CLD-MEC at MEDIQA- CORR 2024 Task: GPT-4 Multi-Stage Clinical Chain of Thought Prompting for Medical Errors Detection and Correction

Renad M. Alzghoul, Abdulrahman Tabaza, Aya Abdelhaq, Ahmad AltamimiNAACL 2024

TL;DR

This paper introduces CLD-MEC, a novel framework leveraging GPT-4 with multi-stage "Chain of Thought" (CoT) prompting to identify and correct medical errors in clinical notes. The system aims to improve patient safety by automating the detection and correction of diagnostic and management errors, demonstrating promising results even with a zero-shot learning approach.

What Problem Does This Paper Address?

Ensuring patient safety in healthcare critically depends on accurately identifying and rectifying medical errors. These errors can occur across various stages of a patient's journey, from initial assessment to diagnosis and ongoing management. Clinical notes, while rich in patient history, are complex and varied in their structure and detail, making automated error detection a significant challenge. Existing methods often struggle with the nuances of clinical context and are prone to hallucinations. This paper tackles the MEDIQA-CORR 2024 shared task, aiming to build a robust system for detecting and correcting medical errors within these complex clinical narratives.

Methods

The CLD-MEC framework employs a multi-stage prompting pipeline powered by GPT-4. This pipeline is designed to handle three core tasks:

  1. Error Existence Classification: Determining if a medical error is present in a clinical note.
  2. Error Location Identification: Pinpointing the exact text span containing the error.
  3. Error Correction: Generating a natural language correction for the identified error.

Key to their approach are:

  • Clinical Chain of Thought (CoT): This technique guides GPT-4 to reason step-by-step through the clinical context, mimicking a clinician's thought process to arrive at a diagnosis and management plan.
  • Chain-of-Verification (CoVe): This method is used to mitigate hallucinations by cross-referencing generated reasoning with external knowledge, specifically clinical practice guidelines.
  • Multi-Stage Prompting: The pipeline is broken down into distinct stages, each with specific prompts designed to progressively refine the analysis, from initial preprocessing and clinical reasoning to verification and final error detection/correction.
  • Retrieval Augmented Generation (RAG) (Experimental): The authors also experimented with a RAG system, integrating clinical practice guidelines as an external knowledge source to enhance accuracy. However, this aspect of their approach did not yield the desired results within the scope of the competition.

Datasets

  • MEDIQA-CORR 2024 Datasets: The primary datasets used for training, validation, and testing were provided by the MEDIQA-CORR 2024 challenge. These datasets consist of clinical notes annotated for medical errors.
  • Training Dataset: Comprised of 2,189 clinical texts, each labeled as either containing an error (1) or not (0). It includes the original note, the erroneous sentence, the corrected sentence, and the fully corrected text.
  • Validation Datasets: Two sets were used: an MS validation set (574 texts) and a UW validation set (160 texts). These were unlabeled for validation purposes.
  • Testing Dataset: Contained 574 clinical texts, providing only the sentences without error flags or corrections, used for evaluating the final model performance.
  • Clinical Guidelines Dataset: A corpus of 47,000 clinical practice guidelines from 17 reputable online medical references was utilized for the RAG component.

Key Insights and Ideas

  • GPT-4's Reasoning Power: The study highlights GPT-4's impressive ability to perform complex clinical reasoning tasks, even in a zero-shot setting, by leveraging multi-stage CoT prompting.
  • Contextual Error Detection: The framework's success hinges on its ability to understand different levels of clinical context (L1 and L2) and map them to specific error scenarios, demonstrating that error types vary significantly based on the patient journey phase.
  • CoT and CoVe for Reliability: The combination of CoT for in-depth reasoning and CoVe for verification proved crucial in mitigating hallucinations and ensuring that the model's outputs were grounded in clinical logic.
  • RAG Challenges in Clinical Domain: The authors' experience with RAG suggests that while promising, integrating external knowledge effectively into complex clinical NLP tasks requires careful consideration of embedding models, chunking strategies, and the structure of the knowledge base itself. Their RAG system, despite using Bio_ClinicalBERT, did not improve performance, indicating areas for future research.

Key Results

The CLD-MEC system achieved acceptable performance on the MEDIQA-CORR 2024 tasks using a zero-shot learning approach.

  • Task 1 (Error Flag): Accuracy of 0.566.
  • Task 2 (Error Location): Accuracy of 0.49.
  • Task 3 (Error Correction): Performance metrics included:
  • ROUGE-1-F: 0.427
    • BERTScore: 0.48
    • BLEURT: 0.53
    • Aggregate Score (Mean of ROUGE-1-F, BERTScore, BLEURT): 0.48
    • Composite Score: 0.34

These metrics indicate the model's capability to produce contextually relevant corrections for identified clinical errors.

Limitations and Future Work

The authors acknowledge several limitations:

  • RAG System Performance: The RAG system did not perform as expected, failing to retrieve relevant queries and thus not improving the overall performance. This is attributed to time constraints, the choice of embedding model (Bio_ClinicalBERT), a naive chunking strategy, and the complex structure of the clinical guidelines dataset.
  • Limited LLM Experimentation: Due to hardware and financial constraints, the team could not experiment with other massive LLMs or more advanced fine-tuning techniques.
  • Algorithmic Scope: The approach, while effective for the three case scenarios, might not fully capture all possible medical error incidents.

Future work suggested includes:

  • Investigating the performance of public LLMs with more accessible hardware.
  • Exploring advanced chunking strategies and fine-tuning domain-specific models for retrieval.
  • Further research into optimizing RAG systems for clinical applications, potentially by using retrieval-oriented embedding models and more sophisticated knowledge graph structures.

Summary updated

BibTeX
@inproceedings{alzghoul2024cld,
  author = {Renad M. Alzghoul and Abdulrahman Tabaza and Aya Abdelhaq and Ahmad Altamimi},
  title = {{CLD-MEC at MEDIQA- CORR 2024 Task: GPT-4 Multi-Stage Clinical Chain of Thought Prompting for Medical Errors Detection and Correction}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure