| # OHCA Classifier v3.0 - Trained Model | |
| ## Model Description | |
| This is a trained BERT-based classifier for detecting Out-of-Hospital Cardiac Arrest (OHCA) cases in medical discharge notes. The model is fine-tuned from PubMedBERT and achieves high sensitivity for OHCA detection with configurable thresholds for different clinical needs. | |
| ## Model Details | |
| - **Model Name**: OHCA Classifier v3.0 - Trained | |
| - **Base Model**: microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract | |
| - **Task**: Binary text classification (OHCA vs Non-OHCA) | |
| - **Language**: English | |
| - **Domain**: Medical/Clinical text | |
| - **Model Version**: 3.0 | |
| - **Author**: Mona Moukaddem | |
| - **Model Size**: 109M parameters | |
| - **License**: MIT | |
| ## Performance Metrics | |
| | Metric | Value | Description | | |
| |---|---|---| | |
| | Optimal Threshold | 0.996 | Found via validation set optimization | | |
| | F1-Score | 0.632 | Harmonic mean of precision and recall | | |
| | Sensitivity (Recall) | 1.000 | 100% - Catches all OHCA cases at optimal threshold | | |
| | Specificity | 0.741 | 74.1% - Correctly identifies non-OHCA cases | | |
| | AUC-ROC | High | Excellent discrimination ability | | |
| ## Threshold Selection Guide | |
| **For Clinical Screening (Recommended): 0.90** | |
| - Good balance of sensitivity and specificity | |
| - Reduces false positives while maintaining high sensitivity | |
| - Suitable for most clinical workflows and screening applications | |
| **For Ultra-Conservative Screening: 0.996** | |
| - Optimal threshold from validation set optimization | |
| - Maximizes sensitivity (100%) | |
| - May produce more false positives in some populations | |
| - Use when missing OHCA cases is extremely costly | |
| **For Research/Validation: Variable** | |
| - Adjust based on your specific requirements | |
| - Consider your population's OHCA prevalence | |
| - Validate performance on your own dataset | |
| ## Training Data | |
| | Dataset Characteristic | Value | | |
| |---|---| | |
| | Total Cases | 330 | | |
| | OHCA Cases | 59 (17.9%) | | |
| | Non-OHCA Cases | 271 (82.1%) | | |
| | Training Split | 264 cases | | |
| | Validation Split | 66 cases | | |
| | Data Source | MIMIC-III derived discharge notes | | |
| ## Usage | |
| ### Quick Start | |
| ```python | |
| from transformers import AutoTokenizer, AutoModelForSequenceClassification | |
| import torch | |
| # Load the model | |
| model_name = "monajm36/ohca-classifier-v3-trained" | |
| tokenizer = AutoTokenizer.from_pretrained(model_name) | |
| model = AutoModelForSequenceClassification.from_pretrained(model_name) | |
| # Threshold options | |
| recommended_threshold = 0.90 # Recommended for clinical screening | |
| optimal_threshold = 0.996 # From validation set optimization | |
| def predict_ohca(text, threshold=0.90): | |
| """Predict OHCA from medical text""" | |
| inputs = tokenizer(text, truncation=True, padding=True, | |
| max_length=512, return_tensors="pt") | |
| with torch.no_grad(): | |
| outputs = model(**inputs) | |
| probs = torch.softmax(outputs.logits, dim=-1) | |
| ohca_prob = probs[0][1].item() | |
| prediction = 1 if ohca_prob >= threshold else 0 | |
| # Clinical priority based on probability | |
| if ohca_prob >= 0.996: | |
| priority = "Immediate Review" | |
| elif ohca_prob >= 0.90: | |
| priority = "Priority Review" | |
| elif ohca_prob >= 0.70: | |
| priority = "Consider Review" | |
| else: | |
| priority = "Routine" | |
| confidence = "High" if ohca_prob >= 0.90 else "Medium" if ohca_prob >= 0.50 else "Low" | |
| return { | |
| "prediction": "OHCA" if prediction == 1 else "Non-OHCA", | |
| "probability": ohca_prob, | |
| "confidence": confidence, | |
| "clinical_priority": priority, | |
| "threshold_used": threshold | |
| } | |
| # Example usage | |
| text = "Patient presents with cardiac arrest at home, found down by family" | |
| result = predict_ohca(text) # Uses recommended 0.90 threshold | |
| print(f"Prediction: {result['prediction']}") | |
| print(f"Probability: {result['probability']:.3f}") | |
| print(f"Clinical Priority: {result['clinical_priority']}") | |
| ``` | |
| ### Pipeline Usage | |
