ACL_24_with_limitation/ACL_24_1064.json

{
  "File Number": "1064",
  "Title": "MedExQA: Medical Question Answering Benchmark with Multiple Explanations",
  "Limitation": "Limitation\nWhile MedExQA provides a robust benchmark for evaluating LLMs in the context of the medical domain, the current version only tests the model’s ability in QA task, limiting its applicability in realworld clinical scenarios to a few applications. This limitation results from the manual collection process. Future work will extend our benchmark to include tasks such as summarizing clinical notes with accompanying explanations. Though we performed the human evaluation of generated explanations of different LLMs through three authors, we performed this at a small scale, at 5 samples per specialty. Future work will seek to\nincrease both the volume of samples and the number of annotators to provide a more robust method of assessing models’ performance. Broader Impacts and Ethics Statement\nWe release MedExQa under a Creative Commons Attribution-Non Commercial-ShareAlike 4.0 International License. MedPhi-2 follows the MIT license as it is based on Phi-2. License and copyright information and Terms of Use will be shared when the dataset and model are released. The dataset may be used for non-commercial purposes and any models trained using the dataset should be used only for research purposes. Our work does not raise any major ethical concerns. All LLMs tested, including Phi-2, were used for research purposes only. While MedPhi-2 outperformed all medical variants of Llama2 models in generating accurate medical answers and explanations, MedPhi-2 is not rigorously tested for use in real-world clinical applications or scenarios. Thus, MedPhi-2 is not suitable for use in the clinical decision making process. This restriction of usage in clinical care is to mitigate any potential risks or harms such as wrong decisions from hallucinations which can lead to unwanted situations.",
  "abstractText": "This paper introduces MedExQA, a novel benchmark in medical question-answering, to evaluate large language models’ (LLMs) understanding of medical knowledge through explanations. By constructing datasets across five distinct medical specialties that are underrepresented in current datasets and further incorporating multiple explanations for each questionanswer pair, we address a major gap in current medical QA benchmarks which is the absence of comprehensive assessments of LLMs’ ability to generate nuanced medical explanations. Our work highlights the importance of explainability in medical LLMs, proposes an effective methodology for evaluating models beyond classification accuracy, and sheds light on one specific domain, speech language pathology, where current LLMs including GPT4 lack good understanding. Our results show generation evaluation with multiple explanations aligns better with human assessment, highlighting an opportunity for a more robust automated comprehension assessment for LLMs. To diversify open-source medical LLMs (currently mostly based on Llama2), this work also proposes a new medical model, MedPhi-2, based on Phi-2 (2.7B). The model outperformed medical LLMs based on Llama2-70B in generating explanations, showing its effectiveness in the resource-constrained medical domain. The benchmark datasets and the model can be found at https://github.com/knowlab/MedExQA.",
  "1 Introduction": "Recent advancements in large language models (LLMs) have not only enhanced their understanding of medical domain text but also improved their ability to generate coherent text with correct medical knowledge (Tu et al., 2023; Singhal et al., 2023). Chatbots, powered by these LLMs, have emerged as indispensable tools, offering unprecedented opportunities to enhance patient care, streamline clinical decision-making processes, and medical knowl-\nedge retrieval (Achiam et al., 2023; OpenAI, 2023; Groves et al., 2023). Moreover, open-source medical LLMs further enhance the usability of such technologies in hospitals by resolving the privacy concerns associated with patient data (Toma et al., 2023; Kweon et al., 2023; Chen et al., 2023).\nThis research in medical LLMs has been facilitated by the introduction of question-answering (QA) datasets that serve as benchmarks for evaluating the model’s understanding of medical domain knowledge (Hendrycks et al., 2020; Jin et al., 2021; Pal et al., 2022; Singhal et al., 2023). The benchmark QA datasets typically consist of multiplechoice questions (MCQ), enabling researchers to readily assess the capabilities of LLMs in comprehending and responding to diverse medical inquiries. Thus, the diversity within these datasets is a key component in creating a rigorous assessment benchmark for complex medical concepts. Nonetheless, certain areas within the medical domain, such as speech language pathology, still remain uncovered by the current benchmark datasets.