HiRMD: A System for Mortality Prediction via LLM-Based High-Risk Information Extraction and Diagnosis

Area of Science:

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

Background:

• Accurate in-hospital mortality prediction is crucial for timely clinical intervention.
• Longitudinal Electronic Health Records (EHRs) contain rich patient data but present challenges in extracting high-risk information.
• Existing prediction models often lack interpretability and struggle to integrate heterogeneous clinical evidence.

Purpose of the Study:

• To develop and validate a novel system, HiRMD, for improved in-hospital mortality prediction.
• To enhance the extraction of mortality-related high-risk clinical information from EHRs.
• To integrate diverse clinical evidence for more accurate and interpretable risk assessment.

Main Methods:

• Proposed HiRMD, a system augmenting Large Language Models (LLMs) for mortality prediction.
• Refined patient records (diseases, medications, procedures) to identify mortality-relevant events.
• Utilized LLMs guided by clinical criteria to generate structured diagnostic sequences, combined with Bi-GRU encoding of visit records and ICU scores for prediction via multi-head cross-attention.

Main Results:

• HiRMD demonstrated superior performance across three public EHR datasets (MIMIC-III, MIMIC-IV, eICU) compared to various baseline models.
• Achieved the highest Area Under the Receiver Operating Characteristic (AUROC) and F1 scores on all tested datasets.
• Validated the effectiveness, stability, and clinical relevance of the LLM-derived high-risk diagnostic representation.

Conclusions:

• HiRMD effectively integrates LLM-based high-risk information extraction with longitudinal EHR modeling for accurate and interpretable in-hospital mortality prediction.
• LLM-guided diagnostic semantics show potential for improved high-risk patient identification, enhanced prediction traceability, and clinically meaningful decision-making.