UMD researchers use GenAI to scan ER notes, spotting hidden bird flu exposure risks faster and more efficiently than traditional methods.
The University of Maryland School of Medicine (UMD) has unveiled a novel use of GenAI to detect hidden bird flu risks in emergency departments. Traditional surveillance often misses exposure clues buried in clinical notes, but this system uses GPT-4 Turbo to scan thousands of records quickly.
In a study of 13,494 ER visits, the model flagged 76 cases where patient notes mentioned high-risk exposures such as farm or butcher work. Human reviewers confirmed 14 patients with relevant recent exposure to livestock and poultry, none of whom had been tested for H5N1. The AI uncovered risks that could have slipped through routine care, highlighting its role in bridging gaps in disease monitoring.
The process proved both scalable and affordable, requiring only 26 minutes of human oversight and costing just three cents per note. Performance metrics were strong, with a 90% positive predictive value and 98% negative predictive value. While some flagged cases involved low-risk contact, human review ensured accuracy. The team believes this approach could serve as a nationwide sentinel system for future outbreaks.
With H5N1 spreading in U.S. livestock and poultry, GenAI offers a way to strengthen preparedness. By surfacing hidden risks in real time, clinicians can test patients sooner, isolate potential cases, and prevent wider spread. Researchers now plan to integrate this tool directly into electronic health records for prospective surveillance, creating a frontline defense against emerging epidemics.