\nAs current medical QA benchmarks are often structured as MCQ, classification accuracy is used as an evaluation metric. However, classification accuracy alone may not adequately assess whether LLMs possess the nuanced medical expertise required for reasoned responses. The explanation and rationale behind the selection of a particular choice by an LLM would provide a deeper understanding of the model’s capabilities and limitations in generating responses to intricate medical questions. This comprehensive evaluation, delving into the explanation and rationale, is especially important in clinical settings where misleading information such as hallucinations produced by LLMs can have serious consequences.\nIn order to assess the quality of the model explainability, the dataset should include a golden explanation for the reasoning behind the answer. Additionally, since there are often multiple ways to\nexpress the same rationale in text, an ideal dataset would provide a multiple set of explanations for a single QA pair. However, current benchmark datasets are not focused on providing explanations as they often lack explanations entirely or only a subset of the dataset comes with an explanation (Hendrycks et al., 2020; Jin et al., 2021; Pal et al., 2022). This limitation highlights the need for improved datasets that are explicitly designed to include comprehensive explanations.\nTo address this issue, this paper presents a novel QA benchmark, MedExQA, with two sets of explanations, aiming to provide a more comprehensive evaluation of LLMs in the medical domain. To diversify the knowledge coverage in the current datasets, our proposed benchmark consists of five underrepresented specialties in current datasets: biomedical engineering, clinical laboratory science, clinical psychology, occupational therapy, and speech language pathology. In this work, the datasets were used to benchmark the performance of an extensive list of LLMs, including those trained with medical domain text. With this comprehensive benchmark evaluation, we explored the effects of medical domain-specific training. Additionally, to diversify the pool of open-source medical LLMs which are currently almost all based on the Llama2 model, we introduce our own trained model, MedPhi-2, a Phi-2 model trained with medical domain text. Our MedPhi-2 model outperformed medical LLMs based on the Llama2-70B model in generating explanations for the rationale behind the answer.\nThe contributions of this paper are as follows:\n1. MedExQA novel datasets with explana-\ntions. We constructed a benchmark with 5 distinct specialties within the medical domain. The datasets include two explanations for each question and answer pairs.\n2. Comprehensive Benchmark. We evaluated\nan extensive list of models: 18 baseline opensource models with various sizes (from 2.7B to 70B), 3 OpenAI GPT models, as well as our model (detailed below). In terms of evaluation approach, classification accuracy, generated explanation performance, and human evaluations are considered. To highlight, this is the first benchmark using multiple explanations, and the results demonstrate that our benchmark can better evaluate language models’ understanding of medical domain knowledge.\n3. MedPhi-2 model. We trained a small lan-\nguage model (SLM) based on the Phi-2 model, with medical pretraining corpus and instruction-tuning datasets. The model outperformed medical LLMs based on Llama2 70B in generating explanations.\n4. Open source. We release the datasets, model\nweights, and codes to facilitate the research in medical large language modeling.",
  "2.1 MMLU": "MMLU (Hendrycks et al., 2020) is a benchmark designed to measure the model’s ability in knowledgeintensive QA with four-way MCQs. Within the extensive list of subjects, there are nine healthcarerelated subjects such as professional medicine and medical genetics. Collectively, these nine subjects comprise a total of 1,871 questions in the test set. While MMLU provides a comprehensive set of questions, it lacks explanations for the answers, thereby limiting the dataset’s evaluation to mere multiple-choice classification accuracy.",
  "2.2 MedQA": "MedQA (Jin et al., 2021) is an open-ended MCQ dataset made from professional medical doctor license exams. The dataset contains questions drawn from both real exams and mock tests for the United States Medical License Exams (USMLE). 1,273 questions, each question accompanied by four or five answer choices, are provided as the test dataset. Similar to MMLU, MedQA does not include explanations for assessing the ability to generate rationale behind the answer.",
