test for continuous variables. Comparisons of the variables at baseline across fruit or vegetable consumptions were made using either Chi-square tests or generalized linear models. Sex-stratified analysis was conducted because of the interaction between fruit and vegetable intake and sex. The hazard ratios (HRs) and 95% confidence intervals (CIs) for the risk of MetS and its components according to fruit or vegetable consumption were calculated using multivariate Cox proportional hazards models. Model 1 was unadjusted, and Model 2 was adjusted for age, residential location, household income, education level, smoking status, alcohol intake, physical activity, and BMI. Model 3 was adjusted for all variables in model 2 plus energy and food intakes. For the selection of variables for adjustment in the multivariable model, potential confounders from the previously published scientific literature were taken into account with the statistical approaches such as stepwise procedures or comparing the adjusted and unadjusted effect estimates of potential confounders. The proportional hazard assumption was proved graphically using log\u2013log plots and statistically using Schoenfeld\u2019s residuals. There was no violation of the proportional hazard assumption. A p value < 0.05 was considered statistically significant, and all data were analyzed using the SAS software, version 9.4 (SAS Institute, Cary, NC, USA).",
        "results": "### Results\n\nThe baseline characteristics of the participants according to development of MetS are shown in Table 1. Participants who developed MetS were older, more likely to be women, more likely to live in rural areas, had lower income, were less educated, were more likely to be current smokers, were more physically active, and had higher BMI compared with those who did not develop MetS. In addition, participants who developed MetS had a higher intake of energy from carbohydrate, vegetables, and refined grains and a lower intake of fruit and whole grains than those who did not develop MetS. The baseline characteristics of the participants according to fruit consumption, divided by sex, are shown in supplemental Table 1. Regardless of sex, frequent fruit consumers (\u2265 4 servings/day) were younger, more likely to live in urban areas, had higher income, were more educated, and were more likely non-smokers compared to rare fruit consumers (< 1 serving/day). Frequent fruit consumers also had higher intake of total energy, protein, and fat; more daily servings of vegetables, meat, whole grains, and dairy foods; and lower carbohydrate intake compared to rare fruit consumers. The characteristics of the participants according to vegetable consumption, divided by sex, are shown in supplemental Table 2. Frequent vegetable consumers (\u2265 4 servings/day) were younger, more likely to live in urban areas, had higher income, were more educated, were more likely current drinkers, and had higher BMI compared to rare vegetable consumers (< 1 serving/day). In addition, they had higher intake of total energy, protein, fat, fruits, meat, whole grains, and dairy foods and lower carbohydrate intake compared to rare vegetable consumers.\n\nThe relative risks for MetS and its components according to total fruit consumption, divided by sex, are shown in Table 2. For men, frequent fruit consumers (\u2265 4 servings/day) had a 31\u201345% reduction in the risk of developing MetS (HR = 0.55, 95% CI 0.44\u20130.67, p for trend < 0.0001) or individual components of MetS, such as abdominal obesity (HR = 0.61, 95% CI 0.47\u20130.78, _p < 0.0001),_ hypertriglyceridemia (HR = 0.60, 95% CI 0.48\u20130.76, p for trend < 0.0001), and elevated blood pressure (HR = 0.69, 95% CI 0.55\u20130.88, p for trend < 0.0001), compared with rare fruit consumers after adjusting for confounding factors of age, residence area, income, education level, smoking status, alcohol consumption, physical activity, BMI, total energy intake, and intake of vegetables, meat, refined grains, whole grains, and dairy foods. In women, the risk of developing MetS was reduced by 43% (HR = 0.57, 95% CI 0.47\u20130.70, p for trend < 0.0001) in frequent fruit consumers compared with rare consumers after adjustment for confounding factors. Frequent fruit consumers also had lower risk of abdominal obesity (HR = 0.74, 95% CI 0.58\u20130.94, _p for trend = 0.0125),_ hypertriglyceridemia (HR = 0.49, 95% CI 0.40\u20130.60, p < 0.0001), and elevated blood pressure (HR = 0.72, 95% CI 0.57\u20130.91, _p for trend = 0.0020)_ than rare consumers after adjustment for confounding factors.\n\nThe risks for MetS and its components according to total vegetable consumption by sex are shown in Table 3. There was no association between vegetable consumption and the risk of incident MetS in either men or women. Frequent vegetable consumption was associated with a lower risk of hyperglycemia after adjustment for confounding factors in men only (HR = 0.65, 95% CI 0.44\u20130.96, p for trend = 0.0275).\n\nThe associations between the consumption of fruits and vegetables by subtype and the risk of MetS were further examined. Frequent non-citrus fruits consumers (\u2265 4 servings/day) in men and women had a 31% (HR = 0.69, 95% CI 0.54\u20130.90, p for trend < 0.0001) and 20% (HR = 0.80, 95% CI 0.65\u20130.99, p for trend = 0.0005) reduced risk of developing MetS, respectively (Supplemental Table 3). In addition, frequent consumption of cruciferous vegetables, green leafy vegetables, and carotene-rich vegetables was not associated with risk of MetS, while other vegetables (\u2265 3 servings/day) had a 51% (HR = 1.51, 95% CI 1.22\u20131.86, p for trend = 0.0005) increased risk of developing MetS in men only (Supplemental Table 4).\n\nIn addition, to exclude residual confounding by smoking or alcohol consumption, the associations between the consumption of fruit or vegetable and MetS and its components in never smokers and never drinkers were examined. An inverse association between fruit consumption and MetS and its components were observed in never smokers regardless of sex. However, fruit consumption was not associated with MetS and its components in men who did never drink, although the inverse association was still found in women who did never drink. No association was observed between vegetable consumption and MetS and its components in never smokers and never drinkers regardless of sex (data not shown).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Association between fruit and vegetable consumption and risk of metabolic syndrome determined using the Korean Genome and Epidemiology Study (KoGES).",
        "abstract": "## PURPOSE\nThis study examined the association between fruit and vegetable consumption and the risk of metabolic syndrome (MetS) and its components in middle-aged and older adults using data from the Korean Genome and Epidemiology Study.\n## METHODS\nA total of 5688 participants aged 40-69\u00a0years without MetS at baseline were recruited and followed for 8\u00a0years. Fruit and vegetable consumption was assessed using a semi-quantitative food-frequency questionnaire at baseline and after 4\u00a0years. A multivariate Cox proportional hazards model was used to examine the risk of incident MetS and its components according to fruit and vegetable consumption.\n## RESULTS\nA total of 2067 participants (1020 men and 1047 women) developed MetS during the 8-year follow-up period. Frequent fruit consumers (\u2265\u20094 servings/day) had a lower risk of incident MetS than did rare consumers (<\u20091 serving/day) in both men and women (hazard ratio (HR)\u2009=\u20090.55, 95% CI 0.44-0.67, p\u2009<\u20090.0001 for men; HR\u2009=\u20090.57, 95% CI 0.47-0.70, p for trend\u2009<\u20090.0001 for women) after adjusting for confounders. Frequent fruit consumption was inversely associated with incidence of abdominal obesity, hypertriglyceridemia, and elevated blood pressure. Frequent vegetable consumption was inversely associated with a risk of hyperglycemia only in men (HR\u2009=\u20090.65, 95% CI 0.44-0.96, p for trend\u2009=\u20090.0275), but no association was observed with risk of incident MetS.\n## CONCLUSIONS\nHigh fruit consumption was associated with a reduced risk of incident MetS and its components, whereas high vegetable consumption was associated with a reduced risk of incident hyperglycemia, but not with MetS in middle-aged and older adults.",
        "methods": "### Methods\n\n#### Study population\n\nThe KoGES is a prospective study that explores dietary, environmental, and lifestyle factors and the incidence of chronic diseases in the general Korean population. The study consisted of 10,030 Korean adults aged 40\u201369 years who lived in one of two areas near central city in South Korea. Participants were recruited between 2001 and 2002, which is baseline, and follow-up examinations were conducted every 2 years over an 8-year period (2009\u20132010). All participants completed questionnaires regarding demographic information, lifestyle, medical history, and health conditions at baseline, and anthropometric measurements and biochemical tests were performed biennially. The study protocol was approved by the Institutional Review Boards of the Korea Centers for Disease Control and Prevention and Kyung Hee University (KHSIRB-16-022). Written informed consent was obtained from all participants. For the analysis of incident MetS, of the 10,030 participants, 2977 had MetS at baseline, 747 refused to participate in the follow-up examinations, 95 did not complete a food-frequency questionnaire (FFQ) at both baseline and the follow-up, 133 had cardiovascular disease (CVD) or cancer at baseline, 163 had extraordinary energy intake (< 500 kcal/day or > 6000 kcal/day) at baseline, and 227 had missing information on covariates or MetS parameters at baseline. All of the above participants were excluded from this study. Ultimately, a total of 5688 participants (2973 men and 2715 women) were included in the analysis of MetS. Likewise, for abdominal obesity analysis, 3962 who had abdominal obesity at baseline, 679 refused to participate in the follow-up examinations, 70 did not complete FFQ, 114 had CVD or cancer, 128 had extraordinary energy intake, and 206 had missing information were excluded. For hyperglycemia analysis, 1950 who had hyperglycemia at baseline, 820 refused to participate in the follow-up examinations, 83 did not complete FFQ, 165 had CVD or cancer, 202 had extraordinary energy intake, and 242 had missing information were excluded. For the hypertriglyceridemia, 3706 who had hypertriglyceridemia at baseline, 680 refused to participate in the follow-up examinations, 80 did not complete FFQ, 137 had CVD or cancer, 155 had extraordinary energy intake, and 266 had missing information were excluded. For low HDL-C analysis, 3781 who had low HDL-C at baseline, 649 refused to participate in the follow-up examinations, 123 did not complete FFQ, 143 had CVD or cancer, 139 had extraordinary energy intake, and 251 had missing information were excluded. For elevated blood pressure analysis, 4479 who had elevated blood pressure at baseline, 579 refused to participate in the follow-up examinations, 81 did not complete FFQ, 84 had CVD or cancer, 123 had extraordinary energy intake, and 205 had missing information were excluded. Finally, 4871 participants for abdominal obesity, 6568 participants for hyperglycemia, 5006 participants for hypertriglyceridemia, 4944 participants for low HDL-C, and 4479 participants for elevated blood pressure were included for the analysis of MetS components. A follow-up rate of 66.5% was achieved, accounting for 32,801 person-years. The average follow-up period was 70 months (range 17\u2013104 months).\n\n#### Dietary assessment\n\nAt baseline and second follow-up examinations (2005\u20132006), fruit and vegetable intakes were assessed with a semi-quantitative FFQ, which has been validated. For participants who acquired the MetS or were censored between baseline and the second follow-up examination, intake of fruit and vegetable was evaluated based on the FFQ at baseline; for those who acquired the MetS or were censored after the second follow-up examination, intake of fruit and vegetable was evaluated based on the average of FFQ at baseline and FFQ at the follow-up. The 12 types of fruits (tangerine, orange, persimmon or dried persimmon, watermelon, strawberry, grape, pear, oriental melon or melon, peach or prune, apple, banana, and tomato) and 20 types of vegetables (Korean cabbage, green pepper, red pepper leaf, spinach, lettuce, perilla leaf, chives or water parsley, leaf beet or Chinese mallow, radish or salted radish, deodeok or balloon flower, onion, cucumber, bean sprouts, carrot, pumpkin, zucchini, cabbage, fern or sweet potato stems, oyster mushroom, and black mushroom or button mushroom) were included in the FFQ. To examine the effect of fruits by subtype, fruits were classified into subgroups of citrus fruits (tangerines and oranges), non-citrus fruits (Persimmon/dried persimmon, watermelon, strawberry, grape, pear, oriental melon/melon, peach/prune, apple, and banana), and carotene-rich fruits (Tangerines, oranges, watermelon, persimmon, and tomato). In addition, to examine the effect of vegetables by subtype, vegetables were classified into subgroups of cruciferous vegetables (Korean cabbage, cabbage, radish, lettuce, and broccoli), green leafy vegetables (spinach, perilla leaf, chive/water parsley, and red pepper leaf), carotene-rich vegetables (carrots and pumpkin), and other vegetables (bean sprouts, fern/sweet potato stems, cucumber, onion, green chili, zucchini, deodeok/balloonflower, and mushrooms). In the FFQ, participants reported the usual frequency and portion size of fruit and vegetable consumption during the past year. The answer for frequency had nine options for each food (never or rarely, once/month, 2\u20133 times/month, 1\u20132 times/week, 3\u20134 times/week, 5\u20136 times/week, once/day, 2 times/day, and 3 times or more/day). The answer for portion size had three options for each food [\u00bd serving, 1 serving (standard), \u2265 2 servings]. One serving of fruit or vegetable was equal to 100 g or 70 g, respectively. For analysis, fruit and vegetable consumptions were converted to weekly frequencies and then multiplied by the reported portion size of each food. Finally, fruit and vegetable consumptions were classified into four groups (never or rarely, 1 to < 2 servings/day, 2 to < 4 servings/day, and \u2265 4 servings/day). Individual food intake was expressed as food consumption per week, and the total intake of each food was summed. In addition, seven meats (including pork, beef, chicken, and processed meat), refined grains (including white rice, noodles, and breads), whole grains (including barley, multigrain, and mixed grains), and dairy foods (including milk and yogurt) were evaluated. Nutrient intake was calculated using a database developed by the Rural Development Administration.\n\n#### Measurements\n\nHeight and body weight were measured in light clothes without shoes, and BMI was calculated as body weight (kg) divided by square of height (m\u00b2). Waist circumference (WC) was measured at the midpoint between the lowest rib and the iliac crest, and the average of three repeated measurements was used in the study. Blood pressure was measured using a mercury sphygmomanometer (W.A Baum Co. Inc., NY, USA) after at least 5 min rest in a seated position. Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured at phase I and V Korotkoff sound, respectively, and were recorded as the average of both arm readings. Blood samples were collected after at least 8 h of fasting, and plasma was separated for biochemical measurements. The concentrations of glucose, triglycerides, and HDL-C in the plasma were enzymatically measured using an autoanalyzer (Bayer HealthCare, Berlin, Germany).\n\n#### Definition of metabolic syndrome\n\nThe diagnostic criteria of MetS were defined based on the National Cholesterol Education Program Adult Treatment Panel III definition as modified by the American Heart Association and the National Heart Lung and Blood Institute. According to the criteria, MetS was defined as the presence of three or more of the following: (1) abdominal obesity, WC \u2265 90 cm in men or \u2265 80 cm in women; (2) hyperglycemia, fasting blood glucose \u2265 5.6 mmol/L, current use of insulin or oral hypoglycemic medication, or a physician\u2019s diagnosis; (3) hypertriglyceridemia, serum triglyceride concentration \u2265 1.7 mmol/L; (4) low HDL-C, serum HDL-C concentration < 1.0 mmol/L in men or < 1.3 mmol/L in women; and (5) elevated blood pressure, SBP \u2265 130 mmHg, DBP \u2265 85 mmHg, use of antihypertensive medication, or a physician\u2019s diagnosis of hypertension.\n\n#### Covariates\n\nDemographic characteristics, socioeconomic status, and lifestyle factors of the participants were recorded at baseline. Monthly household income was categorized into four groups: < 890 USD (approximately, 1 million Korean won in 2018), 890 to < 1780 USD, 1780 to < 2670 USD, and 2670 USD or higher. Education level was categorized into three groups: \u2264 6 years (elementary school graduates), 7 to \u2264 12 years (middle school or high school graduates), and > 12 years (college graduates or higher). Smoking status was classified into three groups: non-smoker, former smoker, and current smoker. Alcohol consumption was classified into three groups: non-drinker, former drinker, and current drinker. Physical activity was estimated using the metabolic equivalent of task hours per day.\n\n#### Statistical analysis\n\nContinuous variables are expressed as mean and standard deviation, and categorical variables are expressed as number and percentage. The differences in basic characteristics were examined using the Chi-square test for categorical variables and Student\u2019s t test for continuous variables. Comparisons of the variables at baseline across fruit or vegetable consumptions were made using either Chi-square tests or generalized linear models. Sex-stratified analysis was conducted because of the interaction between fruit and vegetable intake and sex. The hazard ratios (HRs) and 95% confidence intervals (CIs) for the risk of MetS and its components according to fruit or vegetable consumption were calculated using multivariate Cox proportional hazards models. Model 1 was unadjusted, and Model 2 was adjusted for age, residential location, household income, education level, smoking status, alcohol intake, physical activity, and BMI. Model 3 was adjusted for all variables in model 2 plus energy and food intakes. For the selection of variables for adjustment in the multivariable model, potential confounders from the previously published scientific literature were taken into account with the statistical approaches such as stepwise procedures or comparing the adjusted and unadjusted effect estimates of potential confounders. The proportional hazard assumption was proved graphically using log\u2013log plots and statistically using Schoenfeld\u2019s residuals. There was no violation of the proportional hazard assumption. A p value < 0.05 was considered statistically significant, and all data were analyzed using the SAS software, version 9.4 (SAS Institute, Cary, NC, USA).",
        "results": "### Results\n\nThe baseline characteristics of the participants according to development of MetS are shown in Table 1. Participants who developed MetS were older, more likely to be women, more likely to live in rural areas, had lower income, were less educated, were more likely to be current smokers, were more physically active, and had higher BMI compared with those who did not develop MetS. In addition, participants who developed MetS had a higher intake of energy from carbohydrate, vegetables, and refined grains and a lower intake of fruit and whole grains than those who did not develop MetS. The baseline characteristics of the participants according to fruit consumption, divided by sex, are shown in supplemental Table 1. Regardless of sex, frequent fruit consumers (\u2265 4 servings/day) were younger, more likely to live in urban areas, had higher income, were more educated, and were more likely non-smokers compared to rare fruit consumers (< 1 serving/day). Frequent fruit consumers also had higher intake of total energy, protein, and fat; more daily servings of vegetables, meat, whole grains, and dairy foods; and lower carbohydrate intake compared to rare fruit consumers. The characteristics of the participants according to vegetable consumption, divided by sex, are shown in supplemental Table 2. Frequent vegetable consumers (\u2265 4 servings/day) were younger, more likely to live in urban areas, had higher income, were more educated, were more likely current drinkers, and had higher BMI compared to rare vegetable consumers (< 1 serving/day). In addition, they had higher intake of total energy, protein, fat, fruits, meat, whole grains, and dairy foods and lower carbohydrate intake compared to rare vegetable consumers.\n\nThe relative risks for MetS and its components according to total fruit consumption, divided by sex, are shown in Table 2. For men, frequent fruit consumers (\u2265 4 servings/day) had a 31\u201345% reduction in the risk of developing MetS (HR = 0.55, 95% CI 0.44\u20130.67, p for trend < 0.0001) or individual components of MetS, such as abdominal obesity (HR = 0.61, 95% CI 0.47\u20130.78, _p < 0.0001),_ hypertriglyceridemia (HR = 0.60, 95% CI 0.48\u20130.76, p for trend < 0.0001), and elevated blood pressure (HR = 0.69, 95% CI 0.55\u20130.88, p for trend < 0.0001), compared with rare fruit consumers after adjusting for confounding factors of age, residence area, income, education level, smoking status, alcohol consumption, physical activity, BMI, total energy intake, and intake of vegetables, meat, refined grains, whole grains, and dairy foods. In women, the risk of developing MetS was reduced by 43% (HR = 0.57, 95% CI 0.47\u20130.70, p for trend < 0.0001) in frequent fruit consumers compared with rare consumers after adjustment for confounding factors. Frequent fruit consumers also had lower risk of abdominal obesity (HR = 0.74, 95% CI 0.58\u20130.94, _p for trend = 0.0125),_ hypertriglyceridemia (HR = 0.49, 95% CI 0.40\u20130.60, p < 0.0001), and elevated blood pressure (HR = 0.72, 95% CI 0.57\u20130.91, _p for trend = 0.0020)_ than rare consumers after adjustment for confounding factors.\n\nThe risks for MetS and its components according to total vegetable consumption by sex are shown in Table 3. There was no association between vegetable consumption and the risk of incident MetS in either men or women. Frequent vegetable consumption was associated with a lower risk of hyperglycemia after adjustment for confounding factors in men only (HR = 0.65, 95% CI 0.44\u20130.96, p for trend = 0.0275).\n\nThe associations between the consumption of fruits and vegetables by subtype and the risk of MetS were further examined. Frequent non-citrus fruits consumers (\u2265 4 servings/day) in men and women had a 31% (HR = 0.69, 95% CI 0.54\u20130.90, p for trend < 0.0001) and 20% (HR = 0.80, 95% CI 0.65\u20130.99, p for trend = 0.0005) reduced risk of developing MetS, respectively (Supplemental Table 3). In addition, frequent consumption of cruciferous vegetables, green leafy vegetables, and carotene-rich vegetables was not associated with risk of MetS, while other vegetables (\u2265 3 servings/day) had a 51% (HR = 1.51, 95% CI 1.22\u20131.86, p for trend = 0.0005) increased risk of developing MetS in men only (Supplemental Table 4).\n\nIn addition, to exclude residual confounding by smoking or alcohol consumption, the associations between the consumption of fruit or vegetable and MetS and its components in never smokers and never drinkers were examined. An inverse association between fruit consumption and MetS and its components were observed in never smokers regardless of sex. However, fruit consumption was not associated with MetS and its components in men who did never drink, although the inverse association was still found in women who did never drink. No association was observed between vegetable consumption and MetS and its components in never smokers and never drinkers regardless of sex (data not shown).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Dietary intake and the development of the metabolic syndrome: the Atherosclerosis Risk in Communities study.",
