[Paper Review] MEDEC: A Benchmark for Medical Error Detection and Correction in Clinical Notes

Asma Ben Abacha, Wen-wai Yim+5 more|arXiv (Cornell University)|Dec 26, 2024

Electronic Health Records Systems8 citations

TL;DR

MEDEC is the first publicly available benchmark for detecting and correcting medical errors in clinical notes, evaluating LLMs and medical doctors on error detection, localization, and correction across five error types.

ABSTRACT

Several studies showed that Large Language Models (LLMs) can answer medical questions correctly, even outperforming the average human score in some medical exams. However, to our knowledge, no study has been conducted to assess the ability of language models to validate existing or generated medical text for correctness and consistency. In this paper, we introduce MEDEC (https://github.com/abachaa/MEDEC), the first publicly available benchmark for medical error detection and correction in clinical notes, covering five types of errors (Diagnosis, Management, Treatment, Pharmacotherapy, and Causal Organism). MEDEC consists of 3,848 clinical texts, including 488 clinical notes from three US hospital systems that were not previously seen by any LLM. The dataset has been used for the MEDIQA-CORR shared task to evaluate seventeen participating systems [Ben Abacha et al., 2024]. In this paper, we describe the data creation methods and we evaluate recent LLMs (e.g., o1-preview, GPT-4, Claude 3.5 Sonnet, and Gemini 2.0 Flash) for the tasks of detecting and correcting medical errors requiring both medical knowledge and reasoning capabilities. We also conducted a comparative study where two medical doctors performed the same task on the MEDEC test set. The results showed that MEDEC is a sufficiently challenging benchmark to assess the ability of models to validate existing or generated notes and to correct medical errors. We also found that although recent LLMs have a good performance in error detection and correction, they are still outperformed by medical doctors in these tasks. We discuss the potential factors behind this gap, the insights from our experiments, the limitations of current evaluation metrics, and share potential pointers for future research.

Motivation & Objective

• Motivate rigorous validation of medical text generated or assisted by LLMs to prevent hallucinations and errors in clinical notes.
• Create a large, diverse dataset of clinical notes with injected errors spanning five error types.
• Provide tasks and evaluation protocols for detecting, localizing, and correcting medical errors in notes.
• Benchmark contemporary LLMs and medical experts to quantify gaps and guide future improvements.

Proposed method

• Construct MEDEC with 3,848 clinical texts from two sources and three US hospital systems, covering five error types (Diagnosis, Management, Treatment, Pharmacotherapy, Causal Organism).
• Two error-injection pipelines create correct and erroneous notes; MS (M#1) uses medical board-exam narratives, while UW (M#2) injects errors into real UW notes with de-identification.
• Define three subtasks: A) error flag prediction, B) error sentence extraction, C) corrected sentence generation.
• Evaluate LLMs and two prompts (P#1, P#2) for the three subtasks and compute error-correction metrics (ROUGE-1, BLEURT, BERTScore, AggregateScore).
• Compare LLM performance against two physicians on the same MEDEC test set to quantify gaps between models and expert clinicians.

Experimental results

Research questions

• RQ1 Can LLMs reliably detect whether a clinical note contains an error?
• RQ2 Can LLMs identify the exact sentence containing the error in a note?
• RQ3 Can LLMs generate corrected sentences that align with medical expert corrections?
• RQ4 How do LLMs compare to medical doctors in error detection, localization, and correction across five error types?

Key findings

• Claude 3.5 Sonnet achieves the highest error-flag detection accuracy among LLMs (70.16%) and error-sentence detection accuracy (65.62%).
• o1-mini is the second-best for error-flag detection (69.08%).
• In error correction, the best Aggregate Score is achieved by o1-preview (0.698), ahead of GPT-4o variants (≈0.639).
• Medical doctors outperform LLMs overall on error-flag and sentence-detection accuracy, with Doctor #2 reaching 0.7161 (error flag) and 0.6677 (error sentence) in the MS/UW split, and higher correction scores.
• There is a ranking mismatch between detection accuracy and correction quality across models, highlighting evaluation challenges and the need for improved metrics in clinical text correction.
• MEDEC surface shows model performance gaps likely due to limited exposure to rare medical error types in pretraining data and the complexity of editing existing notes rather than drafting from scratch.