Identifying Early Signals From Emerging Public Health Events Using Natural Language Processing

Timely detection of emerging public health threats is challenging because the surveillance infrastructure is not yet tuned to the emerging threat. We attempt to identify three nonspecific early signals that might be common across emerging events: public health authority communication, zoonotic exposure mentions, and other pathogen exposure mentions. Methods: Data from U.S. Department of Veterans Affairs emergency department visits between 2004 and 2024 were used to construct training and validation sets from reportable or emerging infectious diseases identified by historical diagnoses and laboratories. Not all early signal types were extracted using the same method. Rule-based and transformer models were used in a way to minimize developer and chart reviewer time. We then extracted cases from historic documents among selected diseases. Results: Positive predictive values for public health authority communication, zoonotic exposure, and other pathogen exposure ranged from 0.615 to 1.0. Target concepts were extracted from over 33 million emergency department visits. Distributions of extracted exposures generally matched expectations for the identified pathogen. Conclusion: Automated natural language processing methods allow surveillance scaling to large amounts of clinical documents to identify relevant cases. Initial validation compared to manual text review shows that accuracy is acceptable for initial feasibility exploration in biosurveillance efforts.