Causal-chain-aligned lightweight LLM for interpretable early sepsis warning

Sepsis remains a major global health threat with high mortality. Early warning is particularly challenging due to nonspecific clinical symptoms and the limited specificity of traditional, correlation-based diagnostic approaches. Existing machine learning methods often suffer from high missing-data rates and limited interpretability. To address these challenges, this study develops an interpretable and resource-efficient early sepsis warning framework by integrating Structural Causal Models (SCMs) that capture clinical causal chains with a lightweight fine-tuned Large Language Model (LLM). We propose a causal–semantic fusion framework that aligns clinical reasoning with underlying causal chains. Clinical time-series data are first transformed into natural-language descriptions through predefined semantic rules. Then a SCM is constructed using a causal discovery algorithm to capture temporal and causal dependencies among clinical variables. These resulting causal chains are incorporated into chain-of-thought prompts to guide a lightweight 7B-parameter LLM, which is fine-tuned using Low-Rank Adaptation (LoRA) for interpretable early-sepsis risk assessment. Across the PhysioNet/Computing in Cardiology Challenge 2019 (PhysioNet/CinC Challenge 2019) and MIMIC-IV datasets, the proposed model achieves AUCs of 0.8118 and 0.7780, respectively, surpassing both traditional machine-learning baselines and larger-parameter LLMs.Beyond predictive performance, the model demonstrates enhanced causal-chain clarity, higher data fidelity, and stronger alignment with expert clinical reasoning. By aligning LLM reasoning with data-driven causal chains, the proposed method delivers accurate, interpretable, and resource-efficient early sepsis warning. This causal-chain-aligned LLM framework offers a trustworthy and deployable solution for real-time clinical decision support in critical care settings.