        "abstract": "## BACKGROUND\nThe role of diet in the origin of metabolic syndrome (MetSyn) is not well understood; thus, we sought to evaluate the relationship between incident MetSyn and dietary intake using prospective data from 9514 participants (age, 45 to 64 years) enrolled in the Atherosclerosis Risk in Communities (ARIC) study.\n## METHODS AND RESULTS\nDietary intake was assessed at baseline via a 66-item food frequency questionnaire. We used principal-components analysis to derive \"Western\" and \"prudent\" dietary patterns from 32 food groups and evaluated 10 food groups used in previous studies of the ARIC cohort. MetSyn was defined by American Heart Association guidelines. Proportional-hazards regression was used. Over 9 years of follow-up, 3782 incident cases of MetSyn were identified. After adjustment for demographic factors, smoking, physical activity, and energy intake, consumption of a Western dietary pattern (P(trend)=0.03) was adversely associated with incident MetSyn. After further adjustment for intake of meat, dairy, fruits and vegetables, refined grains, and whole grains, analysis of individual food groups revealed that meat (P(trend)<0.001), fried foods (P(trend)=0.02), and diet soda (P(trend)=< 0.001) also were adversely associated with incident MetSyn, whereas dairy consumption (P(trend)=0.006) was beneficial. No associations were observed between incident MetSyn and a prudent dietary pattern or intakes of whole grains, refined grains, fruits and vegetables, nuts, coffee, or sweetened beverages.\n## CONCLUSIONS\nThese prospective findings suggest that consumption of a Western dietary pattern, meat, and fried foods promotes the incidence of MetSyn, whereas dairy consumption provides some protection. The diet soda association was not hypothesized and deserves further study.",
        "methods": "#### Methods\n###### Study Population\nThe ARIC study is a multicenter population-based prospective cohort study designed to investigate the origin and natural history of atherosclerosis in middle-aged adults. Participants were recruited from 4 US communities: Forsyth County, North Carolina; Jackson, Miss; suburbs of Minneapolis, Minn; and Washington County, Maryland. The study cohort included 15 972 white and black men and women 45 to 64 years of age at baseline in 1987 to 1989 (visit 1). Cohort reexaminations took place at 3-year intervals; response rates were 93%, 86%, and 81% at visits 2 (1990 to 1992), 3 (1993 to 1995), and 4 (1996 to 1998), respectively. Local institutional review boards approved the ARIC protocol, and all subjects gave informed consent. All participants with MetSyn at baseline were excluded (n=5879), as were those with cardiovascular disease (n=776), missing diet data (n=11), or implausible energy intakes defined as \u2264500 or \u22643500 kcal/d in women and \u2264700 or \u22644500 kcal/d in men (n=361). The exclusion criteria were not mutually exclusive; thus, data from 9514 participants were included in the analyses.\n\n###### Data Collection\nARIC participants underwent interviews, fasting venipuncture, and measurement of blood pressure and anthropometrics at each examination. Trained interviewers ascertained basic demographic data, medical history, and information about personal habits, including diet (2 exams only), smoking, physical activity (with the Baecke et al questionnaire), and medication use.\n\n###### Dietary Assessment\nAt baseline and examination 3 (6 years later), usual dietary intake was assessed by a 66-item interviewer-administered semiquantitative food-frequency questionnaire (FFQ). The questionnaire was a modified version of the 61-item instrument developed by Willett et al. Participants reported the frequency of consumption of each food on the basis of 9 levels of frequency, ranging from never or \u22641 time a month to \u22656 times a day. Standard serving sizes and food models were provided as a reference for intake estimation. Interviewers also obtained additional information, including the brand name of the breakfast cereal usually consumed. Food and beverages from the FFQ were categorized into 29 food subgroups, which were used to derive dietary patterns via principal-components analysis. The subgroups were further collapsed into 5 major food groups: meat, dairy, fruits and vegetables, refined grains, and whole grains (see the online Data Supplement). Additionally, associations with the following individual food items were assessed: fried foods, sweetened beverages (regular soda and fruit drinks), diet soda, nuts, and coffee.\n\n###### Metabolic Syndrome\nMetSyn, defined according to current American Heart Association guidelines, was characterized by the presence of any 3 of the following risk factors: waist circumference \u2265102 cm in men and \u226588 cm in women; systolic blood pressure \u2265130 mm Hg, diastolic blood pressure \u226585 mm Hg, or current use of antihypertensive medication; and fasting glucose \u2265100 mg/dL or current drug treatment for elevated glucose. Waist circumference was measured with a cloth tape at the umbilicus. Fasting blood samples were drawn from an antecubital vein into tubes containing EDTA (lipids) or a serum separator gel (glucose). Total plasma cholesterol and triglycerides were determined by enzymatic methods. HDL cholesterol was measured after dextran-magnesium precipitation. Serum glucose levels were assessed with a hexokinase/glucose-6-phosphate dehydrogenase method. Sitting blood pressure was measured by trained technicians 3 times on each participant with a random-zero sphygmomanometer after a 5-minute rest. Because of substantial intraindividual variation in blood pressure measurements, the first blood pressure reading was dropped and the second and third readings were averaged for use in this study. Study participants were asked to bring all medications, vitamins, and supplements taken in the 2 weeks before the examination. Information on pharmacological treatment of hypertension and elevated glucose was based on the participant\u2019s self-reported use of any medication to treat high blood pressure or elevated glucose and the transcription and coding of all medication names.\n\n###### Statistical Analysis\nAll analyses were performed with SAS (version 9.1, SAS Institute Inc, Cary, NC). Principal-components analysis (PROC FACTOR) was used to derive dietary patterns and to determine factor loadings for each of the 29 food subgroups. Factors were rotated with varimax rotation to maintain uncorrelated factors and enhance interpretability. After evaluation of the eigenvalues and interpretability, all values \u22652.0 were retained, resulting in a 2-factor solution, with factor 1 explaining 10.2% of the variability and factor 2 explaining 9.7%. Dietary patterns were named according to the nature of the food groups loading highest for each of the factors. Interval-censored Cox proportional-hazards regression was used to assess the relation between diet and incident MetSyn. The dietary patterns and major food groups (meat, dairy, fruit and vegetables, whole grains, and refined grains) were categorized into quintiles; fried foods, nuts, sweetened beverages, diet soda, and coffee were classified into tertiles, given their low level of consumption. Hazard ratios (HRs) were obtained by entering the quintiles (or tertiles) into the model as indicator variables, whereas tests for trend across the quintiles were conducted by entering the median level of intake at each quintile (or tertile) into the model as a continuous variable. Cumulative average diet, which has been shown to increase the precision of dietary data, was used to create food groups and patterns. Specifically, the time-varying dietary covariates were modeled as follows: Between baseline and examination 3, dietary exposures were based on diet as measured at the baseline examination; after examination 3, dietary exposures were based on the mean of baseline diet and examination 3 diet. Spearman partial correlation coefficients, adjusted for age, sex, race, education, center, and total energy intake, were computed between examinations 1 and 3 for meat (r=0.42), dairy (r=0.42), fruit and vegetables (r=0.50), whole grains (r=0.45), and refined grains (r=0.36) (P<0.001 for all). Considering well-known within-person variation in response to diet questionnaires of the FFQ type, the correlations suggest considerable tracking (lack of change over time), although they also could represent some change in diet over 6 years. Model 1 adjusted for baseline age, sex, race, center, education (less than high school, high school or vocational, college and higher), and energy intake (kcal/d); model 2 further adjusted for behavioral characteristics, including smoking (current, former, never), pack-years (continuous), and physical activity (continuous). For analyses of individual food groups, a third model was considered that also adjusted for the major food groups: meat, dairy, fruit and vegetables, whole grains, and refined grains. The authors had full access to and take responsibility for the integrity of the data. All authors have read and agree to the manuscript as written.",
        "results": "#### Results\nThe cohort of 9514 ARIC participants had a mean age at baseline of 53.6 years, and 44.1% were male. Over 9 years of follow-up, 3782 participants (39.8%) developed incident MetSyn. On diagnosis, the most frequent contributor to MetSyn was abdominal obesity (79.8%), followed by elevated fasting glucose (77.5%), hypertension (62.4%), low HDL (58.5%), and high levels of triglycerides (48.5%). Sex-stratified baseline demographic, behavioral, and dietary variables are presented in Table 1. Factor analysis revealed 2 main dietary patterns for the population under study; food groups and factor loadings for those with absolute values of \u22650.20 are presented in Table 2. Factor 1, identified as a Western dietary pattern, was characterized by high intakes of refined grains, processed meat, and red meat, while the prudent dietary pattern was typified by greater consumption of cruciferous and carotinoid vegetables, fruit, fish, and poultry. Higher scores for the Western dietary pattern were associated with greater risk of developing MetSyn, even after adjustment for behavioral characteristics (Ptrend=0.03) (Table 3). Participants in the highest quintile of Western dietary pattern scores had an 18% greater risk (HR, 1.18; 95% CI, 1.03 to 1.37) of developing MetSyn than those in the lowest quintile. Consumption of a prudent dietary pattern was not associated with incident MetSyn in either model. In the analyses of major food groups, results from models 1 and 2 were similar; thus, only model 1 is shown in the tables. After adjustment for demographics, behaviors, and dietary pattern scores, meat consumption was positively associated with higher incidence of MetSyn (Ptrend=0.001), whereas dairy consumption conferred some protection (Ptrend=0.001) (Table 3). There was no association between incident MetSyn and consumption of whole grains, refined grains, or fruits and vegetables. The HR for the consumption of extreme quintiles of meat was 1.26 (95% CI, 1.11 to 1.43). Individual foods in the meat group also were adversely associated with MetSyn: hamburger (Ptrend=0.002), hot dogs (Ptrend=0.002), and processed meats (Ptrend=0.03). The HR for the extreme quintiles of dairy consumption was 0.87 (95% CI, 0.77 to 0.98); however, individual dairy foods were not associated with incident MetSyn. As shown in Table 4, consumption of fried foods was associated with increased risk of MetSyn (Ptrend=0.02). The HR for the extreme tertiles was 1.25 (95% CI, 1.14 to 1.37). Sweetened beverages also were associated with incident MetSyn after base adjustments (HR, 1.17; 95% CI, 1.07 to 1.28; Ptrend=0.02), although the association became nonsignificant with additional adjustments. Diet soda intake was strongly associated with increased risk across all models (HR, 1.34; 95% CI, 1.11 to 1.29). No relations were observed.\n\nIn sensitivity analyses, similar results were observed when diet relations with MetSyn were examined even after exclusion of individuals with prevalent diabetes at baseline (n=1870) and after omitting individuals diagnosed with incident MetSyn at 1 examination who reverted at a later examination (n=1161). We also evaluated whether gender or baseline overweight status modified the relation between dietary intake and incident MetSyn, but no interactions were observed.",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Dietary intake and the development of the metabolic syndrome: the Atherosclerosis Risk in Communities study.",
        "abstract": "## BACKGROUND\nThe role of diet in the origin of metabolic syndrome (MetSyn) is not well understood; thus, we sought to evaluate the relationship between incident MetSyn and dietary intake using prospective data from 9514 participants (age, 45 to 64 years) enrolled in the Atherosclerosis Risk in Communities (ARIC) study.\n## METHODS AND RESULTS\nDietary intake was assessed at baseline via a 66-item food frequency questionnaire. We used principal-components analysis to derive \"Western\" and \"prudent\" dietary patterns from 32 food groups and evaluated 10 food groups used in previous studies of the ARIC cohort. MetSyn was defined by American Heart Association guidelines. Proportional-hazards regression was used. Over 9 years of follow-up, 3782 incident cases of MetSyn were identified. After adjustment for demographic factors, smoking, physical activity, and energy intake, consumption of a Western dietary pattern (P(trend)=0.03) was adversely associated with incident MetSyn. After further adjustment for intake of meat, dairy, fruits and vegetables, refined grains, and whole grains, analysis of individual food groups revealed that meat (P(trend)<0.001), fried foods (P(trend)=0.02), and diet soda (P(trend)=< 0.001) also were adversely associated with incident MetSyn, whereas dairy consumption (P(trend)=0.006) was beneficial. No associations were observed between incident MetSyn and a prudent dietary pattern or intakes of whole grains, refined grains, fruits and vegetables, nuts, coffee, or sweetened beverages.\n## CONCLUSIONS\nThese prospective findings suggest that consumption of a Western dietary pattern, meat, and fried foods promotes the incidence of MetSyn, whereas dairy consumption provides some protection. The diet soda association was not hypothesized and deserves further study.",
        "methods": "#### Methods\n###### Study Population\nThe ARIC study is a multicenter population-based prospective cohort study designed to investigate the origin and natural history of atherosclerosis in middle-aged adults. Participants were recruited from 4 US communities: Forsyth County, North Carolina; Jackson, Miss; suburbs of Minneapolis, Minn; and Washington County, Maryland. The study cohort included 15 972 white and black men and women 45 to 64 years of age at baseline in 1987 to 1989 (visit 1). Cohort reexaminations took place at 3-year intervals; response rates were 93%, 86%, and 81% at visits 2 (1990 to 1992), 3 (1993 to 1995), and 4 (1996 to 1998), respectively. Local institutional review boards approved the ARIC protocol, and all subjects gave informed consent. All participants with MetSyn at baseline were excluded (n=5879), as were those with cardiovascular disease (n=776), missing diet data (n=11), or implausible energy intakes defined as \u2264500 or \u22643500 kcal/d in women and \u2264700 or \u22644500 kcal/d in men (n=361). The exclusion criteria were not mutually exclusive; thus, data from 9514 participants were included in the analyses.\n\n###### Data Collection\nARIC participants underwent interviews, fasting venipuncture, and measurement of blood pressure and anthropometrics at each examination. Trained interviewers ascertained basic demographic data, medical history, and information about personal habits, including diet (2 exams only), smoking, physical activity (with the Baecke et al questionnaire), and medication use.\n\n###### Dietary Assessment\nAt baseline and examination 3 (6 years later), usual dietary intake was assessed by a 66-item interviewer-administered semiquantitative food-frequency questionnaire (FFQ). The questionnaire was a modified version of the 61-item instrument developed by Willett et al. Participants reported the frequency of consumption of each food on the basis of 9 levels of frequency, ranging from never or \u22641 time a month to \u22656 times a day. Standard serving sizes and food models were provided as a reference for intake estimation. Interviewers also obtained additional information, including the brand name of the breakfast cereal usually consumed. Food and beverages from the FFQ were categorized into 29 food subgroups, which were used to derive dietary patterns via principal-components analysis. The subgroups were further collapsed into 5 major food groups: meat, dairy, fruits and vegetables, refined grains, and whole grains (see the online Data Supplement). Additionally, associations with the following individual food items were assessed: fried foods, sweetened beverages (regular soda and fruit drinks), diet soda, nuts, and coffee.\n\n###### Metabolic Syndrome\nMetSyn, defined according to current American Heart Association guidelines, was characterized by the presence of any 3 of the following risk factors: waist circumference \u2265102 cm in men and \u226588 cm in women; systolic blood pressure \u2265130 mm Hg, diastolic blood pressure \u226585 mm Hg, or current use of antihypertensive medication; and fasting glucose \u2265100 mg/dL or current drug treatment for elevated glucose. Waist circumference was measured with a cloth tape at the umbilicus. Fasting blood samples were drawn from an antecubital vein into tubes containing EDTA (lipids) or a serum separator gel (glucose). Total plasma cholesterol and triglycerides were determined by enzymatic methods. HDL cholesterol was measured after dextran-magnesium precipitation. Serum glucose levels were assessed with a hexokinase/glucose-6-phosphate dehydrogenase method. Sitting blood pressure was measured by trained technicians 3 times on each participant with a random-zero sphygmomanometer after a 5-minute rest. Because of substantial intraindividual variation in blood pressure measurements, the first blood pressure reading was dropped and the second and third readings were averaged for use in this study. Study participants were asked to bring all medications, vitamins, and supplements taken in the 2 weeks before the examination. Information on pharmacological treatment of hypertension and elevated glucose was based on the participant\u2019s self-reported use of any medication to treat high blood pressure or elevated glucose and the transcription and coding of all medication names.\n\n###### Statistical Analysis\nAll analyses were performed with SAS (version 9.1, SAS Institute Inc, Cary, NC). Principal-components analysis (PROC FACTOR) was used to derive dietary patterns and to determine factor loadings for each of the 29 food subgroups. Factors were rotated with varimax rotation to maintain uncorrelated factors and enhance interpretability. After evaluation of the eigenvalues and interpretability, all values \u22652.0 were retained, resulting in a 2-factor solution, with factor 1 explaining 10.2% of the variability and factor 2 explaining 9.7%. Dietary patterns were named according to the nature of the food groups loading highest for each of the factors. Interval-censored Cox proportional-hazards regression was used to assess the relation between diet and incident MetSyn. The dietary patterns and major food groups (meat, dairy, fruit and vegetables, whole grains, and refined grains) were categorized into quintiles; fried foods, nuts, sweetened beverages, diet soda, and coffee were classified into tertiles, given their low level of consumption. Hazard ratios (HRs) were obtained by entering the quintiles (or tertiles) into the model as indicator variables, whereas tests for trend across the quintiles were conducted by entering the median level of intake at each quintile (or tertile) into the model as a continuous variable. Cumulative average diet, which has been shown to increase the precision of dietary data, was used to create food groups and patterns. Specifically, the time-varying dietary covariates were modeled as follows: Between baseline and examination 3, dietary exposures were based on diet as measured at the baseline examination; after examination 3, dietary exposures were based on the mean of baseline diet and examination 3 diet. Spearman partial correlation coefficients, adjusted for age, sex, race, education, center, and total energy intake, were computed between examinations 1 and 3 for meat (r=0.42), dairy (r=0.42), fruit and vegetables (r=0.50), whole grains (r=0.45), and refined grains (r=0.36) (P<0.001 for all). Considering well-known within-person variation in response to diet questionnaires of the FFQ type, the correlations suggest considerable tracking (lack of change over time), although they also could represent some change in diet over 6 years. Model 1 adjusted for baseline age, sex, race, center, education (less than high school, high school or vocational, college and higher), and energy intake (kcal/d); model 2 further adjusted for behavioral characteristics, including smoking (current, former, never), pack-years (continuous), and physical activity (continuous). For analyses of individual food groups, a third model was considered that also adjusted for the major food groups: meat, dairy, fruit and vegetables, whole grains, and refined grains. The authors had full access to and take responsibility for the integrity of the data. All authors have read and agree to the manuscript as written.",
        "results": "#### Results\nThe cohort of 9514 ARIC participants had a mean age at baseline of 53.6 years, and 44.1% were male. Over 9 years of follow-up, 3782 participants (39.8%) developed incident MetSyn. On diagnosis, the most frequent contributor to MetSyn was abdominal obesity (79.8%), followed by elevated fasting glucose (77.5%), hypertension (62.4%), low HDL (58.5%), and high levels of triglycerides (48.5%). Sex-stratified baseline demographic, behavioral, and dietary variables are presented in Table 1. Factor analysis revealed 2 main dietary patterns for the population under study; food groups and factor loadings for those with absolute values of \u22650.20 are presented in Table 2. Factor 1, identified as a Western dietary pattern, was characterized by high intakes of refined grains, processed meat, and red meat, while the prudent dietary pattern was typified by greater consumption of cruciferous and carotinoid vegetables, fruit, fish, and poultry. Higher scores for the Western dietary pattern were associated with greater risk of developing MetSyn, even after adjustment for behavioral characteristics (Ptrend=0.03) (Table 3). Participants in the highest quintile of Western dietary pattern scores had an 18% greater risk (HR, 1.18; 95% CI, 1.03 to 1.37) of developing MetSyn than those in the lowest quintile. Consumption of a prudent dietary pattern was not associated with incident MetSyn in either model. In the analyses of major food groups, results from models 1 and 2 were similar; thus, only model 1 is shown in the tables. After adjustment for demographics, behaviors, and dietary pattern scores, meat consumption was positively associated with higher incidence of MetSyn (Ptrend=0.001), whereas dairy consumption conferred some protection (Ptrend=0.001) (Table 3). There was no association between incident MetSyn and consumption of whole grains, refined grains, or fruits and vegetables. The HR for the consumption of extreme quintiles of meat was 1.26 (95% CI, 1.11 to 1.43). Individual foods in the meat group also were adversely associated with MetSyn: hamburger (Ptrend=0.002), hot dogs (Ptrend=0.002), and processed meats (Ptrend=0.03). The HR for the extreme quintiles of dairy consumption was 0.87 (95% CI, 0.77 to 0.98); however, individual dairy foods were not associated with incident MetSyn. As shown in Table 4, consumption of fried foods was associated with increased risk of MetSyn (Ptrend=0.02). The HR for the extreme tertiles was 1.25 (95% CI, 1.14 to 1.37). Sweetened beverages also were associated with incident MetSyn after base adjustments (HR, 1.17; 95% CI, 1.07 to 1.28; Ptrend=0.02), although the association became nonsignificant with additional adjustments. Diet soda intake was strongly associated with increased risk across all models (HR, 1.34; 95% CI, 1.11 to 1.29). No relations were observed.\n\nIn sensitivity analyses, similar results were observed when diet relations with MetSyn were examined even after exclusion of individuals with prevalent diabetes at baseline (n=1870) and after omitting individuals diagnosed with incident MetSyn at 1 examination who reverted at a later examination (n=1161). We also evaluated whether gender or baseline overweight status modified the relation between dietary intake and incident MetSyn, but no interactions were observed.",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Dietary intake and the development of the metabolic syndrome: the Atherosclerosis Risk in Communities study.",
