OBJECTIVE: To improve in-hospital mortality prediction from longitudinal electronic health records (EHRs) by extracting mortality-related high-risk clinical information and integrating heterogeneous clinical evidence. METHODS: We propose HiRMD, an LLM-augmented mortality prediction system that combines high-risk information extraction, structured medical diagnosis, and cross-attention-based risk prediction. HiRMD first refines patients' disease, medication, and procedure records to retain mortality-relevant clinical events. It then guides LLMs with predefined clinical criteria to generate structured diagnostic sequences reflecting high-risk disease progression and treatment status. For final prediction, longitudinal visit records are encoded by a Bi-GRU, while ICU severity scores and LLM-derived diagnostic outputs are used as clinical priors to guide temporal representation learning through multi-head cross-attention. RESULTS: Experiments on three public EHR datasets, including MIMIC-III, MIMIC-IV, and eICU, demonstrate the effectiveness of HiRMD compared with machine learning, deep learning, LLM-based, and ICU scoring baselines. HiRMD obtains the best AUROC and F1 scores on all three datasets. Additional analyses further support the effectiveness, stability, and clinical relevance of the proposed high-risk diagnostic representation. CONCLUSION: HiRMD effectively integrates LLM-based high-risk information extraction with longitudinal EHR modeling for accurate and interpretable in-hospital mortality prediction. SIGNIFICANCE: This study demonstrates the potential of LLM-guided diagnostic semantics to better identify high-risk patients, enhance prediction traceability, and support clinically meaningful decision-making in high-risk healthcare scenarios.
Meng et al. (Thu,) studied this question.