Integrating Electronic Health Records and Large Language Models for Coarse-to-Fine Hybrid Disease Prediction

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Electronic Health Records (EHRs) contain high-dimensional, long-tailed diagnosis codes that challenge traditional data-driven disease prediction models.
• Large Language Models (LLMs) show promise for disease prediction due to their knowledge and reasoning, but face challenges in vast prediction spaces and domain adaptation.

Purpose of the Study:

• To develop and evaluate a coarse-to-fine hybrid disease prediction framework integrating EHR data and LLMs.
• To improve the effectiveness and interpretability of disease prediction models in clinical settings.

Main Methods:

• A hybrid framework combining an EHR-driven base prediction module for coarse predictions.
• Serialization of patient information using tailored prompts for fine-tuned LLM assessment of candidate diseases.
• An aggregation module to integrate outputs from both modules for refined disease predictions.

Main Results:

• The proposed framework consistently outperformed existing non-LLM and LLM-based approaches on the eICU and MIMIC-IV databases.
• Significant improvements were observed in various metrics including P@5, P@10, P@20, R@5, R@10, R@20, w-F1, and Jaccard index.
• Further analyses confirmed the effectiveness of individual components within the hybrid framework.

Conclusions:

• Integrating EHR-based predictions with LLM assessments in a coarse-to-fine manner offers an effective disease prediction strategy.
• The developed framework provides a practical solution for enhancing clinical decision support systems.
• This approach addresses limitations of traditional models and direct LLM applications in disease prediction.