  "2.3 MedMCQA": "MedMCQA (Pal et al., 2022) is a benchmark with questions sourced from postgraduate-level Indian medical school entrance exams (AIIMS and NEET PG). The dataset covers a breadth of medical specialties, 2,400 healthcare topics and 21 subjects and provides 4,183 MCQ with four answer choices for evaluation. Although MedMCQA is known to have explanations, nearly half of the evaluation dataset lacks explanations and instances of duplicate explanations are also observed. In fact, accuracy is only reported as the evaluation metric and explanation is not used in their paper entirely. Therefore, MedMCQA is not primarily designed for the assessment of generating explanations.",
  "3 MedExQA Datasets": "We introduce MedExQA, a novel QA benchmark designed to tackle the limitations of existing benchmarks by incorporating two sets of explanations. This approach aims to offer a more thorough evaluation of performance in five underrepresented specialties in the medical domain: Biomedical Engineering, Clinical Laboratory Science, Clinical Psychology, Occupational Therapy, and Speech Language Pathology.",
  "3.1 Datasets Preparation": "The raw data was manually collected from diverse freely accessible online sources, including mock tests and online exams tailored to each medical professional specialty. Some questions of the mock tests and online exams have explanations for the answers, which we used the creation of the MedExQA datasets. The pass mark for the collected mock tests and online exams was 60 percent.\nTo ensure data integrity, rigorous preprocessing was conducted, including the removal of duplicate questions and explanations. Additionally, similar questions were identified and eliminated using BERT cosine similarity analysis (Devlin et al., 2018). Questions containing keywords specific to laws or regulations were filtered out using a manually curated list of words. Following fair use regulations1, answer options were systematically shuffled to maintain fairness and uphold the integrity of the dataset. Furthermore, to enhance the quality and coherence of the datasets, two sets of explanations as well as the questions underwent thorough human validation. This validation process aimed to ensure that the explanations exhibited distinct writing styles and provided comprehensible reasoning for the correct answer selection.\n1https://www.copyright.gov/fair-use/more-info.html\nThe resulting datasets have a total of 965 questions. Table 1 provides a detailed breakdown of the number of questions for each specialty. These datasets were split into a few-shot development set and a test set. Specifically, the few-shot development set has 5 questions per specialty, while the test set consists of 940 questions in total. It is noteworthy that each subject contains a minimum of 100 test examples, a length surpassing that of most exams tailored for human assessment.\nAlso, to validate that each pair of explanations is different sufficiently at the individual question level, Table 1 also provides the average cosine similarity of the pairs. The overall similarity is 78.7% which indicates the lexical difference of the two corresponding versions of explanations for each question. An example of the dataset as well as the difference in the pair of example can be found in the Appendix Figure 4.",
  "3.2 Comparison of benchmark datasets": "We compared MedExQA with existing benchmark datasets by visualizing their questions in the same vector space. Using t-distributed Stochastic Neighbour Embedding (tSNE), each question is represented as a point in the vector space. We used the ‘all-mpnet-base-v2’ sentence transformer model in sklearn package tSNE to retrieve vectors from questions. 965 questions were randomly sampled from each dataset. There is a cluster towards the top region mainly composed of questions from MedExQA, which clearly demonstrates its novelty compared to existing medical QA datasets.",
  "4 Methods": "For all the experiments in this paper, both training and evaluation, we used 8 A6000 GPUs.",
  "4.1 Baseline Models": "We explored 18 baseline models with different sizes from 2.7B to 70B. Table 2 provides a comprehensive overview of the baseline models used in this paper, while more detailed descriptions of each model are available in the appendix. In cases where multiple sizes of a model are used, we distinguish each version by appending the model size to the model name. For example, the Llama2 models with sizes 7, 13, and 70B are denoted as Llama2-7B, Llama2-13B, and Llama2-70B, respectively. On the other hand, when a model has only one size, we refer to it solely by its name. For instance, ClinicalCamel denotes the ClinicalCamel 70B model.",
