A sepsis diagnosis method based on Chain-of-Thought reasoning using Large Language Models

Abstract

Sepsis is a severe infectious disease with high incidence and mortality rates globally. Early diagnosis of sepsis is crucial for improving patient outcomes. Previous diagnostic methods heavily relied on subjective clinical experience, while the machine learning-based methods can only learn knowledge from a specific dataset. Recently, the rapid development of Large Language Models (LLMs) has significantly enhanced various downstream dialogue tasks by leveraging prior semantic knowledge. Therefore, it is of great interest to explore the potential of LLMs in sepsis diagnosis. This study proposed an early sepsis diagnosis method based on the Chain of Thought (CoT) reasoning using LLMs. First, the clinical data of a patients were transformed into a textual representation to form the prompt. Subsequently, a CoT was created to simulate the reasoning process of human medical experts and utilized the prior semantic knowledge in LLMs to achieve sepsis diagnosis. The proposed method was validated using real clinical data, demonstrating high classification performance with an accuracy of 0.87, recall of 0.98, and F1 score of 0.88. These metrics showed an improvement in F1 score by 7 to 8 percentage points compared to commonly used machine learning classifiers. The experimental results indicated that the proposed method can enhance the performance of early sepsis diagnosis, and the introduction of CoT enhanced the interpretability of diagnostic results, contributing to the application of LLMs in clinical diagnosis.