        "abstract": "## BACKGROUND\nThe role of diet in the origin of metabolic syndrome (MetSyn) is not well understood; thus, we sought to evaluate the relationship between incident MetSyn and dietary intake using prospective data from 9514 participants (age, 45 to 64 years) enrolled in the Atherosclerosis Risk in Communities (ARIC) study.\n## METHODS AND RESULTS\nDietary intake was assessed at baseline via a 66-item food frequency questionnaire. We used principal-components analysis to derive \"Western\" and \"prudent\" dietary patterns from 32 food groups and evaluated 10 food groups used in previous studies of the ARIC cohort. MetSyn was defined by American Heart Association guidelines. Proportional-hazards regression was used. Over 9 years of follow-up, 3782 incident cases of MetSyn were identified. After adjustment for demographic factors, smoking, physical activity, and energy intake, consumption of a Western dietary pattern (P(trend)=0.03) was adversely associated with incident MetSyn. After further adjustment for intake of meat, dairy, fruits and vegetables, refined grains, and whole grains, analysis of individual food groups revealed that meat (P(trend)<0.001), fried foods (P(trend)=0.02), and diet soda (P(trend)=< 0.001) also were adversely associated with incident MetSyn, whereas dairy consumption (P(trend)=0.006) was beneficial. No associations were observed between incident MetSyn and a prudent dietary pattern or intakes of whole grains, refined grains, fruits and vegetables, nuts, coffee, or sweetened beverages.\n## CONCLUSIONS\nThese prospective findings suggest that consumption of a Western dietary pattern, meat, and fried foods promotes the incidence of MetSyn, whereas dairy consumption provides some protection. The diet soda association was not hypothesized and deserves further study.",
        "methods": "#### Methods\n###### Study Population\nThe ARIC study is a multicenter population-based prospective cohort study designed to investigate the origin and natural history of atherosclerosis in middle-aged adults. Participants were recruited from 4 US communities: Forsyth County, North Carolina; Jackson, Miss; suburbs of Minneapolis, Minn; and Washington County, Maryland. The study cohort included 15 972 white and black men and women 45 to 64 years of age at baseline in 1987 to 1989 (visit 1). Cohort reexaminations took place at 3-year intervals; response rates were 93%, 86%, and 81% at visits 2 (1990 to 1992), 3 (1993 to 1995), and 4 (1996 to 1998), respectively. Local institutional review boards approved the ARIC protocol, and all subjects gave informed consent. All participants with MetSyn at baseline were excluded (n=5879), as were those with cardiovascular disease (n=776), missing diet data (n=11), or implausible energy intakes defined as \u2264500 or \u22643500 kcal/d in women and \u2264700 or \u22644500 kcal/d in men (n=361). The exclusion criteria were not mutually exclusive; thus, data from 9514 participants were included in the analyses.\n\n###### Data Collection\nARIC participants underwent interviews, fasting venipuncture, and measurement of blood pressure and anthropometrics at each examination. Trained interviewers ascertained basic demographic data, medical history, and information about personal habits, including diet (2 exams only), smoking, physical activity (with the Baecke et al questionnaire), and medication use.\n\n###### Dietary Assessment\nAt baseline and examination 3 (6 years later), usual dietary intake was assessed by a 66-item interviewer-administered semiquantitative food-frequency questionnaire (FFQ). The questionnaire was a modified version of the 61-item instrument developed by Willett et al. Participants reported the frequency of consumption of each food on the basis of 9 levels of frequency, ranging from never or \u22641 time a month to \u22656 times a day. Standard serving sizes and food models were provided as a reference for intake estimation. Interviewers also obtained additional information, including the brand name of the breakfast cereal usually consumed. Food and beverages from the FFQ were categorized into 29 food subgroups, which were used to derive dietary patterns via principal-components analysis. The subgroups were further collapsed into 5 major food groups: meat, dairy, fruits and vegetables, refined grains, and whole grains (see the online Data Supplement). Additionally, associations with the following individual food items were assessed: fried foods, sweetened beverages (regular soda and fruit drinks), diet soda, nuts, and coffee.\n\n###### Metabolic Syndrome\nMetSyn, defined according to current American Heart Association guidelines, was characterized by the presence of any 3 of the following risk factors: waist circumference \u2265102 cm in men and \u226588 cm in women; systolic blood pressure \u2265130 mm Hg, diastolic blood pressure \u226585 mm Hg, or current use of antihypertensive medication; and fasting glucose \u2265100 mg/dL or current drug treatment for elevated glucose. Waist circumference was measured with a cloth tape at the umbilicus. Fasting blood samples were drawn from an antecubital vein into tubes containing EDTA (lipids) or a serum separator gel (glucose). Total plasma cholesterol and triglycerides were determined by enzymatic methods. HDL cholesterol was measured after dextran-magnesium precipitation. Serum glucose levels were assessed with a hexokinase/glucose-6-phosphate dehydrogenase method. Sitting blood pressure was measured by trained technicians 3 times on each participant with a random-zero sphygmomanometer after a 5-minute rest. Because of substantial intraindividual variation in blood pressure measurements, the first blood pressure reading was dropped and the second and third readings were averaged for use in this study. Study participants were asked to bring all medications, vitamins, and supplements taken in the 2 weeks before the examination. Information on pharmacological treatment of hypertension and elevated glucose was based on the participant\u2019s self-reported use of any medication to treat high blood pressure or elevated glucose and the transcription and coding of all medication names.\n\n###### Statistical Analysis\nAll analyses were performed with SAS (version 9.1, SAS Institute Inc, Cary, NC). Principal-components analysis (PROC FACTOR) was used to derive dietary patterns and to determine factor loadings for each of the 29 food subgroups. Factors were rotated with varimax rotation to maintain uncorrelated factors and enhance interpretability. After evaluation of the eigenvalues and interpretability, all values \u22652.0 were retained, resulting in a 2-factor solution, with factor 1 explaining 10.2% of the variability and factor 2 explaining 9.7%. Dietary patterns were named according to the nature of the food groups loading highest for each of the factors. Interval-censored Cox proportional-hazards regression was used to assess the relation between diet and incident MetSyn. The dietary patterns and major food groups (meat, dairy, fruit and vegetables, whole grains, and refined grains) were categorized into quintiles; fried foods, nuts, sweetened beverages, diet soda, and coffee were classified into tertiles, given their low level of consumption. Hazard ratios (HRs) were obtained by entering the quintiles (or tertiles) into the model as indicator variables, whereas tests for trend across the quintiles were conducted by entering the median level of intake at each quintile (or tertile) into the model as a continuous variable. Cumulative average diet, which has been shown to increase the precision of dietary data, was used to create food groups and patterns. Specifically, the time-varying dietary covariates were modeled as follows: Between baseline and examination 3, dietary exposures were based on diet as measured at the baseline examination; after examination 3, dietary exposures were based on the mean of baseline diet and examination 3 diet. Spearman partial correlation coefficients, adjusted for age, sex, race, education, center, and total energy intake, were computed between examinations 1 and 3 for meat (r=0.42), dairy (r=0.42), fruit and vegetables (r=0.50), whole grains (r=0.45), and refined grains (r=0.36) (P<0.001 for all). Considering well-known within-person variation in response to diet questionnaires of the FFQ type, the correlations suggest considerable tracking (lack of change over time), although they also could represent some change in diet over 6 years. Model 1 adjusted for baseline age, sex, race, center, education (less than high school, high school or vocational, college and higher), and energy intake (kcal/d); model 2 further adjusted for behavioral characteristics, including smoking (current, former, never), pack-years (continuous), and physical activity (continuous). For analyses of individual food groups, a third model was considered that also adjusted for the major food groups: meat, dairy, fruit and vegetables, whole grains, and refined grains. The authors had full access to and take responsibility for the integrity of the data. All authors have read and agree to the manuscript as written.",
        "results": "#### Results\nThe cohort of 9514 ARIC participants had a mean age at baseline of 53.6 years, and 44.1% were male. Over 9 years of follow-up, 3782 participants (39.8%) developed incident MetSyn. On diagnosis, the most frequent contributor to MetSyn was abdominal obesity (79.8%), followed by elevated fasting glucose (77.5%), hypertension (62.4%), low HDL (58.5%), and high levels of triglycerides (48.5%). Sex-stratified baseline demographic, behavioral, and dietary variables are presented in Table 1. Factor analysis revealed 2 main dietary patterns for the population under study; food groups and factor loadings for those with absolute values of \u22650.20 are presented in Table 2. Factor 1, identified as a Western dietary pattern, was characterized by high intakes of refined grains, processed meat, and red meat, while the prudent dietary pattern was typified by greater consumption of cruciferous and carotinoid vegetables, fruit, fish, and poultry. Higher scores for the Western dietary pattern were associated with greater risk of developing MetSyn, even after adjustment for behavioral characteristics (Ptrend=0.03) (Table 3). Participants in the highest quintile of Western dietary pattern scores had an 18% greater risk (HR, 1.18; 95% CI, 1.03 to 1.37) of developing MetSyn than those in the lowest quintile. Consumption of a prudent dietary pattern was not associated with incident MetSyn in either model. In the analyses of major food groups, results from models 1 and 2 were similar; thus, only model 1 is shown in the tables. After adjustment for demographics, behaviors, and dietary pattern scores, meat consumption was positively associated with higher incidence of MetSyn (Ptrend=0.001), whereas dairy consumption conferred some protection (Ptrend=0.001) (Table 3). There was no association between incident MetSyn and consumption of whole grains, refined grains, or fruits and vegetables. The HR for the consumption of extreme quintiles of meat was 1.26 (95% CI, 1.11 to 1.43). Individual foods in the meat group also were adversely associated with MetSyn: hamburger (Ptrend=0.002), hot dogs (Ptrend=0.002), and processed meats (Ptrend=0.03). The HR for the extreme quintiles of dairy consumption was 0.87 (95% CI, 0.77 to 0.98); however, individual dairy foods were not associated with incident MetSyn. As shown in Table 4, consumption of fried foods was associated with increased risk of MetSyn (Ptrend=0.02). The HR for the extreme tertiles was 1.25 (95% CI, 1.14 to 1.37). Sweetened beverages also were associated with incident MetSyn after base adjustments (HR, 1.17; 95% CI, 1.07 to 1.28; Ptrend=0.02), although the association became nonsignificant with additional adjustments. Diet soda intake was strongly associated with increased risk across all models (HR, 1.34; 95% CI, 1.11 to 1.29). No relations were observed.\n\nIn sensitivity analyses, similar results were observed when diet relations with MetSyn were examined even after exclusion of individuals with prevalent diabetes at baseline (n=1870) and after omitting individuals diagnosed with incident MetSyn at 1 examination who reverted at a later examination (n=1161). We also evaluated whether gender or baseline overweight status modified the relation between dietary intake and incident MetSyn, but no interactions were observed.",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Effects of energy-dense nutrient-poor snacks on the incidence of metabolic syndrome: a prospective approach in Tehran Lipid and Glucose Study.",
        "abstract": "## OBJECTIVES\nIncreased consumption of energy-dense, nutrient-poor snacks is one of the major, growing concerns in relation to the alarming trend of overweight, obesity, and metabolic disorders worldwide. The aim of this study was to investigate whether consumption of energy-dense snacks could affect the occurrence of metabolic syndrome after 3 y of follow-up in adults.\n## METHODS\nThis longitudinal study was conducted within the framework of the Tehran Lipid and Glucose Study between 2006 and 2008 and 2009 and 2011, on 1466 adults, ages 19 to 70 y. The usual intake of participants was measured using a validated semiquantitative food frequency questionnaire at baseline. Biochemical and anthropometric measurements were assessed at baseline and 3 y later. Multiple logistic regression models were used to estimate the occurrence of metabolic syndrome (MetS) in each quartile of energy-dense snacks.\n## RESULTS\nParticipants in the highest quartile of energy-dense snack consumption were significantly younger (33.8 versus 43.1 y; P < 0.01). There were non-significant increased risks for MetS in the highest quartile of biscuits and cakes (odds ratio [OR], 1.13; 95% confidence interval [CI], 0.77-1.66), and candies and chocolates (OR, 1.31; 95% CI, 0.88-1.94). The risk for MetS in participants with highest consumption of salty snacks increased more than 50% (OR, 1.56; 95% CI, 1.01-2.40). Consumption of soft drinks had a borderline effect on the risk for MetS. More than 361 kcal/d from total energy-dense snacks independently increased the occurrence of MetS in the fourth compared the first quartile category (OR, 1.53; 95% CI, 1.03-2.29).\n## CONCLUSION\nThe findings of this study demonstrated that higher consumption of energy-dense snacks could be a dietary risk factor for development of MetS.",
        "methods": "### Methods\n\nStudy population\n\nThis study was conducted within the framework of the Tehran Lipid and Glucose Study (TLGS) in a 3-y follow-up period; baseline examination of the present study was adapted on Phase III of the TLGS (2006\u20132008), and the outcome examination was adapted Phase IV of the TLGS (2009\u20132011). Briefly, the TLGS is a community-based prospective study being conducted to investigate and prevent non-communicable diseases, in a representative sample of residents, ages \u22653 y, from district 13 of Tehran, the capital city of Iran. The first phase of the TLGS began in March 1999 to December 2001 and data collection, at 3-y intervals, is ongoing. In the first phase, a multistage stratified cluster random sampling technique was used to select more than 15,000 people ages \u22653 y from Tehran urban district 13. Phases II, III, and IV were prospective follow-up studies were conducted in 2002 to 2004, 2006 to 2008, and 2009 to 2011, respectively. During Phase III, 12,523 participants were examined. Of these, 4920 participants were randomly selected for dietary assessment, categorized by age and sex; dietary data were available for 3462 participants. Of these, 2799 adults (1129 men and 1438 women) ages 19 to 70 y with completed data (demographic characteristics, anthropometrics, biochemical, and dietary data) were included in baseline examination of the present study. Participants were excluded from the final analysis if they reported implausible energy intakes (<800 kcal/d or \u22654200 kcal/d), were on specific diets due to diabetes, cardiovascular disease, other non-communicable disease or any disorder (n = 262), had no follow-up information on anthropometrics and biochemical measurements at the second examination of this study (Phase IV of TLGS [2009\u20132011]) (n = 629), or were consistent with the definition of MetS at baseline (n = 432). Finally, 1476 participants were included in the analysis. The mean duration of the follow-up was approximately 3 y. Informed written consents were obtained from all participants and the study protocol was approved by the research council of the Research Institute for Endocrine Sciences, Shahid Beheshti University of Medical Sciences. MetS components were defined according to the diagnostic criteria proposed by NCEP ATP III, and new cut-off points of waist circumference for Iranian adults. Participants were considered to have MetS at baseline if they have at least three of the following metabolic abnormalities:\n\n1. Hyperglycemia as fasting plasma glucose (FPG) \u2265100 mg/dL (5.6 mmol/L) or drug treatment of impaired fasting glucose;\n2. Hypertriacylglycerolmia as serum triacylglycerols (TGs) \u2265150 mg/dL (1.69 mmol/L) or drug treatment;\n3. Low high-density lipoprotein cholesterol (HDL-C) as serum HDL-C <40 mg/dL (1.04 mmol/L) for men, and <50 mg/dL (1.29 mmol/L) for women or drug treatment;\n4. Hypertension as blood pressure (BP) \u2265130/85 mm Hg or drug treatment for hypertension; and\n5. Abdominal obesity as waist circumference (WC) \u226595 cm for men and women.\n\nThe incidence of metabolic syndrome after 3 y of follow-up was determined by the same criteria used at baseline. After 3 y, 249 participants were diagnosed.\n\nData collection\n\nAt baseline, some known or suspected risk factors for MetS including age, smoking status, educational level, body mass index (BMI), WC, and physical activity were assessed. Smoking status was obtained using face-to-face interviews and individuals who smoked daily or occasionally were considered current smokers, whereas non-smokers included those who had never smoked or those who had quit smoking. Weight was measured to the nearest 100 g using digital scales while the individuals were minimally clothed, without shoes. Height was measured to the nearest 0.5 cm, in a standing position without shoes, using a tape measure. BMI was calculated as weight (kg) divided by square of the height (m\u00b2). Physical activity level was assessed based on the frequency and time spent on light-, moderate-, high-, and very high-intensity activities according to the list of common activities of daily life over the past year. Physical activity levels were expressed as metabolic equivalent hours per week (METs h/wk). To identify MetS, its components including WC, FPG, serum TGs, HDL-C, and BP were assessed at baseline and again after 3 y. WC was measured to the nearest 0.1 cm (at anatomic landmarks), at the widest portion, over light clothing, using a soft, tape measure without any pressure to the body. Fasting blood samples were taken after 12 to 14 h, from all study participants at baseline and after a 3-y follow-up. FPG was measured by the enzymatic colorimetric method using glucose oxidase. TG level was measured by enzymatic colorimetric analysis with glycerol phosphate oxidase. HDL-C was measured after precipitation of the apolipoprotein B containing lipoproteins with phosphotungstic acid. Analyses were performed using Pars Azmoon kits (Pars Azmoon Inc., Tehran, Iran) and a Selectra 2 auto-analyzer (Vital Scientific, Spankeren, Netherlands). Inter- and intraassay coefficient of variation of all assays was <5%. After a 15-min rest in the sitting position, two measurements of BP were taken, on the right arm, using a standardized mercury sphygmomanometer, and the mean of the two measurements was considered as the participant\u2019s BP. A 168-item food frequency questionnaire (FFQ) was used to assess typical food intakes over the previous year. The validity of the FFQ was previously evaluated by comparing food groups and nutrient values determined from the questionnaire with values estimated from the average of twelve 24-h dietary recall surveys. Trained dietitians, with at least 5 y of experience in the TLGS survey, asked participants to designate their intake frequency for each food item consumed during the past year on a daily, weekly, or monthly basis. Portion sizes of consumed foods reported in household measures were then converted to grams. Snacks were determined as foods that were not consumed as part of the main daily meals and the principal dietary exposure of interest was considered daily energy intake from most common energy-dense snacks, including biscuits, crackers, cakes, cookies, candies, chocolates, potato chips, puff snacks, and soft drink. Foods with low or medium energy density and high nutritional value, such as fruits and vegetables, dairy products, and cereals were not considered as snacks. To analyze foods and beverages for their energy and nutrient contents we used the U.S. Department of Agriculture food composition tables (USDA national nutrient database for standard reference) because the Iranian FCT is incomplete, and has limited data on nutrient content of raw foods and beverages. Participants\u2019 dietary intakes, including dietary energy and energy density, macronutrients, micronutrients, and food groups, were determined. Dietary phytochemical index and dietary total antioxidant capacity were calculated as important background of diet quality.\n\nStatistical analysis\n\nAll statistical analyses were conducted using SPSS (Version 16.0; Chicago, IL, USA), and P-values at <0.05 were considered significant. Daily consumption of energy-dense snacks at baseline and the categories of snacks were divided into quartiles based on their 25th, 50th, and 75th percentile values. Total daily consumption of snacks in the first, second, third, and fourth quartiles were \u226466, 66 to 119, 120 to 213, and >213 kcal/d. Participant characteristics, baseline MetS components and 3-y changes, and the prevalence of MetS at follow-up examination were compared across quartile categories of energy-dense snack consumption, using the \u03c7\u00b2 test or general linear models with adjustment for sex and age. Mean dietary intakes of participants were compared across quartile categories of snack consumption using the general linear model with adjustment for sex, age (y, continuous), and energy intakes (kcal/d). To estimate the odds ratio of MetS at the second examination in each quartile category of energy-dense snack consumption, multivariable logistic regression models with adjustment for sex, age (y, continuous), BMI (kg/m\u00b2, continuous), education (four categories), smoking (yes or no), physical activity (MET-h/wk, continuous), total energy intake (kcal/d), energy density (kcal/100 g of foods), total fiber (g/d), dietary intake of whole grains (g/d), fruits (g/d), vegetables (g/d), dairy (g/d), total meat (g/d), phytochemical index, and dietary total antioxidant capacity (mmolTE/100 g). To assess the overall trends of the odds ratios of MetS across quartiles, the median of energy-dense snacks consumption in each quartile was used.",
