Exploring TopicGPT for Identification of Themes of Emerging Pathogens in Biosurveillance Data.
Characterization of emerging diseases often has to occur when analysts who are in charge of tracking emerging threats have limited time. Topic modeling approaches that require limited analyst input and can be applied on electronic health records (EHR) may provide earlier insights. We applied large language model (LLM) based topic modeling to text from emerging pathogens to discover themes.
EHR notes were extracted from Veterans Affairs (VA) Corporate Data Warehouse. We included up to 200 notes from 5 cohorts including Nationally Notifiable Disease (2005-2007), early COVID-19 cases, and outbreak-related notes from Zika, leptospirosis, COVID-19, and mpox. Notes were sampled from early and late outbreak periods, with a separate cohort that was matched to the early period. Clinical sections were identified using medspaCy to extract tokens from the History of Present Illness (HPI). Topic modeling was conducted using TopicGPT (Mistral-7B-Instruct-0.3), which was modified to perform inferences on local infrastructure to protect sensitive data.
A total of 1,841 clinical documents were processed for this analysis on an NVIDIA A40 GPU. The most frequent topic category surfaced by TopicGPT in each cohort was "Infectious Disease." Expert review on a sample of topics demonstrated a 61.5% appropriate representation of topic and explanation based on the original HPI.
Automated topic modeling methods permit extraction and summary of crucial emerging themes. Still, they would benefit from optimization to provide timely analysis of high-volume documents of emerging diseases. VA data were used to evaluate these capabilities for early risk and exposure detection, but they could also be useful to recognize and monitor for early signal of disease among Service Members.
EHR notes were extracted from Veterans Affairs (VA) Corporate Data Warehouse. We included up to 200 notes from 5 cohorts including Nationally Notifiable Disease (2005-2007), early COVID-19 cases, and outbreak-related notes from Zika, leptospirosis, COVID-19, and mpox. Notes were sampled from early and late outbreak periods, with a separate cohort that was matched to the early period. Clinical sections were identified using medspaCy to extract tokens from the History of Present Illness (HPI). Topic modeling was conducted using TopicGPT (Mistral-7B-Instruct-0.3), which was modified to perform inferences on local infrastructure to protect sensitive data.
A total of 1,841 clinical documents were processed for this analysis on an NVIDIA A40 GPU. The most frequent topic category surfaced by TopicGPT in each cohort was "Infectious Disease." Expert review on a sample of topics demonstrated a 61.5% appropriate representation of topic and explanation based on the original HPI.
Automated topic modeling methods permit extraction and summary of crucial emerging themes. Still, they would benefit from optimization to provide timely analysis of high-volume documents of emerging diseases. VA data were used to evaluate these capabilities for early risk and exposure detection, but they could also be useful to recognize and monitor for early signal of disease among Service Members.
Authors
Callahan Callahan, Peterson Peterson, Kalvesmaki Kalvesmaki, Willson Willson, Dalton Dalton, Vuong Vuong, Gordon Gordon, Barker Barker, Ajinwo Ajinwo, Humpherys Humpherys, Stevens Stevens, Pugh Pugh, Jones Jones
View on Pubmed