  "4.2 Training MedPhi-2": "As far as we know, all the publicly available opensource medical LLMs are based on Llama models, we further extended our work to test the effect of medical domain training on a different foundational model. Phi-2 model was further trained using the medical datasets that are publicly available. We pretrained Phi-2 with a 110M medical-related corpus. We further finetuned the continued pretrained model with 239K instructions. We refer to the resulting model as MedPhi-2 throughout our paper. Table 3 summarizes the detailed composition of our training dataset. We used LLaMaFactory2 and used\n2https://github.com/hiyouga/LLaMA-Factory\nDeep3 for efficient training. For both pretraining and finetuning, We trained the model with a batch size of 16 and a learning rate of 1e-5 with 3 epochs, which took 36 hours in total.",
  "4.3 Evaluation": "We evaluated all models with test datasets except for human evaluation, which was performed on the development datasets. For all the evaluations, we used zero-shot, a batch size of 1, temperature of 0. To benchmark the performance of closed source models we further extended to include OpenAI’s GPT models. We used GPT3.5_1106, GPT4.0_1106, and GPT4.0_0125 APIs11.",
  "4.3.1 Classification Accuracy - Logits": "Classification accuracy of MCQ for generative models relies on classifying the next token using logits. In other words, the token with the highest logit value is selected as the model’s predicted answer. However, this approach cannot assess the model’s understanding of the rationale behind the answer. We exclude GPT models for this evaluation, as we are not able to get the logit value for the next token.\n3https://huggingface.co/datasets/epfl-llm/guidelines 4https://huggingface.co/datasets/FremyCompany/AGCT-\nDataset 5https://huggingface.co/datasets/paniniDot/sci_lay 6https://huggingface.co/datasets/gamino/wiki_medical_terms 7https://huggingface.co/datasets/zhengyun21/pmcpatients 8https://huggingface.co/datasets/starmpcc/asclepiussynthetic-clinical-notes 9https://huggingface.co/datasets/caseymartin/medinstruct 10https://github.com/CogStack/OpenGPT 11https://platform.openai.com/docs/models",
  "4.3.2 Classification Accuracy - Chat": "We utilize string-matching using regular expressions and thefuzz package to assess the model’s proficiency in generating accurate textual responses. This approach involves searching the specific phrase for the answer choice or the choice letter within the generated response, enabling a more realistic evaluation for the model’s performance.",
  "4.4 Explanation Generation": "The quality of generated explanations is further assessed using a combination of general lexical metrics. BLEU (Papineni et al., 2002). measures the geometric mean of precision scores of the generated explanations compared to reference explanations based on n-gram matches. ROUGE (Lin, 2004). assesses the similarity between generated and reference explanations, with ROUGE-L, providing a score that combines precision and recall based on the longest common subsequence. METEOR (Banerjee and Lavie, 2005). considers the semantic similarity and lexical variations with WordNet. BERTScore (Zhang et al., 2019). uses contextual embeddings, scibert embedding (Beltagy et al., 2019) for our work, to capture nuances in the semantics of the explanations. All the metrics are calculated using evaluate package.\nWe propose an enhanced methodology for evaluating models’ understanding of medical domain knowledge by incorporating classification accuracy based on string matches into calculating these metrics. We assign a score of 0 to responses with incorrect answers based on string-matching classification results.",
  "4.5 Evaluation - Human Evaluation": "For human evaluation, three human annotators with MSc degrees in health-related subjects participated in assessing the quality of generated explanations. The evaluation process involved assigning a score for each explanation-answer pair based on the following rules:\n1. Score 0 the answer was incorrect, no explana-\ntion was provided, and/or the explanation is fully irrelevant.\n2. Score 0.5 the answer was correct, but the ex-\nplanation or rationale was incorrect. Also, an incomplete explanation that ended with an incomplete sentence.\n3. Score 1.0 when both the answer and explana-\ntion were correct.\nAlthough this human evaluation was performed on a small scale (development dataset: 5 samples for each specialty), this systematic evaluation process ensured a comprehensive assessment of the models’ performance in providing accurate and coherent explanations.",
