PromptMind Team at MEDIQA-CORR 2024: Improving Clinical Text Correction with Error Categorization and LLM Ensembles

Satya Gundabathula, Sriram KolarNAACL 2024

TL;DR

This paper introduces a novel prompt engineering strategy for Large Language Models (LLMs) to improve the detection and correction of errors in clinical notes. By incorporating error categorization and a chain-of-thought approach within a single prompt, the authors achieve significant performance gains, demonstrating the potential of LLMs for enhancing clinical documentation accuracy.

What Problem Does This Paper Address?

Clinical notes are crucial for patient care, but they often contain errors that can have serious consequences. Manually reviewing these notes is time-consuming and prone to human error. While Natural Language Processing (NLP) offers potential solutions, the inherent complexity of medical jargon, abbreviations, and context-specific nuances makes accurate error detection and correction a significant challenge. Existing LLMs, trained on vast internet data, possess general knowledge but may struggle with the specialized reasoning and accuracy required in the healthcare domain, especially when dealing with potentially unreliable training data. The MEDIQA-CORR shared task specifically aims to address this by focusing on three subtasks: detecting errors, identifying the erroneous sentence, and correcting it.

Methods

The core of the proposed approach lies in a sophisticated prompt engineering strategy for LLMs. Instead of treating the three subtasks (error detection, span identification, and correction) separately, the authors tackle them concurrently within a single prompt. This is achieved through a "chain-of-thought" approach, where the LLM is guided to reason through the problem step-by-step.

Key methodological innovations include:

  • Error Categorization: Before prompting, the authors perform an error analysis on the training data to identify common error types. This information is then used to create a more structured and informative prompt, guiding the LLM to classify errors into predefined categories (e.g., Medications, Medical Conditions, Clinical Plans). This categorization acts as an intermediate reasoning step, enhancing the LLM's understanding.
  • Single-Prompt, In-Context Learning (ICL): The entire task is framed within a single prompt, leveraging few-shot learning with carefully selected examples. This approach aims for efficiency and ease of adoption, avoiding the need for multiple specialized models.
  • Self-Consistency and Ensemble: To further enhance robustness and accuracy, the authors employ two techniques:
  • Self-Consistency: For a given input, the LLM generates multiple outputs, and the majority prediction is used. This is particularly applied to GPT-4.
  • Ensemble: Predictions from different LLMs (GPT-4 and Claude-3 Opus) are combined to produce a final, more reliable output.
  • Prompt Template Design: A detailed prompt template is provided, guiding the LLM through identifying errors, categorizing them, providing a reason, and generating a corrected sentence. This template is crucial for directing the LLM's reasoning process.

Datasets

The primary dataset used is from the MEDIQA-CORR shared task. The training data consists of clinical texts from the MS (Medical School) collection, while the validation and test data include both MS and UW (University of Washington) collections. The paper notes that the dataset composition is detailed in Table 1. No new benchmarks were introduced; the work focuses on applying LLMs to the existing MEDIQA-CORR dataset.

Key Insights and Ideas

  • Error Categorization as a Chain-of-Thought Enabler: The most significant insight is that explicitly categorizing errors within the prompt acts as a powerful intermediate reasoning step for LLMs. This "intermediate chain-of-thought" significantly boosts performance, suggesting that structured guidance is more effective than simply asking for error correction.
  • Single-Prompt Efficiency: Tackling all three subtasks (detection, identification, correction) within a single prompt is not only efficient in terms of computational cost and development effort but also leads to a more holistic evaluation of the LLM's capabilities.
  • Self-Consistency and Ensembling are Crucial for Robustness: In the high-stakes domain of clinical notes, where hallucinations can be detrimental, self-consistency and ensemble methods are vital for improving the reliability and accuracy of LLM outputs.
  • LLM Capabilities are Task-Dependent: The paper highlights that while LLMs have impressive general knowledge, their performance on specialized tasks like clinical error correction is highly dependent on prompt design and the LLM's ability to follow complex instructions.

Key Results

The paper reports significant performance improvements due to the proposed methods:

  • Error Categorization Boost: Integrating error categorization into the prompt led to performance improvements of nearly 10% for error detection (Task-1) and 16% for text span identification (Task-2) when compared to prompts without categorization.
  • GPT-4 Outperforms GPT-3.5: The advanced GPT-4 model demonstrated superior performance over GPT-3.5, confirming its enhanced reasoning capabilities for this task.
  • Ensemble Performance: The ensemble of GPT-4 and Claude-3 Opus achieved strong results. Specifically, the best subtask-3 score (correction) was ranked 2nd among all participants in the competition, and the best subtask-2 performance (span identification) was among the top 3.
  • Specific Metrics: Table 3 shows that Claude-3 Opus surpassed GPT-4 in error detection and span identification, while GPT-4 was more verbose in error correction. Table 4 indicates that using a majority ratio of 3/4 for GPT-4 in the ensemble yielded the best Task-1 and Task-2 results, while a 4/4 majority ratio achieved the best Task-3 aggregate score.

Limitations and Future Work

  • Limitations:
    • General Data vs. Medical Knowledge: LLMs trained on general data may lack specific medical knowledge, leading to misinterpretations and inaccurate corrections.
  • Hallucinations and Inconsistencies: The risk of LLMs generating factually incorrect or inconsistent outputs remains a concern.
  • Prompt Engineering Dependency: The effectiveness of the approach heavily relies on prompt engineering expertise, which may not be easily generalizable.
  • Black Box Nature: The inherent "black box" nature of LLMs poses challenges for explainability and building trust in clinical settings.
  • Future Work:
    • Continuous Improvement: Ongoing efforts are needed to improve training data, develop more robust evaluation metrics, and incorporate human oversight.
  • Exploring Full Potential: Further research is required to fully understand the capabilities and limitations of LLMs in healthcare to ensure their safe and responsible application.
  • Ethical Considerations: Addressing potential biases in training data, ensuring transparency in prompt design, and implementing robust data security and de-identification practices are crucial for ethical deployment. The authors emphasize that LLMs should augment, not replace, healthcare professionals.

Summary updated

BibTeX
@inproceedings{gundabathula2024promptmind,
  author = {Satya Gundabathula and Sriram Kolar},
  title = {{PromptMind Team at MEDIQA-CORR 2024: Improving Clinical Text Correction with Error Categorization and LLM Ensembles}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure