End-to-end Chinese clinical event extraction based on large language model

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence in Healthcare

Background:

• Clinical event extraction is vital for organizing medical data and powering healthcare services.
• Pipeline-based methods for event extraction face challenges like error propagation and information loss.
• Existing methods often struggle with suboptimal performance in complex clinical text.

Purpose of the Study:

• To propose an advanced, end-to-end clinical event extraction method utilizing large language models (LLMs).
• To address the limitations of traditional pipeline approaches in clinical event extraction.
• To enhance the accuracy and efficiency of structuring medical data for improved clinical decision-making.

Main Methods:

• Developed LMCEE, an end-to-end method transforming clinical event extraction into a text generation task.
• Employed a prompt learning strategy tailored for LLMs to perform clinical event extraction.
• Evaluated the method's performance against traditional pipeline and existing generative-based approaches.

Main Results:

• LMCEE achieved a significant 12% increase in F1 score compared to traditional pipeline methods.
• The proposed method outperformed the UIE generative-based method by 5.7% in F1 score.
• Identified limitations including sensitivity to prompt templates and LLM type dependency.

Conclusions:

• End-to-end LLM-based methods offer superior performance for clinical event extraction.
• LMCEE demonstrates a promising advancement in structuring clinical data.
• Further research is needed to optimize prompt templates and LLM selection for enhanced robustness.