  "5.1 Classification Accuracy - Logits": "Table 4 shows the detailed results of all models. As expected, smaller language models demonstrated lower accuracy across specialties than larger models. Med42 showed the best overall performance. It showed outstanding performance in Biomedical Engineering and Clinical Laboratory Science (83.2% and 84.9% respectively). It performed on par with Meditron-70B in Clinical Psychology (84.9%). In Occupational Therapy, Llama2-70B showed the highest accuracy (80.4%). All models underperformed in Speech Language Pathology, with SOLAR performing the best (33.1%).\nThe effect of continued training is observed only in some models. MedPhi-2 demonstrated better performance than Phi-2, and this improvement was also found in AlpaCare-13B compared to Llama213B and Med42 compared to Llama2-70B. However, ClinicalCamel and Meditron-70B performed worse than Llama2-70B. This drop in performance could be due to task-specific challenges as some models may not effectively handle varied levels of specificity in MedExQA.",
  "5.2 Classification Accuracy - Chat": "Classification accuracy using chat decreased in most of the models (Table 4). Phi-2, Llama2-13B, Yi, InternLM2, and Meditron-70B did not pass the pass mark indicating these models are not robust. Meditron-70B showed the biggest performance drop by 29.3%. Llama2-70B also showed a significant performance drop in this testing by 28.5%, although it passed in Biomedical Engineering. Of the 70B models we tested, ClinicalCamel was the most robust model (7.7% decrease), and it scored higher than Med42 by 0.7%.\nOur model, MedPhi-2 was the most robust model among the passed ones (0.2% decrease), and it outperformed AlpaCare-13B, Meditron-70B, Llama270B. This result highlights the importance of the supervised finetuning with in-domain instructions of high quality as more robust models, such as AlpaCare, ClinicalCamel, and MedPhi-2, were\ninstruction-tuned with medical domain data, while Meditron-70B was just further pretrained.\nGPT4_0125, GPT4_1106, and GPT3.5_1106 outperformed all the open-source models. Even with the addition of high-performing closed-source models, there is still a universal failure in performance for Speech Language Pathology.",
  "5.3 Combining Classification Accuracy with Generated Explanation Performance": "Figure 2 shows the relationship between model size and accuracy achieved in both MCQ (using logits) and generation performance. Generally, larger models tend to exhibit better performance as 70B models perform better than most of the other smaller models. However, SOLAR, Yi, and Mistral stand out as these smaller general domain models demonstrate competitive performance to the 70B medical LLMs. Further training on these foundation models holds great promise as we have seen with the Phi-2 model.\nAll medical LLMs with 13B (AlpaCare, Asclepius, and Meditron) exhibit worse performance in both MCQ accuracy and generation performance compared to their 7B counterparts. In\nfact, Medinote-13B is the worst-performing model. Also, 70B models do not always perform better than smaller models as Meditron-70B and Llama270B performed worse than many smaller models including AlpaCare and our model in the generation of reasonable explanations.\nThe performance evaluation presented in Table 5 also provides valuable insights into the efficacy of various models in generating explanations. Among the models evaluated, our model, MedPhi-2 stands out in generating reasonable explanations as it outperformed all medical LLMs including 70B models. This result confirms the findings of Section 5.2 which highlighted the importance of supervised finetuning with in-domain instructions.\nThe SOLAR model performed the best among the open-source models, suggesting its competitive capability in explanation generation although it was not trained specifically for the medical domain. However, even this best-performing open-source model demonstrates a significant performance gap (5.64) compared to the worst-performing closedsource model, GPT3.5_1106, indicating the substantial advancements in OpenAI’s GPT models.\nInterestingly, despite the recent release of GPT4,\nthe performance varies across different evaluation metrics. While the most recent release outperforms GPT4_1106 on average, GPT4_1106 still shows superior performance in METEOR and BERTScore. This highlights the importance of considering multiple metrics and nuances in model performance assessment, as different models may excel in distinct aspects of explanation generation.",