        "results": "##### Results\n\nThe mean age of participants was 37.8 \u00b1 12.3 y, and mean BMI was 26 \u00b1 4.5 kg/m\u00b2 at baseline. Thirty-nine percent of the participants were men. The mean weight gain was 1.87 \u00b1 5.14 kg (2.20 \u00b1 5.3 kg in men and 1.65 \u00b1 4.93 kg in women) during the 3-y period; there was no significant difference in mean weight gain across quartiles of energy-dense snack consumption. The mean daily calorie intake from energy-dense snacks was 163 \u00b1 155 kcal (173 \u00b1 163 and 157 \u00b1 149 kcal, in men and women, respectively). The incidence of MetS after 3 y follow-up was 16.9%. Compared with participants in the lowest quartile, those in the highest quartile of energy-dense snack consumption were significantly younger (33.8 versus 43.1 y; P < 0.01). There was no significant difference between baseline characteristics of participants and MetS components across categories of energy-dense snacks (Table 1). The mean dietary intakes of participants across quartile categories of energy-dense snack consumption are shown in Table 2. Daily energy intake and dietary energy density, as well as intake of total fat, simple sugars, and sodium were significantly higher in participants with the highest consumption of energy-dense snacks (P < 0.05). There was a significant decreasing trend in dietary intake of carbohydrates, protein, calcium, zinc, selenium, and total fiber across increasing consumption of energy-dense snacks (P < 0.05). Participants with the highest compared with the lowest consumption of energy-dense snacks consumed less whole grains, vegetables, and dairy products (P < 0.05). There was a decreasing trend in the dietary phytochemical index (percent of energy intake from phytochemical-rich foods) along with increased consumption of energy-dense snacks (P < 0.05). The odds ratio and 95% confidence interval for occurrence of the MetS in each quartile category of energy-dense snacks after the 3-y follow-up are presented in Table 3. After adjustment for potential confounding variables, there was a moderate non-significant increase in the risk for MetS in the highest quartile categories of consumption of both biscuits and cakes (odds ratio [OR], 1.13; 95% confidence interval [CI], 0.77\u20131.66), and candies and chocolates (OR, 1.31; 95% CI, 0.88\u20131.94). The risk for MetS in participants with the highest consumption of salty snacks significantly increased the risk by 56% after the 3-y follow-up (OR, 1.56; 95% CI, 1.01\u20132.40). Consumption of soft drinks and carbonated beverages was also associated with a borderline increase in the risk for MetS in the fourth quartile category. Total consumption of energy-dense snacks independently increased the occurrence of MetS in the fourth quartile category (OR, 1.53; 95% CI, 1.03\u20132.29); also, there was a significant increasing trend for the occurrence of MetS across increasing total consumption of energy-dense snacks (P for trend < 0.05).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Effects of energy-dense nutrient-poor snacks on the incidence of metabolic syndrome: a prospective approach in Tehran Lipid and Glucose Study.",
        "abstract": "## OBJECTIVES\nIncreased consumption of energy-dense, nutrient-poor snacks is one of the major, growing concerns in relation to the alarming trend of overweight, obesity, and metabolic disorders worldwide. The aim of this study was to investigate whether consumption of energy-dense snacks could affect the occurrence of metabolic syndrome after 3 y of follow-up in adults.\n## METHODS\nThis longitudinal study was conducted within the framework of the Tehran Lipid and Glucose Study between 2006 and 2008 and 2009 and 2011, on 1466 adults, ages 19 to 70 y. The usual intake of participants was measured using a validated semiquantitative food frequency questionnaire at baseline. Biochemical and anthropometric measurements were assessed at baseline and 3 y later. Multiple logistic regression models were used to estimate the occurrence of metabolic syndrome (MetS) in each quartile of energy-dense snacks.\n## RESULTS\nParticipants in the highest quartile of energy-dense snack consumption were significantly younger (33.8 versus 43.1 y; P < 0.01). There were non-significant increased risks for MetS in the highest quartile of biscuits and cakes (odds ratio [OR], 1.13; 95% confidence interval [CI], 0.77-1.66), and candies and chocolates (OR, 1.31; 95% CI, 0.88-1.94). The risk for MetS in participants with highest consumption of salty snacks increased more than 50% (OR, 1.56; 95% CI, 1.01-2.40). Consumption of soft drinks had a borderline effect on the risk for MetS. More than 361 kcal/d from total energy-dense snacks independently increased the occurrence of MetS in the fourth compared the first quartile category (OR, 1.53; 95% CI, 1.03-2.29).\n## CONCLUSION\nThe findings of this study demonstrated that higher consumption of energy-dense snacks could be a dietary risk factor for development of MetS.",
        "methods": "### Methods\n\nStudy population\n\nThis study was conducted within the framework of the Tehran Lipid and Glucose Study (TLGS) in a 3-y follow-up period; baseline examination of the present study was adapted on Phase III of the TLGS (2006\u20132008), and the outcome examination was adapted Phase IV of the TLGS (2009\u20132011). Briefly, the TLGS is a community-based prospective study being conducted to investigate and prevent non-communicable diseases, in a representative sample of residents, ages \u22653 y, from district 13 of Tehran, the capital city of Iran. The first phase of the TLGS began in March 1999 to December 2001 and data collection, at 3-y intervals, is ongoing. In the first phase, a multistage stratified cluster random sampling technique was used to select more than 15,000 people ages \u22653 y from Tehran urban district 13. Phases II, III, and IV were prospective follow-up studies were conducted in 2002 to 2004, 2006 to 2008, and 2009 to 2011, respectively. During Phase III, 12,523 participants were examined. Of these, 4920 participants were randomly selected for dietary assessment, categorized by age and sex; dietary data were available for 3462 participants. Of these, 2799 adults (1129 men and 1438 women) ages 19 to 70 y with completed data (demographic characteristics, anthropometrics, biochemical, and dietary data) were included in baseline examination of the present study. Participants were excluded from the final analysis if they reported implausible energy intakes (<800 kcal/d or \u22654200 kcal/d), were on specific diets due to diabetes, cardiovascular disease, other non-communicable disease or any disorder (n = 262), had no follow-up information on anthropometrics and biochemical measurements at the second examination of this study (Phase IV of TLGS [2009\u20132011]) (n = 629), or were consistent with the definition of MetS at baseline (n = 432). Finally, 1476 participants were included in the analysis. The mean duration of the follow-up was approximately 3 y. Informed written consents were obtained from all participants and the study protocol was approved by the research council of the Research Institute for Endocrine Sciences, Shahid Beheshti University of Medical Sciences. MetS components were defined according to the diagnostic criteria proposed by NCEP ATP III, and new cut-off points of waist circumference for Iranian adults. Participants were considered to have MetS at baseline if they have at least three of the following metabolic abnormalities:\n\n1. Hyperglycemia as fasting plasma glucose (FPG) \u2265100 mg/dL (5.6 mmol/L) or drug treatment of impaired fasting glucose;\n2. Hypertriacylglycerolmia as serum triacylglycerols (TGs) \u2265150 mg/dL (1.69 mmol/L) or drug treatment;\n3. Low high-density lipoprotein cholesterol (HDL-C) as serum HDL-C <40 mg/dL (1.04 mmol/L) for men, and <50 mg/dL (1.29 mmol/L) for women or drug treatment;\n4. Hypertension as blood pressure (BP) \u2265130/85 mm Hg or drug treatment for hypertension; and\n5. Abdominal obesity as waist circumference (WC) \u226595 cm for men and women.\n\nThe incidence of metabolic syndrome after 3 y of follow-up was determined by the same criteria used at baseline. After 3 y, 249 participants were diagnosed.\n\nData collection\n\nAt baseline, some known or suspected risk factors for MetS including age, smoking status, educational level, body mass index (BMI), WC, and physical activity were assessed. Smoking status was obtained using face-to-face interviews and individuals who smoked daily or occasionally were considered current smokers, whereas non-smokers included those who had never smoked or those who had quit smoking. Weight was measured to the nearest 100 g using digital scales while the individuals were minimally clothed, without shoes. Height was measured to the nearest 0.5 cm, in a standing position without shoes, using a tape measure. BMI was calculated as weight (kg) divided by square of the height (m\u00b2). Physical activity level was assessed based on the frequency and time spent on light-, moderate-, high-, and very high-intensity activities according to the list of common activities of daily life over the past year. Physical activity levels were expressed as metabolic equivalent hours per week (METs h/wk). To identify MetS, its components including WC, FPG, serum TGs, HDL-C, and BP were assessed at baseline and again after 3 y. WC was measured to the nearest 0.1 cm (at anatomic landmarks), at the widest portion, over light clothing, using a soft, tape measure without any pressure to the body. Fasting blood samples were taken after 12 to 14 h, from all study participants at baseline and after a 3-y follow-up. FPG was measured by the enzymatic colorimetric method using glucose oxidase. TG level was measured by enzymatic colorimetric analysis with glycerol phosphate oxidase. HDL-C was measured after precipitation of the apolipoprotein B containing lipoproteins with phosphotungstic acid. Analyses were performed using Pars Azmoon kits (Pars Azmoon Inc., Tehran, Iran) and a Selectra 2 auto-analyzer (Vital Scientific, Spankeren, Netherlands). Inter- and intraassay coefficient of variation of all assays was <5%. After a 15-min rest in the sitting position, two measurements of BP were taken, on the right arm, using a standardized mercury sphygmomanometer, and the mean of the two measurements was considered as the participant\u2019s BP. A 168-item food frequency questionnaire (FFQ) was used to assess typical food intakes over the previous year. The validity of the FFQ was previously evaluated by comparing food groups and nutrient values determined from the questionnaire with values estimated from the average of twelve 24-h dietary recall surveys. Trained dietitians, with at least 5 y of experience in the TLGS survey, asked participants to designate their intake frequency for each food item consumed during the past year on a daily, weekly, or monthly basis. Portion sizes of consumed foods reported in household measures were then converted to grams. Snacks were determined as foods that were not consumed as part of the main daily meals and the principal dietary exposure of interest was considered daily energy intake from most common energy-dense snacks, including biscuits, crackers, cakes, cookies, candies, chocolates, potato chips, puff snacks, and soft drink. Foods with low or medium energy density and high nutritional value, such as fruits and vegetables, dairy products, and cereals were not considered as snacks. To analyze foods and beverages for their energy and nutrient contents we used the U.S. Department of Agriculture food composition tables (USDA national nutrient database for standard reference) because the Iranian FCT is incomplete, and has limited data on nutrient content of raw foods and beverages. Participants\u2019 dietary intakes, including dietary energy and energy density, macronutrients, micronutrients, and food groups, were determined. Dietary phytochemical index and dietary total antioxidant capacity were calculated as important background of diet quality.\n\nStatistical analysis\n\nAll statistical analyses were conducted using SPSS (Version 16.0; Chicago, IL, USA), and P-values at <0.05 were considered significant. Daily consumption of energy-dense snacks at baseline and the categories of snacks were divided into quartiles based on their 25th, 50th, and 75th percentile values. Total daily consumption of snacks in the first, second, third, and fourth quartiles were \u226466, 66 to 119, 120 to 213, and >213 kcal/d. Participant characteristics, baseline MetS components and 3-y changes, and the prevalence of MetS at follow-up examination were compared across quartile categories of energy-dense snack consumption, using the \u03c7\u00b2 test or general linear models with adjustment for sex and age. Mean dietary intakes of participants were compared across quartile categories of snack consumption using the general linear model with adjustment for sex, age (y, continuous), and energy intakes (kcal/d). To estimate the odds ratio of MetS at the second examination in each quartile category of energy-dense snack consumption, multivariable logistic regression models with adjustment for sex, age (y, continuous), BMI (kg/m\u00b2, continuous), education (four categories), smoking (yes or no), physical activity (MET-h/wk, continuous), total energy intake (kcal/d), energy density (kcal/100 g of foods), total fiber (g/d), dietary intake of whole grains (g/d), fruits (g/d), vegetables (g/d), dairy (g/d), total meat (g/d), phytochemical index, and dietary total antioxidant capacity (mmolTE/100 g). To assess the overall trends of the odds ratios of MetS across quartiles, the median of energy-dense snacks consumption in each quartile was used.",
        "results": "##### Results\n\nThe mean age of participants was 37.8 \u00b1 12.3 y, and mean BMI was 26 \u00b1 4.5 kg/m\u00b2 at baseline. Thirty-nine percent of the participants were men. The mean weight gain was 1.87 \u00b1 5.14 kg (2.20 \u00b1 5.3 kg in men and 1.65 \u00b1 4.93 kg in women) during the 3-y period; there was no significant difference in mean weight gain across quartiles of energy-dense snack consumption. The mean daily calorie intake from energy-dense snacks was 163 \u00b1 155 kcal (173 \u00b1 163 and 157 \u00b1 149 kcal, in men and women, respectively). The incidence of MetS after 3 y follow-up was 16.9%. Compared with participants in the lowest quartile, those in the highest quartile of energy-dense snack consumption were significantly younger (33.8 versus 43.1 y; P < 0.01). There was no significant difference between baseline characteristics of participants and MetS components across categories of energy-dense snacks (Table 1). The mean dietary intakes of participants across quartile categories of energy-dense snack consumption are shown in Table 2. Daily energy intake and dietary energy density, as well as intake of total fat, simple sugars, and sodium were significantly higher in participants with the highest consumption of energy-dense snacks (P < 0.05). There was a significant decreasing trend in dietary intake of carbohydrates, protein, calcium, zinc, selenium, and total fiber across increasing consumption of energy-dense snacks (P < 0.05). Participants with the highest compared with the lowest consumption of energy-dense snacks consumed less whole grains, vegetables, and dairy products (P < 0.05). There was a decreasing trend in the dietary phytochemical index (percent of energy intake from phytochemical-rich foods) along with increased consumption of energy-dense snacks (P < 0.05). The odds ratio and 95% confidence interval for occurrence of the MetS in each quartile category of energy-dense snacks after the 3-y follow-up are presented in Table 3. After adjustment for potential confounding variables, there was a moderate non-significant increase in the risk for MetS in the highest quartile categories of consumption of both biscuits and cakes (odds ratio [OR], 1.13; 95% confidence interval [CI], 0.77\u20131.66), and candies and chocolates (OR, 1.31; 95% CI, 0.88\u20131.94). The risk for MetS in participants with the highest consumption of salty snacks significantly increased the risk by 56% after the 3-y follow-up (OR, 1.56; 95% CI, 1.01\u20132.40). Consumption of soft drinks and carbonated beverages was also associated with a borderline increase in the risk for MetS in the fourth quartile category. Total consumption of energy-dense snacks independently increased the occurrence of MetS in the fourth quartile category (OR, 1.53; 95% CI, 1.03\u20132.29); also, there was a significant increasing trend for the occurrence of MetS across increasing total consumption of energy-dense snacks (P for trend < 0.05).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Effects of energy-dense nutrient-poor snacks on the incidence of metabolic syndrome: a prospective approach in Tehran Lipid and Glucose Study.",
        "abstract": "## OBJECTIVES\nIncreased consumption of energy-dense, nutrient-poor snacks is one of the major, growing concerns in relation to the alarming trend of overweight, obesity, and metabolic disorders worldwide. The aim of this study was to investigate whether consumption of energy-dense snacks could affect the occurrence of metabolic syndrome after 3 y of follow-up in adults.\n## METHODS\nThis longitudinal study was conducted within the framework of the Tehran Lipid and Glucose Study between 2006 and 2008 and 2009 and 2011, on 1466 adults, ages 19 to 70 y. The usual intake of participants was measured using a validated semiquantitative food frequency questionnaire at baseline. Biochemical and anthropometric measurements were assessed at baseline and 3 y later. Multiple logistic regression models were used to estimate the occurrence of metabolic syndrome (MetS) in each quartile of energy-dense snacks.\n## RESULTS\nParticipants in the highest quartile of energy-dense snack consumption were significantly younger (33.8 versus 43.1 y; P < 0.01). There were non-significant increased risks for MetS in the highest quartile of biscuits and cakes (odds ratio [OR], 1.13; 95% confidence interval [CI], 0.77-1.66), and candies and chocolates (OR, 1.31; 95% CI, 0.88-1.94). The risk for MetS in participants with highest consumption of salty snacks increased more than 50% (OR, 1.56; 95% CI, 1.01-2.40). Consumption of soft drinks had a borderline effect on the risk for MetS. More than 361 kcal/d from total energy-dense snacks independently increased the occurrence of MetS in the fourth compared the first quartile category (OR, 1.53; 95% CI, 1.03-2.29).\n## CONCLUSION\nThe findings of this study demonstrated that higher consumption of energy-dense snacks could be a dietary risk factor for development of MetS.",
        "methods": "### Methods\n\nStudy population\n\nThis study was conducted within the framework of the Tehran Lipid and Glucose Study (TLGS) in a 3-y follow-up period; baseline examination of the present study was adapted on Phase III of the TLGS (2006\u20132008), and the outcome examination was adapted Phase IV of the TLGS (2009\u20132011). Briefly, the TLGS is a community-based prospective study being conducted to investigate and prevent non-communicable diseases, in a representative sample of residents, ages \u22653 y, from district 13 of Tehran, the capital city of Iran. The first phase of the TLGS began in March 1999 to December 2001 and data collection, at 3-y intervals, is ongoing. In the first phase, a multistage stratified cluster random sampling technique was used to select more than 15,000 people ages \u22653 y from Tehran urban district 13. Phases II, III, and IV were prospective follow-up studies were conducted in 2002 to 2004, 2006 to 2008, and 2009 to 2011, respectively. During Phase III, 12,523 participants were examined. Of these, 4920 participants were randomly selected for dietary assessment, categorized by age and sex; dietary data were available for 3462 participants. Of these, 2799 adults (1129 men and 1438 women) ages 19 to 70 y with completed data (demographic characteristics, anthropometrics, biochemical, and dietary data) were included in baseline examination of the present study. Participants were excluded from the final analysis if they reported implausible energy intakes (<800 kcal/d or \u22654200 kcal/d), were on specific diets due to diabetes, cardiovascular disease, other non-communicable disease or any disorder (n = 262), had no follow-up information on anthropometrics and biochemical measurements at the second examination of this study (Phase IV of TLGS [2009\u20132011]) (n = 629), or were consistent with the definition of MetS at baseline (n = 432). Finally, 1476 participants were included in the analysis. The mean duration of the follow-up was approximately 3 y. Informed written consents were obtained from all participants and the study protocol was approved by the research council of the Research Institute for Endocrine Sciences, Shahid Beheshti University of Medical Sciences. MetS components were defined according to the diagnostic criteria proposed by NCEP ATP III, and new cut-off points of waist circumference for Iranian adults. Participants were considered to have MetS at baseline if they have at least three of the following metabolic abnormalities:\n\n1. Hyperglycemia as fasting plasma glucose (FPG) \u2265100 mg/dL (5.6 mmol/L) or drug treatment of impaired fasting glucose;\n2. Hypertriacylglycerolmia as serum triacylglycerols (TGs) \u2265150 mg/dL (1.69 mmol/L) or drug treatment;\n3. Low high-density lipoprotein cholesterol (HDL-C) as serum HDL-C <40 mg/dL (1.04 mmol/L) for men, and <50 mg/dL (1.29 mmol/L) for women or drug treatment;\n4. Hypertension as blood pressure (BP) \u2265130/85 mm Hg or drug treatment for hypertension; and\n5. Abdominal obesity as waist circumference (WC) \u226595 cm for men and women.\n\nThe incidence of metabolic syndrome after 3 y of follow-up was determined by the same criteria used at baseline. After 3 y, 249 participants were diagnosed.\n\nData collection\n\nAt baseline, some known or suspected risk factors for MetS including age, smoking status, educational level, body mass index (BMI), WC, and physical activity were assessed. Smoking status was obtained using face-to-face interviews and individuals who smoked daily or occasionally were considered current smokers, whereas non-smokers included those who had never smoked or those who had quit smoking. Weight was measured to the nearest 100 g using digital scales while the individuals were minimally clothed, without shoes. Height was measured to the nearest 0.5 cm, in a standing position without shoes, using a tape measure. BMI was calculated as weight (kg) divided by square of the height (m\u00b2). Physical activity level was assessed based on the frequency and time spent on light-, moderate-, high-, and very high-intensity activities according to the list of common activities of daily life over the past year. Physical activity levels were expressed as metabolic equivalent hours per week (METs h/wk). To identify MetS, its components including WC, FPG, serum TGs, HDL-C, and BP were assessed at baseline and again after 3 y. WC was measured to the nearest 0.1 cm (at anatomic landmarks), at the widest portion, over light clothing, using a soft, tape measure without any pressure to the body. Fasting blood samples were taken after 12 to 14 h, from all study participants at baseline and after a 3-y follow-up. FPG was measured by the enzymatic colorimetric method using glucose oxidase. TG level was measured by enzymatic colorimetric analysis with glycerol phosphate oxidase. HDL-C was measured after precipitation of the apolipoprotein B containing lipoproteins with phosphotungstic acid. Analyses were performed using Pars Azmoon kits (Pars Azmoon Inc., Tehran, Iran) and a Selectra 2 auto-analyzer (Vital Scientific, Spankeren, Netherlands). Inter- and intraassay coefficient of variation of all assays was <5%. After a 15-min rest in the sitting position, two measurements of BP were taken, on the right arm, using a standardized mercury sphygmomanometer, and the mean of the two measurements was considered as the participant\u2019s BP. A 168-item food frequency questionnaire (FFQ) was used to assess typical food intakes over the previous year. The validity of the FFQ was previously evaluated by comparing food groups and nutrient values determined from the questionnaire with values estimated from the average of twelve 24-h dietary recall surveys. Trained dietitians, with at least 5 y of experience in the TLGS survey, asked participants to designate their intake frequency for each food item consumed during the past year on a daily, weekly, or monthly basis. Portion sizes of consumed foods reported in household measures were then converted to grams. Snacks were determined as foods that were not consumed as part of the main daily meals and the principal dietary exposure of interest was considered daily energy intake from most common energy-dense snacks, including biscuits, crackers, cakes, cookies, candies, chocolates, potato chips, puff snacks, and soft drink. Foods with low or medium energy density and high nutritional value, such as fruits and vegetables, dairy products, and cereals were not considered as snacks. To analyze foods and beverages for their energy and nutrient contents we used the U.S. Department of Agriculture food composition tables (USDA national nutrient database for standard reference) because the Iranian FCT is incomplete, and has limited data on nutrient content of raw foods and beverages. Participants\u2019 dietary intakes, including dietary energy and energy density, macronutrients, micronutrients, and food groups, were determined. Dietary phytochemical index and dietary total antioxidant capacity were calculated as important background of diet quality.\n\nStatistical analysis\n\nAll statistical analyses were conducted using SPSS (Version 16.0; Chicago, IL, USA), and P-values at <0.05 were considered significant. Daily consumption of energy-dense snacks at baseline and the categories of snacks were divided into quartiles based on their 25th, 50th, and 75th percentile values. Total daily consumption of snacks in the first, second, third, and fourth quartiles were \u226466, 66 to 119, 120 to 213, and >213 kcal/d. Participant characteristics, baseline MetS components and 3-y changes, and the prevalence of MetS at follow-up examination were compared across quartile categories of energy-dense snack consumption, using the \u03c7\u00b2 test or general linear models with adjustment for sex and age. Mean dietary intakes of participants were compared across quartile categories of snack consumption using the general linear model with adjustment for sex, age (y, continuous), and energy intakes (kcal/d). To estimate the odds ratio of MetS at the second examination in each quartile category of energy-dense snack consumption, multivariable logistic regression models with adjustment for sex, age (y, continuous), BMI (kg/m\u00b2, continuous), education (four categories), smoking (yes or no), physical activity (MET-h/wk, continuous), total energy intake (kcal/d), energy density (kcal/100 g of foods), total fiber (g/d), dietary intake of whole grains (g/d), fruits (g/d), vegetables (g/d), dairy (g/d), total meat (g/d), phytochemical index, and dietary total antioxidant capacity (mmolTE/100 g). To assess the overall trends of the odds ratios of MetS across quartiles, the median of energy-dense snacks consumption in each quartile was used.",
