Exploring TopicGPT for Identification of Themes of Emerging Pathogens in Biosurveillance Data

Area of Science:

• Computational epidemiology
• Health informatics
• Natural language processing

Background:

• Characterizing emerging diseases is challenging with limited analyst time.
• Electronic health records (EHR) offer potential for early disease insights.
• Large language models (LLMs) can facilitate automated topic modeling on EHR data.

Purpose of the Study:

• To apply LLM-based topic modeling to emerging pathogen data for theme discovery.
• To evaluate the utility of automated topic modeling for early risk detection in EHRs.

Main Methods:

• Extracted EHR notes from Veterans Affairs (VA) Corporate Data Warehouse.
• Utilized TopicGPT (Mistral-7B-Instruct-0.3) for topic modeling on clinical sections (History of Present Illness).
• Processed 1,841 documents from cohorts including notifiable diseases, COVID-19, Zika, leptospirosis, and mpox.

Main Results:

• The most frequent topic identified across cohorts was "Infectious Disease."
• Expert review showed 61.5% appropriate representation of topics and explanations.
• The method processed data on an NVIDIA A40 GPU.

Conclusions:

• Automated topic modeling can extract and summarize emerging disease themes from EHRs.
• Optimization is needed for timely analysis of high-volume emerging disease data.
• This approach can aid early risk detection and disease monitoring in various populations, including Service Members.