  "5.4 Evaluation - Human Evaluation": "Human evaluation of generated responses reveals that MedPhi-2 has the best quality among the opensource models (Figure 3). Our model was the only open-source model that passed (a score of 3 or above) in all specialties in MedExQA. In fact, MedPhi-2 on par with GPT3.5_1106 in Biomedical Engineering and Clinical Laboratory Science, and with GPT4_1106 in Occupational Therapy.\nThe performance of models in Speech Language Pathology during human evaluation was relatively decent, which contrasts with results obtained through other evaluation methods. Appendix Figure 4 provides an example of generated responses of the models, in the context of Speech-\nLanguage Pathology questions. MedPhi-2 and GPT3.5_1106 generated the most coherent and accurate responses. However, other models generated irrelevant sentences or failed to provide explanations. Medinote-13B generated a case study example instead of answering the question and providing an explanation and Asclepius-13B hallucinated and provided an option for the answer that was not present and generated further incorrect explanations. Appendix Table 6 shows the detailed results.",
  "5.5 Effect of additional explanation": "The effect of adding additional explanation was confirmed by analyzing the Pearson correlation between human evaluation and generation performance. When we used just one set of explanations the correlation was 0.9347, and this correlation increased to 0.9385 when we used two versions of explanations together. Although, the increase is small, this finding still indicates generation evaluation with multiple explanations aligns better with human evaluation, which is usually treated as the gold standard.",
  "6 Conclusion": "Our MedExQA benchmark proposes an effective methodology for evaluating LLMs beyond classification accuracy which can be used to assess the explainability of medical LLMs. While, the findings reveal that the generation of coherent and accurate explanations remains a challenging frontier for the current medical LLMs, the results also highlight an opportunity for a more robust automated comprehension assessment for LLMs because generation evaluation with multiple explanations aligned better with human assessment.\nWe also find that the ‘Speech Language Pathology’ dataset posed challenges for all language models, including GPT4. Speech Language Pathology could potentially be attributed to several factors, with one prominent explanation being the absence of relevant text in the corpora used to train the foundation model. As Speech Language Pathology is a highly specialized field that encompasses a wide range of topics related to rare diseases or disorders of speech and language, the collection of high-quality text for this specialty can be very challenging. However, it is important to acknowledge that confirming this hypothesis definitively poses a challenge due to the proprietary nature of the pretraining corpora used for training LLMs.\nThrough the development and evaluation of our MedPhi-2 model, we underscore the importance of targeted pretraining and fine-tuning strategies in improving explanation quality. The model showed the significant potential of LLMs in enhancing medical QA with explanations. Our benchmark and model will set the foundation for future advancements in medical research by facilitating the development and evaluation of medical LLMs.\nLimitation\nWhile MedExQA provides a robust benchmark for evaluating LLMs in the context of the medical domain, the current version only tests the model’s ability in QA task, limiting its applicability in realworld clinical scenarios to a few applications. This limitation results from the manual collection process. Future work will extend our benchmark to include tasks such as summarizing clinical notes with accompanying explanations.\nThough we performed the human evaluation of generated explanations of different LLMs through three authors, we performed this at a small scale, at 5 samples per specialty. Future work will seek to\nincrease both the volume of samples and the number of annotators to provide a more robust method of assessing models’ performance.\nBroader Impacts and Ethics Statement\nWe release MedExQa under a Creative Commons Attribution-Non Commercial-ShareAlike 4.0 International License. MedPhi-2 follows the MIT license as it is based on Phi-2. License and copyright information and Terms of Use will be shared when the dataset and model are released. The dataset may be used for non-commercial purposes and any models trained using the dataset should be used only for research purposes.\nOur work does not raise any major ethical concerns. All LLMs tested, including Phi-2, were used for research purposes only. While MedPhi-2 outperformed all medical variants of Llama2 models in generating accurate medical answers and explanations, MedPhi-2 is not rigorously tested for use in real-world clinical applications or scenarios. Thus, MedPhi-2 is not suitable for use in the clinical decision making process. This restriction of usage in clinical care is to mitigate any potential risks or harms such as wrong decisions from hallucinations which can lead to unwanted situations."
}