        "results": "##### Results\n\nThe mean age of participants was 37.8 \u00b1 12.3 y, and mean BMI was 26 \u00b1 4.5 kg/m\u00b2 at baseline. Thirty-nine percent of the participants were men. The mean weight gain was 1.87 \u00b1 5.14 kg (2.20 \u00b1 5.3 kg in men and 1.65 \u00b1 4.93 kg in women) during the 3-y period; there was no significant difference in mean weight gain across quartiles of energy-dense snack consumption. The mean daily calorie intake from energy-dense snacks was 163 \u00b1 155 kcal (173 \u00b1 163 and 157 \u00b1 149 kcal, in men and women, respectively). The incidence of MetS after 3 y follow-up was 16.9%. Compared with participants in the lowest quartile, those in the highest quartile of energy-dense snack consumption were significantly younger (33.8 versus 43.1 y; P < 0.01). There was no significant difference between baseline characteristics of participants and MetS components across categories of energy-dense snacks (Table 1). The mean dietary intakes of participants across quartile categories of energy-dense snack consumption are shown in Table 2. Daily energy intake and dietary energy density, as well as intake of total fat, simple sugars, and sodium were significantly higher in participants with the highest consumption of energy-dense snacks (P < 0.05). There was a significant decreasing trend in dietary intake of carbohydrates, protein, calcium, zinc, selenium, and total fiber across increasing consumption of energy-dense snacks (P < 0.05). Participants with the highest compared with the lowest consumption of energy-dense snacks consumed less whole grains, vegetables, and dairy products (P < 0.05). There was a decreasing trend in the dietary phytochemical index (percent of energy intake from phytochemical-rich foods) along with increased consumption of energy-dense snacks (P < 0.05). The odds ratio and 95% confidence interval for occurrence of the MetS in each quartile category of energy-dense snacks after the 3-y follow-up are presented in Table 3. After adjustment for potential confounding variables, there was a moderate non-significant increase in the risk for MetS in the highest quartile categories of consumption of both biscuits and cakes (odds ratio [OR], 1.13; 95% confidence interval [CI], 0.77\u20131.66), and candies and chocolates (OR, 1.31; 95% CI, 0.88\u20131.94). The risk for MetS in participants with the highest consumption of salty snacks significantly increased the risk by 56% after the 3-y follow-up (OR, 1.56; 95% CI, 1.01\u20132.40). Consumption of soft drinks and carbonated beverages was also associated with a borderline increase in the risk for MetS in the fourth quartile category. Total consumption of energy-dense snacks independently increased the occurrence of MetS in the fourth quartile category (OR, 1.53; 95% CI, 1.03\u20132.29); also, there was a significant increasing trend for the occurrence of MetS across increasing total consumption of energy-dense snacks (P for trend < 0.05).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Consumption of sugar sweetened beverage is associated with incidence of metabolic syndrome in Tehranian children and adolescents.",
        "abstract": "## BACKGROUND\nIntakes of high sugar-sweetened beverages (SSBs) in adults can escalate risk of metabolic syndrome (MetS); however, data of longitudinal studies in children and adolescents are lacking. In this study we assessed consumption of SSBs in relation to incidence of MetS among children and adolescents during a 3.6\u00a0year follow-up.\n## METHODS\nThis study was a population-based longitudinal study, in which 424 subjects, aged 6-18 years, from the Tehran Lipid and Glucose Study with complete data on dietary intake, blood pressure, anthropometry, and biochemical indices were followed for 3.6\u00a0years. Dietary intake was collected using a valid and reliable food frequency questionnaire. MetS was defined according to the Cook criteria. Sugar sweetened beverages included all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks.\n## RESULTS\nAverage daily intakes of SSSD and fruit juice drinks were 38.5\u2009\u00b1\u200975.0 and 32.3\u2009\u00b1\u200960.1\u00a0g, respectively. After adjustment for confounders, compared to the first quartile, the odds ratio of incident MetS in the highest quartile of SSB and SSSD was 3.20 (95\u00a0% CI: 1.06-9.90) and 3.01 (95\u00a0% CI: 1.17-7.74), respectively. Regarding incidence of MetS components, compared with the lowest quartile, the highest quartile of SSSDs showed odds ratios of 2.49 (95\u00a0% CI: 1.00-6.53) for abdominal obesity and 2.79 (95\u00a0% CI: 1.02-7.64) for hypertension. No significant association was found between consumption of fruit juice drink and SSSD with other components of MetS.\n## CONCLUSIONS\nChildren and adolescents with high intakes of carbonated beverages could be at increased risk of MetS, abdominal obesity, and hypertension.",
        "methods": "## Methods\n\nThis study was a population-based longitudinal study, in which 424 subjects, aged 6\u201318 years, from the Tehran Lipid and Glucose Study with complete data on dietary intake, blood pressure, anthropometry, and biochemical indices were followed for 3.6 years. Dietary intake was collected using a valid and reliable food frequency questionnaire. MetS was defined according to the Cook criteria. Sugar sweetened beverages included all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks.\n\nTehran Lipid and Glucose Study (TLGS) is a prospective community-based study being conducted to determine the risk factors of and prevent non-communicable diseases by implementing and promoting healthy lifestyles.\n\nBaseline data were collected from 15005 participants, aged 3 years and over, of residents District No. 13 of Tehran under coverage of three medical health centers. The participants were followed up every 3 years to update their data on demographic, lifestyle, biochemical, clinical, and dietary measurements; the baseline survey was a cross-sectional study conducted from 1999 to 2001, and surveys 2 (2002\u20132005), 3 (2006\u20132008), and 4 (2009\u20132011) were prospective follow-up surveys.\n\nIn the third survey of the TLGS (2006\u20132008), from among 12523 examined participants, 3462 were randomly selected for dietary assessment; of these, 621 individuals were aged 6\u201318 y who included in the current study. Participants who had incomplete dietary intakes or had missing measures of MetS components (n = 29) were excluded. Additionally, participants with reported energy intakes (kcal/day) to energy requirements ratios beyond \u00b13SD range, were also excluded (n = 6) because they fell into the over-or under-reporting categories. Participants were followed until survey 4 (2009\u20132011), an average of 3.6 years (follow-up rate: 86 %). Depending on the outcome variable under analysis, we excluded persons whose data at survey 3 (2006\u20132008) indicated prevalent MetS (n = 69), hypertension (n = 53), high triglycerides (TGs) (n = 168), low high density lipoprotein cholesterol (HDL-C) (n = 242), high fasting plasma glucose (FPG) (n = 12), or abdominal obesity (n = 145); some individuals fell into more than one exclusion category. The final sample sizes in fully adjusted models varied by outcomes as follows: MetS (n = 424), abdominal obesity (n = 327), high triglycerides (TGs) (n = 347), hypertension (n = 439), low HDL-C (n = 290), and high fasting plasma glucose (n = 476).\n\nThe design of this study was approved by the institutional ethics committee of the Research Institute for Endocrine Sciences, Shahid Beheshti University of Medical Sciences, and written informed consent was obtained from participants\u2019 parents.\n\nDietary intake\nA valid and reliable 168-item semi-quantitative food frequency questionnaire (FFQ) was used by trained dietitians, during face-to face interviews, to evaluate the usual dietary intakes of participants. Participants were asked to report their consumption frequency of food items during the previous year on a daily, weekly, or monthly basis. Mothers were asked about the type and quantity of meals and snacks when children were unable to recall. Portion sizes of consumed foods were specified according to the US Department of Agriculture (USDA) standard portion sizes or household measures (e.g. beans, 1 tablespoon; chicken meat, 1 leg or wing; rice, 1 large or small plate). Obtained data were converted to mean daily intake in grams or milliliters. This FFQ included questions on the frequency of consumption and usual portion size of all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks (both 100 % fruit juice and sugar sweetened synthetic juice drinks that are not 100 % juice). According to Althuis et al., these items were combined to estimate the daily intake of SSBs (ml/d). SSB intakes were divided by using quartile cutoffs, with the assumption that one serving was equivalent to 1 cup (250 mL).\n\nTo study the reproducibility of the FFQ, 132 subjects (61 men and 71 women) twice completed a 168-item FFQ (FFQ1, FFQ2), with a 14-month interval between FFQ1 and FFQ2, and to assess the validity, 12 dietary recalls (DRs) were collected (1 each month) over the 1-year interval. The age-and energy-adjusted and deattenuated Spearman correlation coefficients used to assess validity of the FFQ for soft drinks were 0.54 and 0.37 in men and women, respectively. The age-and energy-adjusted intraclass correlation coefficients, which reflect the reproducibility of food groups in the FFQ, were 0.61 and 0.65 in men and in women, respectively.\n\nMeasurements\nInformation about physical activity was collected using the modifiable activity questionnaire (MAQ) to calculate metabolic equivalent task (MET) minutes per week. High reliability (98 %) and moderate validity (47 %) were ascertained for the MAQ translated into Persian. Levels of physical activity were categorized as light (MET <600 min/wk), moderate (MET 600\u20131,499 min/wk), and vigorous (MET \u22651,500 min/wk).\n\nWeight was measured, while participants were minimally clothed without shoes, using digital scales (Seca 707, Seca Corp., Hanover, MD; range 0.1\u2013150 kg) and recorded to the nearest 100 g. Height was measured in a standing position without shoes, using a stadiometer, with shoulders in normal alignment. Body mass index (BMI) was calculated as weight (kg) divided by the square of height (m[2]). Waist circumference (WC) was measured at the umbilicus, using a measuring tape, without pressure to body surfaces, and was recorded to the nearest 0.5 cm.\n\nBlood pressure was measured twice, after participants were seated for 15 min, using a standard mercury sphygmomanometer; there was at least a 30-s interval between these two separate measurements, and the mean of two measurements was considered the patient\u2019s blood pressure.\n\nBlood samples, at baseline and follow up, were drawn between 7:00 and 9:00 AM from all study participants after 12\u201314 h overnight fasting; all blood analyses were done at the TLGS research laboratory on the day of blood collection. FPG was measured by the enzymatic colorimetric method using glucose oxidase. Serum HDL-C was measured after precipitation of the apolipoprotein B-containing lipoproteins with phosphotungstic acid and serum TGs were assayed using an enzymatic colorimetric method with glycerol phosphate oxidase. These analyses were performed using commercial kits (Pars Azmoon Inc., Tehran, Iran) and a Selectra 2 auto analyzer (Vital Scientific, Spankeren, The Netherlands). Inter-and intra-assay coefficients of variations at baseline were both 2.2 % for FPG, 2 and 0.5 % for HDL-C and 1.6 and 0.6 % for TGs, respectively.\n\nDefinitions\nIn children and adolescents, MetS was defined according to definition proposed by Cook et al. as 3 or more of the following: Fasting TGs \u2265110 mg/dl; HDL-C < 40 mg/dl; WC \u2265 90th percentile for age and sex, according to national reference curves; systolic blood pressure and diastolic blood pressure \u226590th percentile for sex, age and height, from the National Heart, Lung, and Blood Institute\u2019s recommended cut-off points, and FPG \u2265 100 mg/dl, according to the recommendations of American Diabetes Association. In subjects aged > 18 years after follow-up, MetS was defined according to the joint interim statement as the presence of any 3 of 5 risk factors of the following: 1) Abdominal obesity as WC \u226591 and men \u226589 cm for women and men, respectively, according to Iranians cut-off points; 2) FPG \u2265100 mg/dl or drug treatment; 3) fasting TGs \u2265150 mg/dl or drug treatment; 4) fasting HDL-C <50 mg/dl for women and <40 mg/dl for men or drug treatment, and 5) hypertension was defined as systolic blood pressure \u2265130 mmHg, diastolic blood pressure \u226585 mmHg, or antihypertensive drug treatment.\n\nStatistical analysis\nAll analyses were conducted using the Statistical Package for Social Sciences (version 15.0; SPSS Inc, Chicago IL). Baseline characteristics of subjects were expressed as mean \u00b1 SD or median (interquartile 25\u201375) for continuous variables and percentages for categorical variables. To investigate the trend of continuous and categorical variables according to quartiles of SSB, linear regression and chi-square test were used, respectively. There were no significant differences in baseline anthropometric and biochemical assessments between subjects who provided follow-up assessments and those lost to follow-up (data not shown).\n\nIntakes of SSBs (<19.2, 19.2-43.8, 43.9-91.5, and >91.6 ml/day), SSSDs (<7.6, 7.6\u201315.3, 15.3\u201340.0, and >40.0 ml/day), and fruit juice drinks (<4.1, 4.1\u201313.5, 13.5\u201334.9, and 34.9 ml/day) were categorized into the quartile cut off points. To estimate the risk of 3.6-year incident outcomes, separate multiple regression models were used for abdominal obesity, high FPG, high TGs, low HDL-C, hypertension, and MetS with quartiles of SSB, SSSD, and fruit juice drink consumption; odds ratios (OR) and 95 % confidence interval (CI) were reported. The initial model was adjusted for age, sex, total energy intake, physical activity, and family history of diabetes. Further adjustment for dietary fiber, tea and coffee, red and processed meat, fruits, and vegetables was made and the final model was additionally adjusted for BMI.",
        "results": "**Results**\n\nThe mean \u00b1 SD age of study population (68 % girls) was 13.6 \u00b1 3.7 years; the mean daily SSB intake was 98 ml in boys and 70 ml in girls. After a 3.6 year follow up, cumulative incidence of MetS was 11 %.\n\nThe descriptive characteristics of the study population according to quartile of SSB consumption are shown in Table 1. Participants with higher SSB consumption were likely to have higher systolic blood pressure, FPG, and BMI; in addition, subjects with higher SSB consumption had higher intakes of daily energy, saturated fatty acids, trans-fatty acids, glucose, fructose, sucrose, simple sugars, higher glycemic index, and lower intakes of fruits and whole grains.\n\nSugar sweetened beverage intake was inversely related to MetS and its components after adjustment for potential confounders (Table 2). Further adjustment for dietary factors did not have a major impact on the results. In this model, subjects in the highest quartile of SSB consumption had increased risk of incident MetS (OR: 3.16; 95 % CI: 1.02\u20139.78, P for trend = 0.007) and abdominal obesity (OR: 2.97; 95 % CI: 1.23\u20137.19, P for trend = 0.017). Adjustment for BMI slightly strengthened the association; adjusted multivariable ORs of MetS, abdominal obesity, and hypertension were 3.50, 3.66, and 2.90 (all P for trends < 0.05).\n\nHigher intakes of SSSD was associated with significant increase in the risk of incident MetS (OR: 3.58, 95 % CI: 1.09-11.78, P for trend = 0.013), abdominal obesity (OR: 3.78, 95 % CI: 1.08\u201313.27, P for trend = 0.043), and hypertension (OR: 2.74, 95 % CI: 1.05\u20137.19, P for trend = 0.018) in the fully adjusted model. Associations for fruit juice drink intake were not statistically significant (Table 3).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Consumption of sugar sweetened beverage is associated with incidence of metabolic syndrome in Tehranian children and adolescents.",
        "abstract": "## BACKGROUND\nIntakes of high sugar-sweetened beverages (SSBs) in adults can escalate risk of metabolic syndrome (MetS); however, data of longitudinal studies in children and adolescents are lacking. In this study we assessed consumption of SSBs in relation to incidence of MetS among children and adolescents during a 3.6\u00a0year follow-up.\n## METHODS\nThis study was a population-based longitudinal study, in which 424 subjects, aged 6-18 years, from the Tehran Lipid and Glucose Study with complete data on dietary intake, blood pressure, anthropometry, and biochemical indices were followed for 3.6\u00a0years. Dietary intake was collected using a valid and reliable food frequency questionnaire. MetS was defined according to the Cook criteria. Sugar sweetened beverages included all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks.\n## RESULTS\nAverage daily intakes of SSSD and fruit juice drinks were 38.5\u2009\u00b1\u200975.0 and 32.3\u2009\u00b1\u200960.1\u00a0g, respectively. After adjustment for confounders, compared to the first quartile, the odds ratio of incident MetS in the highest quartile of SSB and SSSD was 3.20 (95\u00a0% CI: 1.06-9.90) and 3.01 (95\u00a0% CI: 1.17-7.74), respectively. Regarding incidence of MetS components, compared with the lowest quartile, the highest quartile of SSSDs showed odds ratios of 2.49 (95\u00a0% CI: 1.00-6.53) for abdominal obesity and 2.79 (95\u00a0% CI: 1.02-7.64) for hypertension. No significant association was found between consumption of fruit juice drink and SSSD with other components of MetS.\n## CONCLUSIONS\nChildren and adolescents with high intakes of carbonated beverages could be at increased risk of MetS, abdominal obesity, and hypertension.",
