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

Background: Electronic health records (EHRs) have been widely adopted in recent years. However, diagnosis codes in EHRs are often high-dimensional, abstract, and follow a long-tailed distribution, limiting the effectiveness and interpretability of traditional data-driven models for disease prediction. Large language models (LLMs) offer a promising alternative due to their broad knowledge and strong reasoning capabilities, yet their direct application remains challenging due to the vast prediction space and limited domain adaptation. Methods: To address these challenges, we propose a coarse-to-fine hybrid disease prediction framework. First, an EHR-driven base prediction module is used to generate coarse predictions. Second, patient information is serialized using a tailored prompt template, and a fine-tuned LLM further assesses the presence of high-probability candidate diseases. Finally, an aggregation module integrates outputs from both modules to yield refined predictions. Results: The coarse-to-fine hybrid disease prediction framework consistently outperforms existing non-LLM and LLM-based approaches in disease prediction. Specifically, compared to the second-best baselines, improvements of 1.6% on P@5, 1.1% on P@10, 0.7% on P@20, 0.3% on R@5, 1.6% on R@10, 1.0% on R@20, 1.4% on w-F1, and 1.0% on Jaccard for the eICU Collaborative Research Database and improvements of 2.6% on P@5, 4.3% on P@10, 4.2% on P@20, 4.2% on R@5, 5.9% on R@10, 4.5% on R@20, 6.0% on w-F1, and 2.5% on Jaccard for the Medical Information Mart for Intensive Care IV are achieved, respectively. Further analyses demonstrate the effectiveness of each component of the framework. Conclusion: By integrating EHR-driven base prediction with LLM-based assessments in a coarse-to-fine manner, this study provides an effective and practical disease-prediction framework for clinical decision support.