DTKF: a sepsis early prediction framework based on large language model semantic augmentation and dual-teacher knowledge distillation

Abstract

Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection. Early and accurate prediction is crucial for reducing mortality. However, clinical settings often face challenges such as data scarcity, class imbalance, and the underutilization of unstructured clinical notes. Existing scoring systems and traditional machine learning models primarily rely on structured vital signs, failing to capture the rich semantic risk factors embedded in textual data. Although Large Language Models (LLMs) excel in text understanding, their high computational costs and potential hallucinations hinder direct deployment in resource-constrained emergency settings. To address these limitations, this study proposes a novel Dual-Teacher Knowledge Fusion (DTKF) framework. First, we introduce a semantic feature augmentation module that leverages DeepSeek to extract semantic risk probabilities from clinical notes and explicitly fuses them with structured vital signs to construct high-order hybrid features for the student model. Second, to prevent overfitting under small-sample conditions, we design a dual-source knowledge distillation strategy. This strategy integrates rule-based knowledge and data-driven semantic knowledge to construct hybrid teacher signals, regulating a lightweight logistic regression student model via soft-label supervision. Experiments on the MIMIC-IV dataset demonstrate that DTKF achieves an AUC of 0.796. This result not only achieves a higher AUC than traditional unimodal baselines but also improves performance by 6.6% compared to direct prediction using LLMs. This study effectively integrates rule-based and data-driven knowledge, enhances feature representation using textual information, and improves model performance and robustness through distillation techniques, providing a clinically valuable solution for sepsis early warning.