        "methods": "## Methods\n\nThis study was a population-based longitudinal study, in which 424 subjects, aged 6\u201318 years, from the Tehran Lipid and Glucose Study with complete data on dietary intake, blood pressure, anthropometry, and biochemical indices were followed for 3.6 years. Dietary intake was collected using a valid and reliable food frequency questionnaire. MetS was defined according to the Cook criteria. Sugar sweetened beverages included all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks.\n\nTehran Lipid and Glucose Study (TLGS) is a prospective community-based study being conducted to determine the risk factors of and prevent non-communicable diseases by implementing and promoting healthy lifestyles.\n\nBaseline data were collected from 15005 participants, aged 3 years and over, of residents District No. 13 of Tehran under coverage of three medical health centers. The participants were followed up every 3 years to update their data on demographic, lifestyle, biochemical, clinical, and dietary measurements; the baseline survey was a cross-sectional study conducted from 1999 to 2001, and surveys 2 (2002\u20132005), 3 (2006\u20132008), and 4 (2009\u20132011) were prospective follow-up surveys.\n\nIn the third survey of the TLGS (2006\u20132008), from among 12523 examined participants, 3462 were randomly selected for dietary assessment; of these, 621 individuals were aged 6\u201318 y who included in the current study. Participants who had incomplete dietary intakes or had missing measures of MetS components (n = 29) were excluded. Additionally, participants with reported energy intakes (kcal/day) to energy requirements ratios beyond \u00b13SD range, were also excluded (n = 6) because they fell into the over-or under-reporting categories. Participants were followed until survey 4 (2009\u20132011), an average of 3.6 years (follow-up rate: 86 %). Depending on the outcome variable under analysis, we excluded persons whose data at survey 3 (2006\u20132008) indicated prevalent MetS (n = 69), hypertension (n = 53), high triglycerides (TGs) (n = 168), low high density lipoprotein cholesterol (HDL-C) (n = 242), high fasting plasma glucose (FPG) (n = 12), or abdominal obesity (n = 145); some individuals fell into more than one exclusion category. The final sample sizes in fully adjusted models varied by outcomes as follows: MetS (n = 424), abdominal obesity (n = 327), high triglycerides (TGs) (n = 347), hypertension (n = 439), low HDL-C (n = 290), and high fasting plasma glucose (n = 476).\n\nThe design of this study was approved by the institutional ethics committee of the Research Institute for Endocrine Sciences, Shahid Beheshti University of Medical Sciences, and written informed consent was obtained from participants\u2019 parents.\n\nDietary intake\nA valid and reliable 168-item semi-quantitative food frequency questionnaire (FFQ) was used by trained dietitians, during face-to face interviews, to evaluate the usual dietary intakes of participants. Participants were asked to report their consumption frequency of food items during the previous year on a daily, weekly, or monthly basis. Mothers were asked about the type and quantity of meals and snacks when children were unable to recall. Portion sizes of consumed foods were specified according to the US Department of Agriculture (USDA) standard portion sizes or household measures (e.g. beans, 1 tablespoon; chicken meat, 1 leg or wing; rice, 1 large or small plate). Obtained data were converted to mean daily intake in grams or milliliters. This FFQ included questions on the frequency of consumption and usual portion size of all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks (both 100 % fruit juice and sugar sweetened synthetic juice drinks that are not 100 % juice). According to Althuis et al., these items were combined to estimate the daily intake of SSBs (ml/d). SSB intakes were divided by using quartile cutoffs, with the assumption that one serving was equivalent to 1 cup (250 mL).\n\nTo study the reproducibility of the FFQ, 132 subjects (61 men and 71 women) twice completed a 168-item FFQ (FFQ1, FFQ2), with a 14-month interval between FFQ1 and FFQ2, and to assess the validity, 12 dietary recalls (DRs) were collected (1 each month) over the 1-year interval. The age-and energy-adjusted and deattenuated Spearman correlation coefficients used to assess validity of the FFQ for soft drinks were 0.54 and 0.37 in men and women, respectively. The age-and energy-adjusted intraclass correlation coefficients, which reflect the reproducibility of food groups in the FFQ, were 0.61 and 0.65 in men and in women, respectively.\n\nMeasurements\nInformation about physical activity was collected using the modifiable activity questionnaire (MAQ) to calculate metabolic equivalent task (MET) minutes per week. High reliability (98 %) and moderate validity (47 %) were ascertained for the MAQ translated into Persian. Levels of physical activity were categorized as light (MET <600 min/wk), moderate (MET 600\u20131,499 min/wk), and vigorous (MET \u22651,500 min/wk).\n\nWeight was measured, while participants were minimally clothed without shoes, using digital scales (Seca 707, Seca Corp., Hanover, MD; range 0.1\u2013150 kg) and recorded to the nearest 100 g. Height was measured in a standing position without shoes, using a stadiometer, with shoulders in normal alignment. Body mass index (BMI) was calculated as weight (kg) divided by the square of height (m[2]). Waist circumference (WC) was measured at the umbilicus, using a measuring tape, without pressure to body surfaces, and was recorded to the nearest 0.5 cm.\n\nBlood pressure was measured twice, after participants were seated for 15 min, using a standard mercury sphygmomanometer; there was at least a 30-s interval between these two separate measurements, and the mean of two measurements was considered the patient\u2019s blood pressure.\n\nBlood samples, at baseline and follow up, were drawn between 7:00 and 9:00 AM from all study participants after 12\u201314 h overnight fasting; all blood analyses were done at the TLGS research laboratory on the day of blood collection. FPG was measured by the enzymatic colorimetric method using glucose oxidase. Serum HDL-C was measured after precipitation of the apolipoprotein B-containing lipoproteins with phosphotungstic acid and serum TGs were assayed using an enzymatic colorimetric method with glycerol phosphate oxidase. These analyses were performed using commercial kits (Pars Azmoon Inc., Tehran, Iran) and a Selectra 2 auto analyzer (Vital Scientific, Spankeren, The Netherlands). Inter-and intra-assay coefficients of variations at baseline were both 2.2 % for FPG, 2 and 0.5 % for HDL-C and 1.6 and 0.6 % for TGs, respectively.\n\nDefinitions\nIn children and adolescents, MetS was defined according to definition proposed by Cook et al. as 3 or more of the following: Fasting TGs \u2265110 mg/dl; HDL-C < 40 mg/dl; WC \u2265 90th percentile for age and sex, according to national reference curves; systolic blood pressure and diastolic blood pressure \u226590th percentile for sex, age and height, from the National Heart, Lung, and Blood Institute\u2019s recommended cut-off points, and FPG \u2265 100 mg/dl, according to the recommendations of American Diabetes Association. In subjects aged > 18 years after follow-up, MetS was defined according to the joint interim statement as the presence of any 3 of 5 risk factors of the following: 1) Abdominal obesity as WC \u226591 and men \u226589 cm for women and men, respectively, according to Iranians cut-off points; 2) FPG \u2265100 mg/dl or drug treatment; 3) fasting TGs \u2265150 mg/dl or drug treatment; 4) fasting HDL-C <50 mg/dl for women and <40 mg/dl for men or drug treatment, and 5) hypertension was defined as systolic blood pressure \u2265130 mmHg, diastolic blood pressure \u226585 mmHg, or antihypertensive drug treatment.\n\nStatistical analysis\nAll analyses were conducted using the Statistical Package for Social Sciences (version 15.0; SPSS Inc, Chicago IL). Baseline characteristics of subjects were expressed as mean \u00b1 SD or median (interquartile 25\u201375) for continuous variables and percentages for categorical variables. To investigate the trend of continuous and categorical variables according to quartiles of SSB, linear regression and chi-square test were used, respectively. There were no significant differences in baseline anthropometric and biochemical assessments between subjects who provided follow-up assessments and those lost to follow-up (data not shown).\n\nIntakes of SSBs (<19.2, 19.2-43.8, 43.9-91.5, and >91.6 ml/day), SSSDs (<7.6, 7.6\u201315.3, 15.3\u201340.0, and >40.0 ml/day), and fruit juice drinks (<4.1, 4.1\u201313.5, 13.5\u201334.9, and 34.9 ml/day) were categorized into the quartile cut off points. To estimate the risk of 3.6-year incident outcomes, separate multiple regression models were used for abdominal obesity, high FPG, high TGs, low HDL-C, hypertension, and MetS with quartiles of SSB, SSSD, and fruit juice drink consumption; odds ratios (OR) and 95 % confidence interval (CI) were reported. The initial model was adjusted for age, sex, total energy intake, physical activity, and family history of diabetes. Further adjustment for dietary fiber, tea and coffee, red and processed meat, fruits, and vegetables was made and the final model was additionally adjusted for BMI.",
        "results": "**Results**\n\nThe mean \u00b1 SD age of study population (68 % girls) was 13.6 \u00b1 3.7 years; the mean daily SSB intake was 98 ml in boys and 70 ml in girls. After a 3.6 year follow up, cumulative incidence of MetS was 11 %.\n\nThe descriptive characteristics of the study population according to quartile of SSB consumption are shown in Table 1. Participants with higher SSB consumption were likely to have higher systolic blood pressure, FPG, and BMI; in addition, subjects with higher SSB consumption had higher intakes of daily energy, saturated fatty acids, trans-fatty acids, glucose, fructose, sucrose, simple sugars, higher glycemic index, and lower intakes of fruits and whole grains.\n\nSugar sweetened beverage intake was inversely related to MetS and its components after adjustment for potential confounders (Table 2). Further adjustment for dietary factors did not have a major impact on the results. In this model, subjects in the highest quartile of SSB consumption had increased risk of incident MetS (OR: 3.16; 95 % CI: 1.02\u20139.78, P for trend = 0.007) and abdominal obesity (OR: 2.97; 95 % CI: 1.23\u20137.19, P for trend = 0.017). Adjustment for BMI slightly strengthened the association; adjusted multivariable ORs of MetS, abdominal obesity, and hypertension were 3.50, 3.66, and 2.90 (all P for trends < 0.05).\n\nHigher intakes of SSSD was associated with significant increase in the risk of incident MetS (OR: 3.58, 95 % CI: 1.09-11.78, P for trend = 0.013), abdominal obesity (OR: 3.78, 95 % CI: 1.08\u201313.27, P for trend = 0.043), and hypertension (OR: 2.74, 95 % CI: 1.05\u20137.19, P for trend = 0.018) in the fully adjusted model. Associations for fruit juice drink intake were not statistically significant (Table 3).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Consumption of sugar sweetened beverage is associated with incidence of metabolic syndrome in Tehranian children and adolescents.",
        "abstract": "## BACKGROUND\nIntakes of high sugar-sweetened beverages (SSBs) in adults can escalate risk of metabolic syndrome (MetS); however, data of longitudinal studies in children and adolescents are lacking. In this study we assessed consumption of SSBs in relation to incidence of MetS among children and adolescents during a 3.6\u00a0year follow-up.\n## METHODS\nThis study was a population-based longitudinal study, in which 424 subjects, aged 6-18 years, from the Tehran Lipid and Glucose Study with complete data on dietary intake, blood pressure, anthropometry, and biochemical indices were followed for 3.6\u00a0years. Dietary intake was collected using a valid and reliable food frequency questionnaire. MetS was defined according to the Cook criteria. Sugar sweetened beverages included all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks.\n## RESULTS\nAverage daily intakes of SSSD and fruit juice drinks were 38.5\u2009\u00b1\u200975.0 and 32.3\u2009\u00b1\u200960.1\u00a0g, respectively. After adjustment for confounders, compared to the first quartile, the odds ratio of incident MetS in the highest quartile of SSB and SSSD was 3.20 (95\u00a0% CI: 1.06-9.90) and 3.01 (95\u00a0% CI: 1.17-7.74), respectively. Regarding incidence of MetS components, compared with the lowest quartile, the highest quartile of SSSDs showed odds ratios of 2.49 (95\u00a0% CI: 1.00-6.53) for abdominal obesity and 2.79 (95\u00a0% CI: 1.02-7.64) for hypertension. No significant association was found between consumption of fruit juice drink and SSSD with other components of MetS.\n## CONCLUSIONS\nChildren and adolescents with high intakes of carbonated beverages could be at increased risk of MetS, abdominal obesity, and hypertension.",
        "methods": "## Methods\n\nThis study was a population-based longitudinal study, in which 424 subjects, aged 6\u201318 years, from the Tehran Lipid and Glucose Study with complete data on dietary intake, blood pressure, anthropometry, and biochemical indices were followed for 3.6 years. Dietary intake was collected using a valid and reliable food frequency questionnaire. MetS was defined according to the Cook criteria. Sugar sweetened beverages included all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks.\n\nTehran Lipid and Glucose Study (TLGS) is a prospective community-based study being conducted to determine the risk factors of and prevent non-communicable diseases by implementing and promoting healthy lifestyles.\n\nBaseline data were collected from 15005 participants, aged 3 years and over, of residents District No. 13 of Tehran under coverage of three medical health centers. The participants were followed up every 3 years to update their data on demographic, lifestyle, biochemical, clinical, and dietary measurements; the baseline survey was a cross-sectional study conducted from 1999 to 2001, and surveys 2 (2002\u20132005), 3 (2006\u20132008), and 4 (2009\u20132011) were prospective follow-up surveys.\n\nIn the third survey of the TLGS (2006\u20132008), from among 12523 examined participants, 3462 were randomly selected for dietary assessment; of these, 621 individuals were aged 6\u201318 y who included in the current study. Participants who had incomplete dietary intakes or had missing measures of MetS components (n = 29) were excluded. Additionally, participants with reported energy intakes (kcal/day) to energy requirements ratios beyond \u00b13SD range, were also excluded (n = 6) because they fell into the over-or under-reporting categories. Participants were followed until survey 4 (2009\u20132011), an average of 3.6 years (follow-up rate: 86 %). Depending on the outcome variable under analysis, we excluded persons whose data at survey 3 (2006\u20132008) indicated prevalent MetS (n = 69), hypertension (n = 53), high triglycerides (TGs) (n = 168), low high density lipoprotein cholesterol (HDL-C) (n = 242), high fasting plasma glucose (FPG) (n = 12), or abdominal obesity (n = 145); some individuals fell into more than one exclusion category. The final sample sizes in fully adjusted models varied by outcomes as follows: MetS (n = 424), abdominal obesity (n = 327), high triglycerides (TGs) (n = 347), hypertension (n = 439), low HDL-C (n = 290), and high fasting plasma glucose (n = 476).\n\nThe design of this study was approved by the institutional ethics committee of the Research Institute for Endocrine Sciences, Shahid Beheshti University of Medical Sciences, and written informed consent was obtained from participants\u2019 parents.\n\nDietary intake\nA valid and reliable 168-item semi-quantitative food frequency questionnaire (FFQ) was used by trained dietitians, during face-to face interviews, to evaluate the usual dietary intakes of participants. Participants were asked to report their consumption frequency of food items during the previous year on a daily, weekly, or monthly basis. Mothers were asked about the type and quantity of meals and snacks when children were unable to recall. Portion sizes of consumed foods were specified according to the US Department of Agriculture (USDA) standard portion sizes or household measures (e.g. beans, 1 tablespoon; chicken meat, 1 leg or wing; rice, 1 large or small plate). Obtained data were converted to mean daily intake in grams or milliliters. This FFQ included questions on the frequency of consumption and usual portion size of all kinds of sugar sweetened carbonated soft drinks (SSSDs) and fruit juice drinks (both 100 % fruit juice and sugar sweetened synthetic juice drinks that are not 100 % juice). According to Althuis et al., these items were combined to estimate the daily intake of SSBs (ml/d). SSB intakes were divided by using quartile cutoffs, with the assumption that one serving was equivalent to 1 cup (250 mL).\n\nTo study the reproducibility of the FFQ, 132 subjects (61 men and 71 women) twice completed a 168-item FFQ (FFQ1, FFQ2), with a 14-month interval between FFQ1 and FFQ2, and to assess the validity, 12 dietary recalls (DRs) were collected (1 each month) over the 1-year interval. The age-and energy-adjusted and deattenuated Spearman correlation coefficients used to assess validity of the FFQ for soft drinks were 0.54 and 0.37 in men and women, respectively. The age-and energy-adjusted intraclass correlation coefficients, which reflect the reproducibility of food groups in the FFQ, were 0.61 and 0.65 in men and in women, respectively.\n\nMeasurements\nInformation about physical activity was collected using the modifiable activity questionnaire (MAQ) to calculate metabolic equivalent task (MET) minutes per week. High reliability (98 %) and moderate validity (47 %) were ascertained for the MAQ translated into Persian. Levels of physical activity were categorized as light (MET <600 min/wk), moderate (MET 600\u20131,499 min/wk), and vigorous (MET \u22651,500 min/wk).\n\nWeight was measured, while participants were minimally clothed without shoes, using digital scales (Seca 707, Seca Corp., Hanover, MD; range 0.1\u2013150 kg) and recorded to the nearest 100 g. Height was measured in a standing position without shoes, using a stadiometer, with shoulders in normal alignment. Body mass index (BMI) was calculated as weight (kg) divided by the square of height (m[2]). Waist circumference (WC) was measured at the umbilicus, using a measuring tape, without pressure to body surfaces, and was recorded to the nearest 0.5 cm.\n\nBlood pressure was measured twice, after participants were seated for 15 min, using a standard mercury sphygmomanometer; there was at least a 30-s interval between these two separate measurements, and the mean of two measurements was considered the patient\u2019s blood pressure.\n\nBlood samples, at baseline and follow up, were drawn between 7:00 and 9:00 AM from all study participants after 12\u201314 h overnight fasting; all blood analyses were done at the TLGS research laboratory on the day of blood collection. FPG was measured by the enzymatic colorimetric method using glucose oxidase. Serum HDL-C was measured after precipitation of the apolipoprotein B-containing lipoproteins with phosphotungstic acid and serum TGs were assayed using an enzymatic colorimetric method with glycerol phosphate oxidase. These analyses were performed using commercial kits (Pars Azmoon Inc., Tehran, Iran) and a Selectra 2 auto analyzer (Vital Scientific, Spankeren, The Netherlands). Inter-and intra-assay coefficients of variations at baseline were both 2.2 % for FPG, 2 and 0.5 % for HDL-C and 1.6 and 0.6 % for TGs, respectively.\n\nDefinitions\nIn children and adolescents, MetS was defined according to definition proposed by Cook et al. as 3 or more of the following: Fasting TGs \u2265110 mg/dl; HDL-C < 40 mg/dl; WC \u2265 90th percentile for age and sex, according to national reference curves; systolic blood pressure and diastolic blood pressure \u226590th percentile for sex, age and height, from the National Heart, Lung, and Blood Institute\u2019s recommended cut-off points, and FPG \u2265 100 mg/dl, according to the recommendations of American Diabetes Association. In subjects aged > 18 years after follow-up, MetS was defined according to the joint interim statement as the presence of any 3 of 5 risk factors of the following: 1) Abdominal obesity as WC \u226591 and men \u226589 cm for women and men, respectively, according to Iranians cut-off points; 2) FPG \u2265100 mg/dl or drug treatment; 3) fasting TGs \u2265150 mg/dl or drug treatment; 4) fasting HDL-C <50 mg/dl for women and <40 mg/dl for men or drug treatment, and 5) hypertension was defined as systolic blood pressure \u2265130 mmHg, diastolic blood pressure \u226585 mmHg, or antihypertensive drug treatment.\n\nStatistical analysis\nAll analyses were conducted using the Statistical Package for Social Sciences (version 15.0; SPSS Inc, Chicago IL). Baseline characteristics of subjects were expressed as mean \u00b1 SD or median (interquartile 25\u201375) for continuous variables and percentages for categorical variables. To investigate the trend of continuous and categorical variables according to quartiles of SSB, linear regression and chi-square test were used, respectively. There were no significant differences in baseline anthropometric and biochemical assessments between subjects who provided follow-up assessments and those lost to follow-up (data not shown).\n\nIntakes of SSBs (<19.2, 19.2-43.8, 43.9-91.5, and >91.6 ml/day), SSSDs (<7.6, 7.6\u201315.3, 15.3\u201340.0, and >40.0 ml/day), and fruit juice drinks (<4.1, 4.1\u201313.5, 13.5\u201334.9, and 34.9 ml/day) were categorized into the quartile cut off points. To estimate the risk of 3.6-year incident outcomes, separate multiple regression models were used for abdominal obesity, high FPG, high TGs, low HDL-C, hypertension, and MetS with quartiles of SSB, SSSD, and fruit juice drink consumption; odds ratios (OR) and 95 % confidence interval (CI) were reported. The initial model was adjusted for age, sex, total energy intake, physical activity, and family history of diabetes. Further adjustment for dietary fiber, tea and coffee, red and processed meat, fruits, and vegetables was made and the final model was additionally adjusted for BMI.",
        "results": "**Results**\n\nThe mean \u00b1 SD age of study population (68 % girls) was 13.6 \u00b1 3.7 years; the mean daily SSB intake was 98 ml in boys and 70 ml in girls. After a 3.6 year follow up, cumulative incidence of MetS was 11 %.\n\nThe descriptive characteristics of the study population according to quartile of SSB consumption are shown in Table 1. Participants with higher SSB consumption were likely to have higher systolic blood pressure, FPG, and BMI; in addition, subjects with higher SSB consumption had higher intakes of daily energy, saturated fatty acids, trans-fatty acids, glucose, fructose, sucrose, simple sugars, higher glycemic index, and lower intakes of fruits and whole grains.\n\nSugar sweetened beverage intake was inversely related to MetS and its components after adjustment for potential confounders (Table 2). Further adjustment for dietary factors did not have a major impact on the results. In this model, subjects in the highest quartile of SSB consumption had increased risk of incident MetS (OR: 3.16; 95 % CI: 1.02\u20139.78, P for trend = 0.007) and abdominal obesity (OR: 2.97; 95 % CI: 1.23\u20137.19, P for trend = 0.017). Adjustment for BMI slightly strengthened the association; adjusted multivariable ORs of MetS, abdominal obesity, and hypertension were 3.50, 3.66, and 2.90 (all P for trends < 0.05).\n\nHigher intakes of SSSD was associated with significant increase in the risk of incident MetS (OR: 3.58, 95 % CI: 1.09-11.78, P for trend = 0.013), abdominal obesity (OR: 3.78, 95 % CI: 1.08\u201313.27, P for trend = 0.043), and hypertension (OR: 2.74, 95 % CI: 1.05\u20137.19, P for trend = 0.018) in the fully adjusted model. Associations for fruit juice drink intake were not statistically significant (Table 3).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Association between yogurt consumption and the risk of metabolic syndrome over 6 years in the SUN study.",
        "abstract": "## BACKGROUND\nThe role of yogurt consumption in the development of metabolic syndrome (MetS) is not fully understood and the available epidemiologic evidence is scarce. The aim of our study was to assess the association between total, whole-fat, or low-fat yogurt consumption and the risk of developing MetS.\n## METHODS\nYogurt consumption was assessed at baseline through a 136-item validated FFQ. MetS was defined following the harmonized definition for MetS according to the AHA and the IDF criteria. Logistic regression models were used.\n## RESULTS\nDuring the first 6-y of follow-up of the SUN cohort, 306 incident cases of MetS were identified. Frequent consumption [\u2265875\u00a0g/week (\u22657 servings/week) versus\u2009\u2264\u2009250\u00a0g/week (2 servings/week)] of total, whole-fat and low-fat yogurt consumption showed non-significant inverse associations with MetS [OR\u2009=\u20090.84 (95% CI: 0.60-1.18); 0.98 (95% CI: 0.68-1.41); and 0.63 (95% CI: 0.39-1.02) respectively]. Only one component of the MetS, central adiposity, was inversely associated with total and whole-fat yogurt consumption [OR\u2009=\u20090.85 (95% CI: 0.74-0.98) and 0.85 (95% CI: 0.73-0.99) respectively]. In the joint assessment of exposure to total yogurt consumption and fruit consumption, those in the highest category of total yogurt consumption, and having a high fruit consumption (above the median \u2265264.5\u00a0g/day) exhibited a significantly lower risk of developing MetS [OR\u2009=\u20090.61 (95% CI: 0.38-0.99)] compared with those in the lowest category of total yogurt consumption and had fruit consumption below the study median.\n## CONCLUSION\nNo significant association between yogurt consumption and MetS was apparent. Only one component out of the 5 MetS criteria, central adiposity, was inversely associated with high yogurt consumption. The combination of high consumption of both yogurt and fruit was inversely associated with the development of MetS.",
