Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics

A team of researchers (Weizhi Nie, Zhen Qu, Weijie Wang, Chunpei Li, Ke Lu, Bingyang Zhou, Hongzhi Yu) has developed a novel Large Language Model (LLM)-guided framework for early sepsis warning that explicitly models physiological deterioration trajectories before disease onset. Published on arXiv (2604.20924), the system addresses a critical flaw in existing AI-based sepsis predictors: their black-box nature limits clinician trust and adoption. The framework uses a spatiotemporal feature extraction module to capture dynamic dependencies among multivariate vital signs, a Medical Prompt-as-Prefix module that embeds clinical reasoning cues into LLMs, and an agent-based post-processing component to keep predictions within physiologically plausible ranges.

Evaluated on the MIMIC-IV and eICU critical care databases, the method achieves superior AUC scores ranging from 0.861 to 0.903 across prediction windows of 24 to 4 hours before sepsis onset. This outperforms conventional deep learning approaches and rule-based systems like qSOFA. More importantly, the model provides transparent, interpretable trajectories and risk trends that align with clinical judgment, potentially enabling earlier interventions and personalized decision-making in intensive care environments. The work represents a significant step toward making AI-driven clinical decision support both accurate and trustworthy for frontline clinicians.