Lightweight large language models for early sepsis prediction via a semantic abstraction rule engine

Sepsis remains a major global health challenge due to its high mortality rate and the difficulty of predicting its onset from nonspecific early symptoms. To address this, we explored the feasibility of using fine-tuned small-scale large language models (LLMs) to perform accurate sepsis prediction from structured clinical time-series data. We developed a Semantic Abstraction Rule Engine (SARE) that transforms numerical clinical data into clinician-readable textual descriptions through temporal feature extraction and semantic transformation. Using these representations, four lightweight LLMs were fine-tuned with Low-Rank Adaptation and evaluated on publicly available datasets. The fine-tuned models consistently outperformed traditional machine learning and temporal prediction methods, with notable improvements in sensitivity and area under the curve. The Gemma-2-9B model achieved an AUC of 0.9307 and a sensitivity of 0.8532 on the PhysioNet/Computing in Cardiology Challenge 2019 dataset, surpassing existing state-of-the-art approaches. These findings demonstrate that lightweight LLMs can provide accurate and computationally efficient predictions, offering practical support for early sepsis detection in resource-limited healthcare environments.