AI Reveals Hidden Bird Flu Risks in Maryland ERs

University of Maryland School of Medicine

Researchers from the University of Maryland School of Medicine developed a new and highly effective application of an artificial intelligence (AI) tool to quickly scan notes in electronic medical records and identify high-risk patients who may have been infected with H5N1 avian influenza or "bird flu", according to new findings published in the journal Clinical Infectious Diseases .

Using a generative AI large language model (LLM), the research team analyzed 13,494 visits across University of Maryland Medical System (UMMS) hospital emergency departments from adult patients in urban, suburban, and rural areas in 2024. These patients all had acute respiratory illness (such as, cough, fever, congestion) or conjunctivitis—symptoms consistent with early H5N1 infections. The goal was to assess how well generative AI could find high-risk patients who may have been overlooked at the time of initial treatment.

Scanning all of the emergency department notes, the model flagged 76 because they mentioned a high-risk exposure for bird flu, such as working as a butcher or at a farm with livestock, like chickens or cows. Usually, these exposures were mentioned incidentally—for example, documenting a patient's occupation as a butcher or farmworker—and not because of clinical suspicion for bird flu.

After a brief review by research staff, 14 patients were confirmed to have had recent, relevant exposure to animals known to carry H5N1, including poultry, wild birds, and livestock. These patients were not tested specifically for H5N1, so their potential bird-flu infections were not confirmed, but the model worked to find those "needle in a haystack" cases among thousands of patients treated for seasonal flu and other routine respiratory illnesses.

"This study shows how generative AI can fill a critical gap in our public health infrastructure by detecting high-risk patients that would otherwise go unnoticed," said study corresponding author Katherine E. Goodman, PhD, JD, Assistant Professor of Epidemiology & Public Health at UMSOM and a faculty member of the University of Maryland Institute for Health Computing (UM-IHC) . "With H5N1 continuing to circulate in U.S. animals, our biggest danger nationwide is that we don't know what we don't know. Because we are not tracking how many symptomatic patients have potential bird flu exposures, and how many of those patients are being tested, infections could be going undetected. It's vital for healthcare systems to monitor for potential human exposure and to act quickly on that information."

Since early 2024, H5N1 has infected more than 1,075 dairy herds across 17 states, and over 175 million poultry and wild birds have tested positive during this outbreak period. Identified human cases remain rare, with 70 confirmed infections and just one fatality in the U.S. by mid-2025, according to the Centers for Disease Control and Prevention (CDC). There are, however, likely many more infections that have gone undetected due to a lack of widespread testing. In addition, new strains could arise enabling human-to-human airborne spread, which would lead to an uptick in cases and a potential epidemic.

"The AI review required only 26 minutes of human time and cost just 3 cents per patient note, demonstrating high scalability and efficiency," said study co-author Anthony Harris, MD, MPH, Professor and Acting Chair of Epidemiology & Public Health at UMSOM. "This method has the potential to create a national network of clinical sentinel sites for emerging infectious disease surveillance to help us better monitor newly emerging epidemics."

The LLM (GPT-4 Turbo) demonstrated strong performance in identifying mentions of animal exposure, with a 90% positive predictive value and a 98% negative predictive value when it was evaluated on a sample of 10,000 historical emergency department visits from 2022-2023, before bird flu was circulating in U.S. livestock. However, the model was conservative when identifying exposures specifically relevant to avian influenza—sometimes flagging patients with low-risk animal contact, such as exposure to dogs—underscoring the need for human review of any flagged cases.

As the risk of infections transmitted by animals grows, researchers suggest that large language models could also be used prospectively to alert healthcare providers in real time. This could prompt them to be more vigilant about asking about potential exposure to infected animals, targeted testing, and controlling infections by isolating patients. The CDC currently relies on mandated lab reporting to track avian influenza but lacks systems to assess whether clinicians are asking about or documenting relevant exposures in symptomatic patients.

The researchers hope to next test the large language model for prospective surveillance and deployment within the electronic health record, for faster real-time identification of high-risk patients. As respiratory virus season resumes in the fall, having a fast and accurate way to identify those patients needing special testing for bird flu, or precautionary isolation while receiving treatment, will be especially critical.

"We are at the forefront of a disruptive but incredibly promising revolution around big data and artificial intelligence," said UMSOM Dean Mark T. Gladwin, MD , who is also the Vice President for Medical Affairs, University of Maryland, Baltimore (UMB), and the John Z. and Akiko K. Bowers Distinguished Professor. "The engineer and physician researchers working at the Institute for Health Computing have secure access to medical records from the two million patients that we serve throughout Maryland, and as this study demonstrates, can use AI and big data to identify early signals of emerging infectious diseases like bird flu to enable us to take action sooner to test for these diseases and keep them from spreading."

Other UMSOM faculty co-authors on the paper include Laurence S. Magder, PhD, Professor of Epidemiology & Public Health at UMSOM, Jonathan D. Baghdadi, PhD, MD, Associate Professor of Epidemiology & Public Health at UMSOM who is also on faculty at the UM-IHC, and Daniel J. Morgan, MD, MS, Professor of Epidemiology & Public Health at UMSOM.

The study would not have been possible without the contributions of the UM Institute of Health Computing, which was established two years ago in North Bethesda, Maryland as a collaboration between the University of Maryland, College Park, the University of Maryland, Baltimore, and the University of Maryland Medical System. The Institute merges the computational expertise, clinical expertise, biomedical innovation, health data and academic resources of the three institutions.

"As an academic health system, we have the responsibility to prepare for the cures of tomorrow while delivering the care of today, and have long been a national leader in data driving medical research and patient care," said Mohan Suntha, MD, MBA , University of Maryland Medical System President and CEO. "We also recognize that the value of the data across our System is representative of the diversity of the communities that we are privileged to serve."

Funding for the research was provided by the federal Agency for Healthcare Research and Quality. Computing and data storage costs for LLM analyses were supported by the UM Institute for Health Computing.

/Public Release. This material from the originating organization/author(s) might be of the point-in-time nature, and edited for clarity, style and length. Mirage.News does not take institutional positions or sides, and all views, positions, and conclusions expressed herein are solely those of the author(s).View in full here.