The University of Maryland School of Medicine team applied a generative Artificial Intelligence large language model (GPT-4 Turbo) to 13,494 adult emergency department visits across the University of Maryland Medical System in 2024. The selected visits were for acute respiratory illness or conjunctivitis, symptoms consistent with early H5N1 avian influenza. The model scanned free-text clinical notes to detect mentions of high-risk animal exposures such as work as a butcher or contact with poultry, wild birds, or livestock.
Across the dataset the model flagged 76 visits that mentioned potentially relevant exposures. After brief human review, 14 patients were confirmed to have recent, relevant animal exposure. None of those patients had been specifically tested for H5N1 at the time, so potential infections remain unconfirmed. When evaluated on a separate sample of 10,000 historical emergency visits from 2022 to 2023, the model showed a 90 percent positive predictive value and a 98 percent negative predictive value for identifying mentions of animal exposure. The review process required 26 minutes of human time and cost an estimated three cents per patient note. The researchers noted the model was conservative and sometimes flagged low-risk contacts, supporting the need for human verification of flagged records.
Authors and institutional leaders framed the method as a scalable sentinel surveillance tool to bolster national monitoring of emerging infectious diseases. The team recommends testing prospective deployment within electronic health records to provide real-time alerts that could prompt clinicians to ask about exposures, order targeted testing, or isolate patients. The study notes broader context for concern, reporting that since early 2024 more than 1,075 dairy herds in 17 states and over 175 million poultry and wild birds have tested positive for H5N1, with 70 confirmed human infections and one death in the U.S. by mid-2025 per the Centers for Disease Control and Prevention. Funding came from the Agency for Healthcare Research and Quality, with computing support from the University of Maryland Institute for Health Computing.