        "methods": "## Methods\n\nStudy population  \nThe SUN project is a prospective, multipurpose, and dynamic cohort study conducted in Spain, which was originally designed to establish associations between diet and the occurrence of several diseases and chronic conditions including overweight and obesity [16]. Using biennial mailed questionnaires participants have been continually followed-up. Participants\u2019 recruitment started in December 1999, and it is permanently open. All participants are university graduates, with ages ranging from 20 to 90 years.\n\nFor the present study, we included a subsample of the cohort with a minimum follow-up period of 6 years, therefore only those participants who were recruited before September 2006 could be included (n = 15,909). To avoid reverse causality bias, we excluded 4,084 participants who met at least one MetS criterion at baseline. As recommended in nutritional epidemiology, we also excluded 1,187 participants who reported values for total energy intake at baseline out of predefined limits (less than 800 kcal/d in men and 500 kcal/d in women or more than 4,000 kcal/d in men and 3,500 kcal/d in women). Further exclusions were 578 participants without any follow-up who were considered lost to follow-up (retention rate = 95%), and 1,997 participants who did not provide the relevant information about diagnostic criteria for the MetS at the 6th-year follow-up. After these exclusions a total of 8,063 participants were available for the final analysis. The retention rate of the study was above 95%. A comparison analysis between baseline characteristics of the participants who were included in the study, and those participants who did not have any information at Q_6, did not find meaningful differences between them (data not shown).\n\nThe study was approved by the Human Research Ethical Committee of the University of Navarra. Voluntary completion of the first self-administrated questionnaire was considered to imply informed consent.\n\nDietary assessment  \nA semi-quantitative food frequency questionnaire (FFQ) was included in the baseline questionnaire. The FFQ was repeatedly validated in Spain [17-19] and includes 136 food items. Nutrient composition of specified portion sizes was multiplied by the frequency of consumption in order to calculate each nutrient score. Frequencies of consumption were measured in 9 categories (ranging from never/almost never, to >6 servings/day) for each food item, in order to assess habitual dietary intakes over the previous year. We defined a Mediterranean dietary pattern \u201ca priori\u201d using the score proposed by Trichopoulou et al. [20]. The traditional Mediterranean dietary pattern is characterized by abundant consumption of olive oil (the major source of fat), plant foods (vegetables, fruits, cereals, and nuts), fresh fruit as daily dessert, low-to moderate intake of dairy products (cheese and yogurt), fish and poultry, low intake of red meat, and regular moderate intake of wine generally consumed during meals. For the present study the Mediterranean dietary pattern was appraised combining 8 out of 9 of items of the original score (fruits and nuts; vegetables; fish; legumes; cereals; meat and meat products; alcohol; and the ratio monounsaturated fatty acids/saturated fatty acids). Since yogurt consumption is our exposure of interest, we excluded consumption of dairy products from the score. Therefore, the score ranged from 0 to 8 points. Food composition tables valid for Spain were used to update the nutrient databank [21].\n\nYogurt consumption assessment  \nFrequency of whole-fat and low-fat yogurt consumption in the previous year was reported by each participant in the baseline FFQ. Frequency of total yogurt consumption was estimated using the sum of these two food items (whole-fat and low-fat yogurt consumption). The sample was divided into three categories according to their total, whole-fat and low-fat yogurt consumption: 0\u2013250 g/week (0\u20132 servings/week), >250 to <875 g/week, (>2 to <7 servings/week) and \u2265875 g/week (\u22657 servings/week); (125 g was considered as one serving size).\n\nAssessment of non-dietary variables  \nOther questions (46 items for men and 54 for women) were also included in the baseline questionnaire assessing the participants\u2019 lifestyle, and socio-demographic variables (sex, age, marital status, and employment), health related habits (smoking status, physical activity during leisure time), medical history (prevalence of chronic diseases such as cancer, diabetes and cardiovascular disease), as well as anthropometric data (weight and height) previously validated in the cohort [22]. We calculated body mass index (BMI) as self-reported weight in kilograms divided by the square of the self-reported height in meters. Physical activity was ascertained through a baseline 17-item questionnaire Physical activity was previously validated in a subsample of the cohort, with weekly MET-hours shown to adequately correlate with objectively measured energy expenditure [Spearman \u03c1 = 0.51; 95% confidence interval (CI): 0.232, 0.707] [23].\n\nOutcome assessment  \nWe defined MetS according to the American Heart Association (AHA) and the International Diabetes Federation (IDF) criteria as outlined in the harmonized definition for MetS [4]. According to this harmonized definition, the MetS diagnosis requires the presence of at least 3 of the 5 criteria: central adiposity (in our study waist circumference: \u226594 cm in males and \u226580 cm in females), hypertriglyceridemia (\u2265150 mg/dL or condition-specific medication), low levels of high-density lipoprotein cholesterol (<40 mg/dL for men and < 50 mg/dL for women), elevated blood pressure (systolic \u2265 130 and/or diastolic \u2265 85 mmHg or antihypertensive drug treatment in a patient with a history of hypertension), and impaired glucose metabolism (fasting glucose \u2265 100 mg/dL or drug treatment of elevated glucose). For a sensitivity analysis we used a higher cut-off point for central adiposity (waist circumference: \u2265102 cm in males and \u226588 cm in females) as proposed by the WHO for Caucasian populations [4].\n\nSelf-reported information about each specific MetS criterion was collected in the Q_6 (6th year follow-up questionnaire). With the Q_6 questionnaire, a measuring tape and an explanation of how to measure their own waist were sent to each participant. Diagnosis of MetS itself, the same as each MetS criterion was previously validated in a subsample of the cohort [24,25]. The validation study for MetS diagnosis found agreement between self-reported MetS diagnosis and MetS diagnosis from participant medical records to be 91.2% for confirmed MetS (95% CI: 80.7-97.1%) and 92.2% for non-confirmed MetS (95% CI: 85.1-96.4%) [24].\n\nIncident cases of MetS were defined as all participants who did not have any MetS criterion at baseline, and reported 3 or more criteria of MetS in the 6th year follow-up questionnaire.\n\nStatistical analyses  \nThree categories according to their baseline yogurt consumption were used to classify the participants included in the study: 0\u2013250 g/week (0\u20132 servings/week), >250 to <875 g/week (>2 to <7 servings/week), and \u2265875 g/week (\u22657 servings/week). The group with the lowest frequency of consumption (0\u2013250 g/week) was considered as the reference category.\n\nLogistic regression models were fitted to assess the relationship between total, whole-fat or low-fat yogurt consumption at baseline and the risk of developing incident MetS during the first 6 years of follow-up. The specific relationship between each of the 5 components of the MetS and total, whole-fat and low-fat yogurt consumption were also assessed. Odds Ratios (OR) and their 95% CI were calculated. Tests for linear trend across increasing categories of yogurt consumption were conducted by assigning the median consumption within each category and treating this variable as a continuous variable.\n\nThe interaction between adherence to Mediterranean diet [low (0\u20134 pts)/high (5\u20138 pts)], and fruit consumption [low (under the median: <264.5 g/d)/high (above the median: \u2265264.5 g/d)] and total yogurt consumption were tested using likelihood ratio tests comparing the fully adjusted model and the same model with the interaction product-term (2 degrees of freedom).\n\nAdditional analyses were conducted to test the association of the joint exposure to total yogurt consumption and fruit consumption with the risk of MetS, therefore we built six categories of the joint combined exposure to total yogurt consumption (low, medium and high consumption) and fruit consumption (under the median and at the median or higher) considering as the reference category those who had low yogurt consumption (<250 g/week) and low fruit consumption (under the median).\n\nWe fitted a first model without any adjustment (crude), a second model adjusted for age and sex, and a third multivariable-adjusted model after additional adjustment for the following potential confounders: baseline weight (kg), total energy intake (kcal/d), alcohol (g/d), soft drink consumption (ml/d), red meat consumption (g/d), consumption of French fries (g/d), consumption of fast food (g/d), Mediterranean dietary pattern adherence (3 categories: low, medium, and high adherence), physical activity (METs-hours/week), sitting hours (hr/d), sedentary behavior (hr/d), smoking status (non-smoker, smoker, former smoker), snacking between meals (yes/no), and following a special diet (yes/no). Snacking between meals was defined as those participants who responded affirmatively to the question \u201cDo you eat between main meals (snacking)?\u201d We did not consider if it was healthy or unhealthy snacking. The percentage of participants who reported snacking between meals was 33.3%. We considered that a participant followed a special diet, if she/he answered yes to the question \u201cDo you follow any special diet\u201d. If they had answered yes, they were asked which kind of special diet they followed (such as hypocaloric diet, vegetarian diet, low-fat diet, low sodium diet, high protein diet, among others). We used this variable dichotomously. The percentage of participants following a special diet was 6.1%.\n\nFor the selection of the potential confounders in the multivariate model, and as its currently recommended [26], we took into account the previously published scientific literature including our own results based on the cohort about risk factors for hypertension, overweight/obesity, type 2 diabetes, avoiding exclusively the statistical approach, the stepwise procedures, or the changes in the point estimates after adjusting for potential confounders.\n\nAll p values presented are two-tailed; p < 0.05 was considered statistically significant. Analyses were performed using STATA/SE version 12.0 (StataCorp, College Station, TX, USA).",
        "results": "## Results\nThe main characteristics of participants according to their frequency of yogurt consumption are presented in Table 1. The mean age of the participants was 36.4 y (SD: 11.6) and the mean BMI was 22.7 kg/m[2] (SD: 2.7).\n\nOn average participants in the highest category of total yogurt consumption had the lowest BMI, were more physically active, less likely to be current smokers, spent less hours sitting, and had the highest total energy intake, carbohydrate and protein intake and the lowest fat intake.\n\nThe association between the frequency of yogurt consumption and the risk of incident MetS was assessed in 8,063 participants initially free of MetS. During the first 6 years of follow up we observed 306 incident cases of MetS. As displayed in Table 2, a higher frequency of total, whole-fat or low-fat yogurt consumption was not significantly associated with a lower risk of developing MetS. Participants consuming total, whole-fat and low-fat \u2265875 g/week did not exhibit significantly lower risk of developing MetS compared with those who consumed from 0 to 250 g/week. However, we observed non-significant inverse associations [OR = 0.84 (95% CI: 0.60-1.18); 0.98 (95% CI: 0.68-1.41); and 0.63 (95% CI: 0.39-1.02)] respectively for total, whole-fat and low-fat yogurt consumption, after adjusting for potential confounders (Table 2).\n\nThe risk of developing each component of MetS according to the total, whole-fat and low-fat yogurt consumption is shown in Figure 1. This figure shows the ORs for each component of the MetS separately for total, whole-fat and low-fat yogurt consumption only for the highest category of consumption (\u2265875 g/week) versus the lowest category (\u2264250 g/week). From the 5 components of MetS, we observed that those participants who consumed \u2265875 g/week of total and whole-fat yogurt compared with those who consumed \u2264 250 g/week had a lower risk of developing central adiposity [OR = 0.85 (95% CI: 0.74-0.98) and 0.85 (95% CI: 0.73-0.99) respectively, after adjusting for potential confounders. The results for the rest of the components suggested a non-significant trend for a lower risk (except for blood pressure).\n\nAs a sensitivity analysis we changed the cut-off point for central adiposity (waist circumference: \u2265102 cm in males and \u226588 cm in females) as proposed by the WHO for Caucasian populations. The ORs for total, whole-fat and low-fat yogurt consumption (\u2265875 g/week versus \u2264250 g/week) were 0.74 (95% CI: 0.48-1.14); 0.76 (95% CI: 0.46-1.25); and 0.83 (95% CI: 0.47-1.45), respectively (data not shown).\n\nNo interaction between Mediterranean diet adherence (low/high) and total yogurt consumption was observed (p for interaction = 0.270). The interaction between fruit consumption (below/above median) and total yogurt consumption approached the threshold for statistical significance (p for interaction = 0.070). In any case, the lowest ORs in Table 3 were observed in the groups of participants with both a higher adherence to the Mediterranean diet and higher consumption of total yogurt [OR = 0.77 (95% CI: 0.46-1.30) p for trend = 0.323] and in participants with both above-the-median fruit consumption and high yogurt consumption [OR = 0.69 (95% CI: 0.43-1.10) p for trend = 0.104] (Table 3).\n\nWhen we simultaneously analyzed total yogurt consumption (3 categories) and fruit consumption (2 categories), 6 new categories were created. The category of lower yogurt consumption (\u2264250 g/week) and low fruit consumption (under the median < 264.5 g/d) was considered as the reference group. We observed a significantly lower risk of developing MetS in the group where a higher yogurt consumption (\u2265875 g/week) was accompanied by a high fruit consumption (above the median \u2265264.5 g/d) OR = 0.61 (95% CI: 0.38-0.99) (Table 4).",
        "column_title": "Selection (Maximum 4 points awarded for cohort representativeness, selection of non-exposed cohort, exposure assessment, and demonstration outcome not present at baseline)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3, 4",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Association between yogurt consumption and the risk of metabolic syndrome over 6 years in the SUN study.",
        "abstract": "## BACKGROUND\nThe role of yogurt consumption in the development of metabolic syndrome (MetS) is not fully understood and the available epidemiologic evidence is scarce. The aim of our study was to assess the association between total, whole-fat, or low-fat yogurt consumption and the risk of developing MetS.\n## METHODS\nYogurt consumption was assessed at baseline through a 136-item validated FFQ. MetS was defined following the harmonized definition for MetS according to the AHA and the IDF criteria. Logistic regression models were used.\n## RESULTS\nDuring the first 6-y of follow-up of the SUN cohort, 306 incident cases of MetS were identified. Frequent consumption [\u2265875\u00a0g/week (\u22657 servings/week) versus\u2009\u2264\u2009250\u00a0g/week (2 servings/week)] of total, whole-fat and low-fat yogurt consumption showed non-significant inverse associations with MetS [OR\u2009=\u20090.84 (95% CI: 0.60-1.18); 0.98 (95% CI: 0.68-1.41); and 0.63 (95% CI: 0.39-1.02) respectively]. Only one component of the MetS, central adiposity, was inversely associated with total and whole-fat yogurt consumption [OR\u2009=\u20090.85 (95% CI: 0.74-0.98) and 0.85 (95% CI: 0.73-0.99) respectively]. In the joint assessment of exposure to total yogurt consumption and fruit consumption, those in the highest category of total yogurt consumption, and having a high fruit consumption (above the median \u2265264.5\u00a0g/day) exhibited a significantly lower risk of developing MetS [OR\u2009=\u20090.61 (95% CI: 0.38-0.99)] compared with those in the lowest category of total yogurt consumption and had fruit consumption below the study median.\n## CONCLUSION\nNo significant association between yogurt consumption and MetS was apparent. Only one component out of the 5 MetS criteria, central adiposity, was inversely associated with high yogurt consumption. The combination of high consumption of both yogurt and fruit was inversely associated with the development of MetS.",
        "methods": "## Methods\n\nStudy population  \nThe SUN project is a prospective, multipurpose, and dynamic cohort study conducted in Spain, which was originally designed to establish associations between diet and the occurrence of several diseases and chronic conditions including overweight and obesity [16]. Using biennial mailed questionnaires participants have been continually followed-up. Participants\u2019 recruitment started in December 1999, and it is permanently open. All participants are university graduates, with ages ranging from 20 to 90 years.\n\nFor the present study, we included a subsample of the cohort with a minimum follow-up period of 6 years, therefore only those participants who were recruited before September 2006 could be included (n = 15,909). To avoid reverse causality bias, we excluded 4,084 participants who met at least one MetS criterion at baseline. As recommended in nutritional epidemiology, we also excluded 1,187 participants who reported values for total energy intake at baseline out of predefined limits (less than 800 kcal/d in men and 500 kcal/d in women or more than 4,000 kcal/d in men and 3,500 kcal/d in women). Further exclusions were 578 participants without any follow-up who were considered lost to follow-up (retention rate = 95%), and 1,997 participants who did not provide the relevant information about diagnostic criteria for the MetS at the 6th-year follow-up. After these exclusions a total of 8,063 participants were available for the final analysis. The retention rate of the study was above 95%. A comparison analysis between baseline characteristics of the participants who were included in the study, and those participants who did not have any information at Q_6, did not find meaningful differences between them (data not shown).\n\nThe study was approved by the Human Research Ethical Committee of the University of Navarra. Voluntary completion of the first self-administrated questionnaire was considered to imply informed consent.\n\nDietary assessment  \nA semi-quantitative food frequency questionnaire (FFQ) was included in the baseline questionnaire. The FFQ was repeatedly validated in Spain [17-19] and includes 136 food items. Nutrient composition of specified portion sizes was multiplied by the frequency of consumption in order to calculate each nutrient score. Frequencies of consumption were measured in 9 categories (ranging from never/almost never, to >6 servings/day) for each food item, in order to assess habitual dietary intakes over the previous year. We defined a Mediterranean dietary pattern \u201ca priori\u201d using the score proposed by Trichopoulou et al. [20]. The traditional Mediterranean dietary pattern is characterized by abundant consumption of olive oil (the major source of fat), plant foods (vegetables, fruits, cereals, and nuts), fresh fruit as daily dessert, low-to moderate intake of dairy products (cheese and yogurt), fish and poultry, low intake of red meat, and regular moderate intake of wine generally consumed during meals. For the present study the Mediterranean dietary pattern was appraised combining 8 out of 9 of items of the original score (fruits and nuts; vegetables; fish; legumes; cereals; meat and meat products; alcohol; and the ratio monounsaturated fatty acids/saturated fatty acids). Since yogurt consumption is our exposure of interest, we excluded consumption of dairy products from the score. Therefore, the score ranged from 0 to 8 points. Food composition tables valid for Spain were used to update the nutrient databank [21].\n\nYogurt consumption assessment  \nFrequency of whole-fat and low-fat yogurt consumption in the previous year was reported by each participant in the baseline FFQ. Frequency of total yogurt consumption was estimated using the sum of these two food items (whole-fat and low-fat yogurt consumption). The sample was divided into three categories according to their total, whole-fat and low-fat yogurt consumption: 0\u2013250 g/week (0\u20132 servings/week), >250 to <875 g/week, (>2 to <7 servings/week) and \u2265875 g/week (\u22657 servings/week); (125 g was considered as one serving size).\n\nAssessment of non-dietary variables  \nOther questions (46 items for men and 54 for women) were also included in the baseline questionnaire assessing the participants\u2019 lifestyle, and socio-demographic variables (sex, age, marital status, and employment), health related habits (smoking status, physical activity during leisure time), medical history (prevalence of chronic diseases such as cancer, diabetes and cardiovascular disease), as well as anthropometric data (weight and height) previously validated in the cohort [22]. We calculated body mass index (BMI) as self-reported weight in kilograms divided by the square of the self-reported height in meters. Physical activity was ascertained through a baseline 17-item questionnaire Physical activity was previously validated in a subsample of the cohort, with weekly MET-hours shown to adequately correlate with objectively measured energy expenditure [Spearman \u03c1 = 0.51; 95% confidence interval (CI): 0.232, 0.707] [23].\n\nOutcome assessment  \nWe defined MetS according to the American Heart Association (AHA) and the International Diabetes Federation (IDF) criteria as outlined in the harmonized definition for MetS [4]. According to this harmonized definition, the MetS diagnosis requires the presence of at least 3 of the 5 criteria: central adiposity (in our study waist circumference: \u226594 cm in males and \u226580 cm in females), hypertriglyceridemia (\u2265150 mg/dL or condition-specific medication), low levels of high-density lipoprotein cholesterol (<40 mg/dL for men and < 50 mg/dL for women), elevated blood pressure (systolic \u2265 130 and/or diastolic \u2265 85 mmHg or antihypertensive drug treatment in a patient with a history of hypertension), and impaired glucose metabolism (fasting glucose \u2265 100 mg/dL or drug treatment of elevated glucose). For a sensitivity analysis we used a higher cut-off point for central adiposity (waist circumference: \u2265102 cm in males and \u226588 cm in females) as proposed by the WHO for Caucasian populations [4].\n\nSelf-reported information about each specific MetS criterion was collected in the Q_6 (6th year follow-up questionnaire). With the Q_6 questionnaire, a measuring tape and an explanation of how to measure their own waist were sent to each participant. Diagnosis of MetS itself, the same as each MetS criterion was previously validated in a subsample of the cohort [24,25]. The validation study for MetS diagnosis found agreement between self-reported MetS diagnosis and MetS diagnosis from participant medical records to be 91.2% for confirmed MetS (95% CI: 80.7-97.1%) and 92.2% for non-confirmed MetS (95% CI: 85.1-96.4%) [24].\n\nIncident cases of MetS were defined as all participants who did not have any MetS criterion at baseline, and reported 3 or more criteria of MetS in the 6th year follow-up questionnaire.\n\nStatistical analyses  \nThree categories according to their baseline yogurt consumption were used to classify the participants included in the study: 0\u2013250 g/week (0\u20132 servings/week), >250 to <875 g/week (>2 to <7 servings/week), and \u2265875 g/week (\u22657 servings/week). The group with the lowest frequency of consumption (0\u2013250 g/week) was considered as the reference category.\n\nLogistic regression models were fitted to assess the relationship between total, whole-fat or low-fat yogurt consumption at baseline and the risk of developing incident MetS during the first 6 years of follow-up. The specific relationship between each of the 5 components of the MetS and total, whole-fat and low-fat yogurt consumption were also assessed. Odds Ratios (OR) and their 95% CI were calculated. Tests for linear trend across increasing categories of yogurt consumption were conducted by assigning the median consumption within each category and treating this variable as a continuous variable.\n\nThe interaction between adherence to Mediterranean diet [low (0\u20134 pts)/high (5\u20138 pts)], and fruit consumption [low (under the median: <264.5 g/d)/high (above the median: \u2265264.5 g/d)] and total yogurt consumption were tested using likelihood ratio tests comparing the fully adjusted model and the same model with the interaction product-term (2 degrees of freedom).\n\nAdditional analyses were conducted to test the association of the joint exposure to total yogurt consumption and fruit consumption with the risk of MetS, therefore we built six categories of the joint combined exposure to total yogurt consumption (low, medium and high consumption) and fruit consumption (under the median and at the median or higher) considering as the reference category those who had low yogurt consumption (<250 g/week) and low fruit consumption (under the median).\n\nWe fitted a first model without any adjustment (crude), a second model adjusted for age and sex, and a third multivariable-adjusted model after additional adjustment for the following potential confounders: baseline weight (kg), total energy intake (kcal/d), alcohol (g/d), soft drink consumption (ml/d), red meat consumption (g/d), consumption of French fries (g/d), consumption of fast food (g/d), Mediterranean dietary pattern adherence (3 categories: low, medium, and high adherence), physical activity (METs-hours/week), sitting hours (hr/d), sedentary behavior (hr/d), smoking status (non-smoker, smoker, former smoker), snacking between meals (yes/no), and following a special diet (yes/no). Snacking between meals was defined as those participants who responded affirmatively to the question \u201cDo you eat between main meals (snacking)?\u201d We did not consider if it was healthy or unhealthy snacking. The percentage of participants who reported snacking between meals was 33.3%. We considered that a participant followed a special diet, if she/he answered yes to the question \u201cDo you follow any special diet\u201d. If they had answered yes, they were asked which kind of special diet they followed (such as hypocaloric diet, vegetarian diet, low-fat diet, low sodium diet, high protein diet, among others). We used this variable dichotomously. The percentage of participants following a special diet was 6.1%.\n\nFor the selection of the potential confounders in the multivariate model, and as its currently recommended [26], we took into account the previously published scientific literature including our own results based on the cohort about risk factors for hypertension, overweight/obesity, type 2 diabetes, avoiding exclusively the statistical approach, the stepwise procedures, or the changes in the point estimates after adjusting for potential confounders.\n\nAll p values presented are two-tailed; p < 0.05 was considered statistically significant. Analyses were performed using STATA/SE version 12.0 (StataCorp, College Station, TX, USA).",
