Syndromic Analysis of Sepsis Cohorts Using Large Language Models

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Medicine
• Sepsis Research

Background:

• Extracting patient signs and symptoms from clinical notes is crucial for sepsis care but challenging for large-scale studies.
• Current methods struggle to efficiently process unstructured clinical text for symptom data.

Purpose of the Study:

• To evaluate the efficacy of large language models (LLMs) in extracting presenting signs and symptoms from patient admission notes.
• To analyze the association of extracted symptoms with infectious diagnoses, multidrug-resistant infections, and mortality in sepsis patients.

Main Methods:

• A retrospective cohort study involving over 100,000 adult patients across 5 hospitals.
• Utilized a large language model (LLaMA 3 8B) to extract up to 10 symptoms from admission notes.
• Validated LLM-extracted symptom labels against manual physician review and analyzed associations with outcomes using logistic regression.

Main Results:

• The LLM demonstrated high accuracy (99.3%) in extracting symptoms, with validated labels for 98.7% of patients.
• Extracted symptoms clustered into syndromes that correlated with specific infection sources and multidrug-resistant organisms (MRSA, MDRGN).
• Cardiopulmonary symptoms were linked to increased mortality (AOR 1.30), while skin/soft tissue symptoms increased MRSA risk (AOR 1.73).

Conclusions:

• LLMs can accurately extract patient signs and symptoms from admission notes, forming syndromes associated with infection types and patient outcomes.
• This approach enables large-scale analysis of symptom data, potentially improving sepsis management and antibiotic strategies.
• Further research is recommended to integrate this symptom data into predictive models for antibiotic choice and patient outcomes.