OptimalMEE: Optimizing Large Language Models for Medical Event Extraction Through Fine-Tuning and Post-hoc Verification

Medical event extraction (MEE) aims to identify and extract medical events mentioned in clinical notes, serving as a fundamental task for many clinical applications. Traditional solutions demand significant labor and sophisticated model design. The emerging large language models (LLMs) are considered to be potential on MEE. However, such methods have been shown to encounter issues related to accuracy, interpretability, and generalizability. In this paper, we propose OptimalMEE to optimize LLMs for MEE through fine-tuning and post-hoc verification. We leverage the LLM paired with a parameter-efficient fine-tuning mechanism and develop a four-step post-hoc verification process aimed at refining and interpreting the extracted events. Experimental results on multi-center datasets illustrate the strength of the proposed OptimalMEE on accuracy, interpretability, and generalizability, with 0.902 and 0.809 in MicroF1.