| ```python | |
| from transformers import pipeline | |
| # Create classification pipeline | |
| classifier = pipeline("text-classification", model="monajm36/ohca-classifier-v3-trained") | |
| # Classify medical text | |
| text = "Patient presents with cardiac arrest at home" | |
| result = classifier(text) | |
| print(result) | |
| # Output: [{'label': 'LABEL_1', 'score': 0.998}] | |
| # LABEL_0 = Non-OHCA, LABEL_1 = OHCA | |
| # For clinical use, apply appropriate threshold: | |
| probability = result[0]['score'] if result[0]['label'] == 'LABEL_1' else 1 - result[0]['score'] | |
| is_ohca_90 = probability >= 0.90 # Recommended threshold | |
| is_ohca_996 = probability >= 0.996 # Optimal threshold | |
| ``` | |
| ### Batch Processing | |
| ```python | |
| import pandas as pd | |
| def process_medical_notes(df, text_column='clean_text', threshold=0.90): | |
| """Process multiple medical notes""" | |
| results = [] | |
| for text in df[text_column]: | |
| result = predict_ohca(text, threshold=threshold) | |
| results.append(result) | |
| # Add results to dataframe | |
| df['ohca_prediction'] = [r['prediction'] for r in results] | |
| df['ohca_probability'] = [r['probability'] for r in results] | |
| df['clinical_priority'] = [r['clinical_priority'] for r in results] | |
| return df | |
| # Example with DataFrame | |
| medical_notes = pd.DataFrame({ | |
| 'patient_id': [1, 2, 3], | |
| 'clean_text': [ | |
| "Patient found in cardiac arrest at home by spouse", | |
| "Patient complains of chest pain, vital signs stable", | |
| "Witnessed cardiac arrest in emergency department" | |
| ] | |
| }) | |
| results = process_medical_notes(medical_notes) | |
| print(results[['patient_id', 'ohca_prediction', 'ohca_probability']]) | |
| ``` | |
| ### Compare Different Thresholds | |
| ```python | |
| def compare_thresholds(text): | |
| """Compare predictions at different thresholds""" | |
| thresholds = [0.50, 0.70, 0.90, 0.996] | |
| for threshold in thresholds: | |
| result = predict_ohca(text, threshold=threshold) | |
| print(f"Threshold {threshold}: {result['prediction']} " | |
| f"(p={result['probability']:.3f}, priority={result['clinical_priority']})") | |
| # Example comparison | |
| text = "Patient found down at home, family performed CPR" | |
| compare_thresholds(text) | |
| ``` | |
| ## Clinical Decision Support | |
| The model provides configurable sensitivity for OHCA detection, making it suitable for clinical screening where different thresholds may be appropriate based on clinical context and cost of missed cases. | |
| ### Clinical Workflow Integration | |
| | Probability Range | Clinical Priority | Recommended Action | | |
| |---|---|---| | |
| | ≥ 0.996 | 🔴 Immediate Review | Very high confidence - Urgent review required | | |
| | 0.90 - 0.995 | 🟡 Priority Review | High confidence - Clinical team review | | |
| | 0.70 - 0.89 | 🟠 Consider Review | Moderate confidence - Consider for review | | |
| | < 0.70 | 🟢 Routine | Low probability - Standard processing | | |
| ### Threshold Selection for Clinical Use | |
| **Use 0.90 threshold when:** | |
| - Screening large volumes of discharge notes | |
| - Balancing sensitivity with manageable false positive rates | |
| - Implementing in routine clinical workflows | |
| **Use 0.996 threshold when:** | |
| - Ultra-high sensitivity is required | |
| - Cost of missing OHCA cases is extremely high | |
| - You have resources to review more false positives | |
| ## Quality Assurance | |
| - **High Sensitivity**: Configurable thresholds ensure no OHCA cases are missed | |
| - **Optimal Threshold**: 0.996 maximizes sensitivity on validation data | |
| - **Clinical Threshold**: 0.90 provides practical balance for screening | |
| - **Patient-Level Training**: Prevents data leakage and overfitting | |
| - **Clinical Validation**: Designed for real-world medical text processing | |
| ## Model Architecture | |
| ``` | |
| PubMedBERT (microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract) | |
| ├── 12 Transformer layers | |
| ├── 768 hidden dimensions | |
| ├── 12 attention heads | |
| ├── 109M parameters | |
| └── Classification head (2 classes: OHCA vs Non-OHCA) | |