        "results": "## Results\nThe main characteristics of participants according to their frequency of yogurt consumption are presented in Table 1. The mean age of the participants was 36.4 y (SD: 11.6) and the mean BMI was 22.7 kg/m[2] (SD: 2.7).\n\nOn average participants in the highest category of total yogurt consumption had the lowest BMI, were more physically active, less likely to be current smokers, spent less hours sitting, and had the highest total energy intake, carbohydrate and protein intake and the lowest fat intake.\n\nThe association between the frequency of yogurt consumption and the risk of incident MetS was assessed in 8,063 participants initially free of MetS. During the first 6 years of follow up we observed 306 incident cases of MetS. As displayed in Table 2, a higher frequency of total, whole-fat or low-fat yogurt consumption was not significantly associated with a lower risk of developing MetS. Participants consuming total, whole-fat and low-fat \u2265875 g/week did not exhibit significantly lower risk of developing MetS compared with those who consumed from 0 to 250 g/week. However, we observed non-significant inverse associations [OR = 0.84 (95% CI: 0.60-1.18); 0.98 (95% CI: 0.68-1.41); and 0.63 (95% CI: 0.39-1.02)] respectively for total, whole-fat and low-fat yogurt consumption, after adjusting for potential confounders (Table 2).\n\nThe risk of developing each component of MetS according to the total, whole-fat and low-fat yogurt consumption is shown in Figure 1. This figure shows the ORs for each component of the MetS separately for total, whole-fat and low-fat yogurt consumption only for the highest category of consumption (\u2265875 g/week) versus the lowest category (\u2264250 g/week). From the 5 components of MetS, we observed that those participants who consumed \u2265875 g/week of total and whole-fat yogurt compared with those who consumed \u2264 250 g/week had a lower risk of developing central adiposity [OR = 0.85 (95% CI: 0.74-0.98) and 0.85 (95% CI: 0.73-0.99) respectively, after adjusting for potential confounders. The results for the rest of the components suggested a non-significant trend for a lower risk (except for blood pressure).\n\nAs a sensitivity analysis we changed the cut-off point for central adiposity (waist circumference: \u2265102 cm in males and \u226588 cm in females) as proposed by the WHO for Caucasian populations. The ORs for total, whole-fat and low-fat yogurt consumption (\u2265875 g/week versus \u2264250 g/week) were 0.74 (95% CI: 0.48-1.14); 0.76 (95% CI: 0.46-1.25); and 0.83 (95% CI: 0.47-1.45), respectively (data not shown).\n\nNo interaction between Mediterranean diet adherence (low/high) and total yogurt consumption was observed (p for interaction = 0.270). The interaction between fruit consumption (below/above median) and total yogurt consumption approached the threshold for statistical significance (p for interaction = 0.070). In any case, the lowest ORs in Table 3 were observed in the groups of participants with both a higher adherence to the Mediterranean diet and higher consumption of total yogurt [OR = 0.77 (95% CI: 0.46-1.30) p for trend = 0.323] and in participants with both above-the-median fruit consumption and high yogurt consumption [OR = 0.69 (95% CI: 0.43-1.10) p for trend = 0.104] (Table 3).\n\nWhen we simultaneously analyzed total yogurt consumption (3 categories) and fruit consumption (2 categories), 6 new categories were created. The category of lower yogurt consumption (\u2264250 g/week) and low fruit consumption (under the median < 264.5 g/d) was considered as the reference group. We observed a significantly lower risk of developing MetS in the group where a higher yogurt consumption (\u2265875 g/week) was accompanied by a high fruit consumption (above the median \u2265264.5 g/d) OR = 0.61 (95% CI: 0.38-0.99) (Table 4).",
        "column_title": "Outcome (Maximum 3 points awarded for follow-up length, adequacy of follow-up, and outcome assessment)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2, 3",
        "type": "score",
        "truth": "3"
    },
    {
        "pmid": "32644139",
        "meta_analysis_title": "Association of Major Food Sources of Fructose-Containing Sugars With Incident Metabolic Syndrome: A Systematic Review and Meta-analysis.",
        "title": "Association between yogurt consumption and the risk of metabolic syndrome over 6 years in the SUN study.",
        "abstract": "## BACKGROUND\nThe role of yogurt consumption in the development of metabolic syndrome (MetS) is not fully understood and the available epidemiologic evidence is scarce. The aim of our study was to assess the association between total, whole-fat, or low-fat yogurt consumption and the risk of developing MetS.\n## METHODS\nYogurt consumption was assessed at baseline through a 136-item validated FFQ. MetS was defined following the harmonized definition for MetS according to the AHA and the IDF criteria. Logistic regression models were used.\n## RESULTS\nDuring the first 6-y of follow-up of the SUN cohort, 306 incident cases of MetS were identified. Frequent consumption [\u2265875\u00a0g/week (\u22657 servings/week) versus\u2009\u2264\u2009250\u00a0g/week (2 servings/week)] of total, whole-fat and low-fat yogurt consumption showed non-significant inverse associations with MetS [OR\u2009=\u20090.84 (95% CI: 0.60-1.18); 0.98 (95% CI: 0.68-1.41); and 0.63 (95% CI: 0.39-1.02) respectively]. Only one component of the MetS, central adiposity, was inversely associated with total and whole-fat yogurt consumption [OR\u2009=\u20090.85 (95% CI: 0.74-0.98) and 0.85 (95% CI: 0.73-0.99) respectively]. In the joint assessment of exposure to total yogurt consumption and fruit consumption, those in the highest category of total yogurt consumption, and having a high fruit consumption (above the median \u2265264.5\u00a0g/day) exhibited a significantly lower risk of developing MetS [OR\u2009=\u20090.61 (95% CI: 0.38-0.99)] compared with those in the lowest category of total yogurt consumption and had fruit consumption below the study median.\n## CONCLUSION\nNo significant association between yogurt consumption and MetS was apparent. Only one component out of the 5 MetS criteria, central adiposity, was inversely associated with high yogurt consumption. The combination of high consumption of both yogurt and fruit was inversely associated with the development of MetS.",
        "methods": "## Methods\n\nStudy population  \nThe SUN project is a prospective, multipurpose, and dynamic cohort study conducted in Spain, which was originally designed to establish associations between diet and the occurrence of several diseases and chronic conditions including overweight and obesity [16]. Using biennial mailed questionnaires participants have been continually followed-up. Participants\u2019 recruitment started in December 1999, and it is permanently open. All participants are university graduates, with ages ranging from 20 to 90 years.\n\nFor the present study, we included a subsample of the cohort with a minimum follow-up period of 6 years, therefore only those participants who were recruited before September 2006 could be included (n = 15,909). To avoid reverse causality bias, we excluded 4,084 participants who met at least one MetS criterion at baseline. As recommended in nutritional epidemiology, we also excluded 1,187 participants who reported values for total energy intake at baseline out of predefined limits (less than 800 kcal/d in men and 500 kcal/d in women or more than 4,000 kcal/d in men and 3,500 kcal/d in women). Further exclusions were 578 participants without any follow-up who were considered lost to follow-up (retention rate = 95%), and 1,997 participants who did not provide the relevant information about diagnostic criteria for the MetS at the 6th-year follow-up. After these exclusions a total of 8,063 participants were available for the final analysis. The retention rate of the study was above 95%. A comparison analysis between baseline characteristics of the participants who were included in the study, and those participants who did not have any information at Q_6, did not find meaningful differences between them (data not shown).\n\nThe study was approved by the Human Research Ethical Committee of the University of Navarra. Voluntary completion of the first self-administrated questionnaire was considered to imply informed consent.\n\nDietary assessment  \nA semi-quantitative food frequency questionnaire (FFQ) was included in the baseline questionnaire. The FFQ was repeatedly validated in Spain [17-19] and includes 136 food items. Nutrient composition of specified portion sizes was multiplied by the frequency of consumption in order to calculate each nutrient score. Frequencies of consumption were measured in 9 categories (ranging from never/almost never, to >6 servings/day) for each food item, in order to assess habitual dietary intakes over the previous year. We defined a Mediterranean dietary pattern \u201ca priori\u201d using the score proposed by Trichopoulou et al. [20]. The traditional Mediterranean dietary pattern is characterized by abundant consumption of olive oil (the major source of fat), plant foods (vegetables, fruits, cereals, and nuts), fresh fruit as daily dessert, low-to moderate intake of dairy products (cheese and yogurt), fish and poultry, low intake of red meat, and regular moderate intake of wine generally consumed during meals. For the present study the Mediterranean dietary pattern was appraised combining 8 out of 9 of items of the original score (fruits and nuts; vegetables; fish; legumes; cereals; meat and meat products; alcohol; and the ratio monounsaturated fatty acids/saturated fatty acids). Since yogurt consumption is our exposure of interest, we excluded consumption of dairy products from the score. Therefore, the score ranged from 0 to 8 points. Food composition tables valid for Spain were used to update the nutrient databank [21].\n\nYogurt consumption assessment  \nFrequency of whole-fat and low-fat yogurt consumption in the previous year was reported by each participant in the baseline FFQ. Frequency of total yogurt consumption was estimated using the sum of these two food items (whole-fat and low-fat yogurt consumption). The sample was divided into three categories according to their total, whole-fat and low-fat yogurt consumption: 0\u2013250 g/week (0\u20132 servings/week), >250 to <875 g/week, (>2 to <7 servings/week) and \u2265875 g/week (\u22657 servings/week); (125 g was considered as one serving size).\n\nAssessment of non-dietary variables  \nOther questions (46 items for men and 54 for women) were also included in the baseline questionnaire assessing the participants\u2019 lifestyle, and socio-demographic variables (sex, age, marital status, and employment), health related habits (smoking status, physical activity during leisure time), medical history (prevalence of chronic diseases such as cancer, diabetes and cardiovascular disease), as well as anthropometric data (weight and height) previously validated in the cohort [22]. We calculated body mass index (BMI) as self-reported weight in kilograms divided by the square of the self-reported height in meters. Physical activity was ascertained through a baseline 17-item questionnaire Physical activity was previously validated in a subsample of the cohort, with weekly MET-hours shown to adequately correlate with objectively measured energy expenditure [Spearman \u03c1 = 0.51; 95% confidence interval (CI): 0.232, 0.707] [23].\n\nOutcome assessment  \nWe defined MetS according to the American Heart Association (AHA) and the International Diabetes Federation (IDF) criteria as outlined in the harmonized definition for MetS [4]. According to this harmonized definition, the MetS diagnosis requires the presence of at least 3 of the 5 criteria: central adiposity (in our study waist circumference: \u226594 cm in males and \u226580 cm in females), hypertriglyceridemia (\u2265150 mg/dL or condition-specific medication), low levels of high-density lipoprotein cholesterol (<40 mg/dL for men and < 50 mg/dL for women), elevated blood pressure (systolic \u2265 130 and/or diastolic \u2265 85 mmHg or antihypertensive drug treatment in a patient with a history of hypertension), and impaired glucose metabolism (fasting glucose \u2265 100 mg/dL or drug treatment of elevated glucose). For a sensitivity analysis we used a higher cut-off point for central adiposity (waist circumference: \u2265102 cm in males and \u226588 cm in females) as proposed by the WHO for Caucasian populations [4].\n\nSelf-reported information about each specific MetS criterion was collected in the Q_6 (6th year follow-up questionnaire). With the Q_6 questionnaire, a measuring tape and an explanation of how to measure their own waist were sent to each participant. Diagnosis of MetS itself, the same as each MetS criterion was previously validated in a subsample of the cohort [24,25]. The validation study for MetS diagnosis found agreement between self-reported MetS diagnosis and MetS diagnosis from participant medical records to be 91.2% for confirmed MetS (95% CI: 80.7-97.1%) and 92.2% for non-confirmed MetS (95% CI: 85.1-96.4%) [24].\n\nIncident cases of MetS were defined as all participants who did not have any MetS criterion at baseline, and reported 3 or more criteria of MetS in the 6th year follow-up questionnaire.\n\nStatistical analyses  \nThree categories according to their baseline yogurt consumption were used to classify the participants included in the study: 0\u2013250 g/week (0\u20132 servings/week), >250 to <875 g/week (>2 to <7 servings/week), and \u2265875 g/week (\u22657 servings/week). The group with the lowest frequency of consumption (0\u2013250 g/week) was considered as the reference category.\n\nLogistic regression models were fitted to assess the relationship between total, whole-fat or low-fat yogurt consumption at baseline and the risk of developing incident MetS during the first 6 years of follow-up. The specific relationship between each of the 5 components of the MetS and total, whole-fat and low-fat yogurt consumption were also assessed. Odds Ratios (OR) and their 95% CI were calculated. Tests for linear trend across increasing categories of yogurt consumption were conducted by assigning the median consumption within each category and treating this variable as a continuous variable.\n\nThe interaction between adherence to Mediterranean diet [low (0\u20134 pts)/high (5\u20138 pts)], and fruit consumption [low (under the median: <264.5 g/d)/high (above the median: \u2265264.5 g/d)] and total yogurt consumption were tested using likelihood ratio tests comparing the fully adjusted model and the same model with the interaction product-term (2 degrees of freedom).\n\nAdditional analyses were conducted to test the association of the joint exposure to total yogurt consumption and fruit consumption with the risk of MetS, therefore we built six categories of the joint combined exposure to total yogurt consumption (low, medium and high consumption) and fruit consumption (under the median and at the median or higher) considering as the reference category those who had low yogurt consumption (<250 g/week) and low fruit consumption (under the median).\n\nWe fitted a first model without any adjustment (crude), a second model adjusted for age and sex, and a third multivariable-adjusted model after additional adjustment for the following potential confounders: baseline weight (kg), total energy intake (kcal/d), alcohol (g/d), soft drink consumption (ml/d), red meat consumption (g/d), consumption of French fries (g/d), consumption of fast food (g/d), Mediterranean dietary pattern adherence (3 categories: low, medium, and high adherence), physical activity (METs-hours/week), sitting hours (hr/d), sedentary behavior (hr/d), smoking status (non-smoker, smoker, former smoker), snacking between meals (yes/no), and following a special diet (yes/no). Snacking between meals was defined as those participants who responded affirmatively to the question \u201cDo you eat between main meals (snacking)?\u201d We did not consider if it was healthy or unhealthy snacking. The percentage of participants who reported snacking between meals was 33.3%. We considered that a participant followed a special diet, if she/he answered yes to the question \u201cDo you follow any special diet\u201d. If they had answered yes, they were asked which kind of special diet they followed (such as hypocaloric diet, vegetarian diet, low-fat diet, low sodium diet, high protein diet, among others). We used this variable dichotomously. The percentage of participants following a special diet was 6.1%.\n\nFor the selection of the potential confounders in the multivariate model, and as its currently recommended [26], we took into account the previously published scientific literature including our own results based on the cohort about risk factors for hypertension, overweight/obesity, type 2 diabetes, avoiding exclusively the statistical approach, the stepwise procedures, or the changes in the point estimates after adjusting for potential confounders.\n\nAll p values presented are two-tailed; p < 0.05 was considered statistically significant. Analyses were performed using STATA/SE version 12.0 (StataCorp, College Station, TX, USA).",
        "results": "## Results\nThe main characteristics of participants according to their frequency of yogurt consumption are presented in Table 1. The mean age of the participants was 36.4 y (SD: 11.6) and the mean BMI was 22.7 kg/m[2] (SD: 2.7).\n\nOn average participants in the highest category of total yogurt consumption had the lowest BMI, were more physically active, less likely to be current smokers, spent less hours sitting, and had the highest total energy intake, carbohydrate and protein intake and the lowest fat intake.\n\nThe association between the frequency of yogurt consumption and the risk of incident MetS was assessed in 8,063 participants initially free of MetS. During the first 6 years of follow up we observed 306 incident cases of MetS. As displayed in Table 2, a higher frequency of total, whole-fat or low-fat yogurt consumption was not significantly associated with a lower risk of developing MetS. Participants consuming total, whole-fat and low-fat \u2265875 g/week did not exhibit significantly lower risk of developing MetS compared with those who consumed from 0 to 250 g/week. However, we observed non-significant inverse associations [OR = 0.84 (95% CI: 0.60-1.18); 0.98 (95% CI: 0.68-1.41); and 0.63 (95% CI: 0.39-1.02)] respectively for total, whole-fat and low-fat yogurt consumption, after adjusting for potential confounders (Table 2).\n\nThe risk of developing each component of MetS according to the total, whole-fat and low-fat yogurt consumption is shown in Figure 1. This figure shows the ORs for each component of the MetS separately for total, whole-fat and low-fat yogurt consumption only for the highest category of consumption (\u2265875 g/week) versus the lowest category (\u2264250 g/week). From the 5 components of MetS, we observed that those participants who consumed \u2265875 g/week of total and whole-fat yogurt compared with those who consumed \u2264 250 g/week had a lower risk of developing central adiposity [OR = 0.85 (95% CI: 0.74-0.98) and 0.85 (95% CI: 0.73-0.99) respectively, after adjusting for potential confounders. The results for the rest of the components suggested a non-significant trend for a lower risk (except for blood pressure).\n\nAs a sensitivity analysis we changed the cut-off point for central adiposity (waist circumference: \u2265102 cm in males and \u226588 cm in females) as proposed by the WHO for Caucasian populations. The ORs for total, whole-fat and low-fat yogurt consumption (\u2265875 g/week versus \u2264250 g/week) were 0.74 (95% CI: 0.48-1.14); 0.76 (95% CI: 0.46-1.25); and 0.83 (95% CI: 0.47-1.45), respectively (data not shown).\n\nNo interaction between Mediterranean diet adherence (low/high) and total yogurt consumption was observed (p for interaction = 0.270). The interaction between fruit consumption (below/above median) and total yogurt consumption approached the threshold for statistical significance (p for interaction = 0.070). In any case, the lowest ORs in Table 3 were observed in the groups of participants with both a higher adherence to the Mediterranean diet and higher consumption of total yogurt [OR = 0.77 (95% CI: 0.46-1.30) p for trend = 0.323] and in participants with both above-the-median fruit consumption and high yogurt consumption [OR = 0.69 (95% CI: 0.43-1.10) p for trend = 0.104] (Table 3).\n\nWhen we simultaneously analyzed total yogurt consumption (3 categories) and fruit consumption (2 categories), 6 new categories were created. The category of lower yogurt consumption (\u2264250 g/week) and low fruit consumption (under the median < 264.5 g/d) was considered as the reference group. We observed a significantly lower risk of developing MetS in the group where a higher yogurt consumption (\u2265875 g/week) was accompanied by a high fruit consumption (above the median \u2265264.5 g/d) OR = 0.61 (95% CI: 0.38-0.99) (Table 4).",
        "column_title": "Comparability (Maximum 2 points awarded for controlling for the pre-specified primary confounding variable (age) and additional confounding variables.)",
        "instruction": "Rate the given study and provide the most suitable score from 0, 1, 2",
        "type": "score",
        "truth": "2"
    }
]