LLM-Powered Automated Infection Detection in Cirrhosis: Achieving Expert-Level Accuracy from Clinical Documents.

Area of Science:

• Hepatology
• Infectious Diseases
• Artificial Intelligence in Medicine

Background:

• Systemic infections, especially bacterial, are a major cause of complications and death in cirrhosis patients.
• Current methods for infection identification (ICD codes, manual review) are inefficient and unreliable.
• Accurate infection labeling is vital for clinical decision support and predictive modeling in cirrhosis.

Purpose of the Study:

• To evaluate the efficacy of a large language model (LLM) in automating the identification of infections in hospitalized cirrhosis patients.
• To compare LLM performance against manual physician chart review for infection detection.

Main Methods:

• A state-of-the-art LLM (Claude 3.5 Sonnet) was deployed on a HIPAA-compliant AWS instance.
• The LLM analyzed the first 48 hours of clinical notes from 196 hospitalized cirrhosis patients.
• Gold-standard infection labels were established via manual physician chart reviews.

Main Results:

• The LLM achieved high performance metrics: 89.8% accuracy, 97.9% sensitivity, and 89.1% positive predictive value (PPV).
• The model successfully identified diverse pathogen types and infection locations in complex cases.
• LLM performance was comparable to manual physician chart review.

Conclusions:

• LLMs show significant potential for accurate, scalable, and efficient automated infection identification and classification in cirrhosis.
• This AI-driven approach can enhance infection management strategies for hospitalized patients with cirrhosis.
• LLMs offer a promising advancement over traditional, labor-intensive methods for infection detection.