| ``` | |
| ## Training Details | |
| | Training Parameter | Value | | |
| |---|---| | |
| | Framework | PyTorch + Transformers | | |
| | Optimizer | AdamW | | |
| | Learning Rate | Default (with linear scheduling) | | |
| | Epochs | 3 | | |
| | Batch Size | 8 (with gradient accumulation) | | |
| | Max Sequence Length | 512 tokens | | |
| | Class Balancing | Weighted loss + minority oversampling | | |
| | Validation Strategy | Patient-level splits (prevents data leakage) | | |
| | Hardware | CPU training | | |
| ## Evaluation Strategy | |
| - **Patient-Level Data Splits**: Ensures all notes from the same patient stay in one split | |
| - **Optimal Threshold Finding**: Uses validation set to find best decision threshold | |
| - **Independent Test Set**: Unbiased evaluation on held-out data | |
| - **Clinical Metrics**: Focus on sensitivity for medical screening applications | |
| ## Limitations and Considerations | |
| ### Limitations | |
| - Trained on specific medical text format (discharge notes) | |
| - May not generalize to different hospital systems without fine-tuning | |
| - Performance may vary with different patient populations | |
| - Designed specifically for English medical text | |
| - Limited to text-based OHCA detection (no multimodal inputs) | |
| ### Ethical Considerations | |
| - **Clinical Use**: This model is intended to assist, not replace, clinical judgment | |
| - **Bias Monitoring**: Regular evaluation across different patient demographics recommended | |
| - **Human Oversight**: All high-probability predictions should be reviewed by medical professionals | |
| - **Privacy**: Ensure compliance with healthcare data regulations (HIPAA, etc.) | |
| ### Performance Variations | |
| Model performance may vary across different: | |
| - Hospital systems and documentation styles | |
| - Patient demographics and populations | |
| - Types of cardiac arrest presentations | |
| - Clinical documentation quality and completeness | |
| ## Related Work | |
| This model is based on the OHCA Classifier v3.0 methodology with significant improvements over previous versions: | |
| - **Enhanced Methodology**: Patient-level splits, optimal threshold finding | |
| - **Source Code**: Available at [monajm36/ohca-classifier-3.0](https://github.com/monajm36/ohca-classifier-3.0) | |
| - **Training Pipeline**: Complete v3.0 training workflow for custom model development | |
| - **Research Foundation**: Built on established medical NLP and machine learning best practices | |
| ## Installation and Dependencies | |
| ```bash | |
| pip install transformers torch pandas numpy | |
| ``` | |
| **Minimum Requirements:** | |
| - Python 3.8+ | |
| - PyTorch 1.9+ | |
| - Transformers 4.20+ | |
| - 4GB RAM for inference | |
| - GPU optional (model works on CPU) | |
| ## Citation | |
| If you use this model in your research or clinical work, please cite: | |
| ```bibtex | |
| @software{ohca_classifier_v3_trained, | |
| title={OHCA Classifier v3.0: Trained BERT Model for Cardiac Arrest Detection in Medical Text}, | |
| author={Mona Moukaddem}, | |
| year={2025}, | |
| url={https://huggingface.co/monajm36/ohca-classifier-v3-trained}, | |
| note={High-sensitivity BERT classifier for out-of-hospital cardiac arrest detection in discharge notes} | |
| } | |
| ``` | |
| ## License | |
| This model is released under the MIT License. See LICENSE file for details. | |
| ## Contact and Support | |
| - **Repository**: [GitHub - OHCA Classifier v3.0](https://github.com/monajm36/ohca-classifier-3.0) | |
| - **Issues**: Please report issues on the GitHub repository | |
| - **Model Card**: This model card follows the framework proposed by Mitchell et al. (2019) | |
| ## Acknowledgments | |
| - **Base Model**: Microsoft Research for PubMedBERT | |
| - **Dataset**: MIMIC-III for training data foundation | |
| - **Framework**: Hugging Face Transformers library | |
| - **Medical Domain**: Clinical expertise in cardiac arrest detection | |
| - **Methodology**: Data science community for best practices in medical ML | |
| This model is intended for research and clinical decision support. Always consult with medical professionals for patient care decisions. |