| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| pretty_name: MedPIC-Bench |
| task_categories: |
| - question-answering |
| tags: |
| - medical |
| - medication-safety |
| - multiple-choice |
| - counterfactual-reasoning |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: questions.json |
| --- |
| |
| # MedPIC-Bench |
|
|
| MedPIC-Bench is an English multiple-choice benchmark for evaluating whether |
| large language models apply medication-safety rules to patient-specific |
| clinical contexts. Questions may have one or more correct options. |
|
|
| ## Dataset overview |
|
|
| The dataset contains 467 questions: |
|
|
| | Task family | Questions | |
| |---|---:| |
| | Guideline following | 284 | |
| | Counterfactual reasoning | 183 | |
| | **Total** | **467** | |
|
|
| ## Loading the dataset |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset("TIM0927/MedPIC-Bench", split="test") |
| ``` |
|
|
| ## Data fields |
|
|
| | Field | Type | Description | |
| |---|---|---| |
| | `instance_id` | string | Unique question identifier. | |
| | `patient_vignette` | string | Patient-specific clinical context. | |
| | `question` | string | Medication-safety question. | |
| | `options` | object | Mapping from option letters to answer text. | |
| | `answer` | list[string] | Correct option letter or letters. | |
| | `taxonomy_clinical_department` | string | Clinical department category. | |
| | `taxonomy_clinical_system` | string | Organ-system category. | |
| | `taxonomy_drug_categories` | list[string] | Drug classes represented in the question. | |
| | `taxonomy_patient_info_type` | string | Patient-information or condition type. | |
| | `taxonomy_population_source` | string | Population group represented by the question. | |
| | `taxonomy_reasoning_operation` | string | Required reasoning operation. | |
| | `benchmark_task_family` | string | `guideline_following` or `counterfactual`. | |
|
|
| Example: |
|
|
| ```json |
| { |
| "instance_id": "SS-HF-HFREF-001", |
| "patient_vignette": "A 71-year-old man has heart failure with reduced ejection fraction (LVEF 32%), currently asymptomatic on guideline-directed therapy review. No other significant past medical history; renal and hepatic function normal; not on other prescription medications; no known drug allergies.", |
| "question": "For this patient, which of the following drugs should be avoided?", |
| "options": { |
| "A": "Verapamil", |
| "B": "Dextromethorphan-quinidine", |
| "C": "Diltiazem", |
| "D": "Amlodipine", |
| "E": "Cilostazol" |
| }, |
| "answer": [ |
| "A", |
| "B", |
| "C", |
| "E" |
| ], |
| "taxonomy_clinical_department": "cardiology", |
| "taxonomy_clinical_system": "cardiovascular system", |
| "taxonomy_drug_categories": [ |
| "calcium-channel blockers", |
| "antiarrhythmics", |
| "pde3 inhibitors/peripheral vascular drugs" |
| ], |
| "taxonomy_patient_info_type": "disease presence", |
| "taxonomy_population_source": "older adults", |
| "taxonomy_reasoning_operation": "static risk identification", |
| "benchmark_task_family": "guideline_following" |
| } |
| ``` |
|
|
| ## Ethical statement and disclaimer |
|
|
| MedPIC-Bench is a research benchmark for evaluating model behavior. It is not |
| a clinical decision-support system, medical device, prescribing reference, or |
| source of medical advice. Benchmark questions and model outputs must not be |
| used to make decisions about diagnosis, treatment, medication selection, or |
| patient care. |
|
|
| The benchmark covers a limited set of medication-safety rules, clinical |
| contexts, and patient populations. It is not an exhaustive representation of |
| clinical practice. Performance on this dataset does not establish that a model |
| is medically accurate, safe, unbiased, or suitable for deployment. Model |
| outputs may contain clinically harmful errors even when aggregate benchmark |
| scores are high. |
|
|
| Users are responsible for appropriate expert review, risk assessment, and |
| compliance with applicable ethical, institutional, and legal requirements. |
|
|
| ## Acknowledgments |
|
|
| We gratefully acknowledge Yuting Long (龙宇婷) and Xinyao Ma (马馨瑶) for |
| their valuable support in dataset validation. Their careful review helped |
| improve the consistency and reliability of the benchmark. |
|
|