AI detects exposure risks of avian flu in the Maryland emergency department
Researchers at the University of Maryland School of Medicine have developed a new and highly effective application of artificial intelligence (AI) tools, quickly scan notes in electronic medical records, and identify high-risk patients who may have been infected with H5N1 avian flu or “bird flu.” Clinical infections.
Using the Generated AI Lead Language Model (LLM), the research team analyzed 13,494 visits to the University of Maryland Health System (UMMS) Hospital emergency department from urban, suburban and rural adult patients in 2024. All of these patients have early membranitis, such as acute respiratory disease (cough, fever, congestion, etc.). The goal was to assess the extent to which high-risk patients could be produced that could have been overlooked during initial treatment.
Scanning all the memos in the emergency department, the model flagged 76 for mentioning high-risk exposure to avian flu, including working on farms with livestock like meat influ and cattle. These exposures were usually mentioned by chance. For example, documenting a patient's occupation as a butcher or farm worker is not due to clinical suspicion of avian influenza.
After a brief review by research staff, 14 patients were identified to have been exposed to animals that are recently associated with animals known to carry H5N1, including poultry, wild birds and livestock. Although these patients were not specifically tested for H5N1, no potential bird fluorescence infections were identified, the model worked to find “haystack needles” cases in thousands of patients treated for seasonal influenza and other routine respiratory diseases.
“This study shows how generative AI can bridge critical gaps in public health infrastructure by detecting patients at high risk otherwise unnoticed,” the corresponding author said. Catherine E. Goodman, J.D.Assistant Professor of Epidemiology and Public Health at UMSOM and University of Maryland Health Computing Institute (UM-IHC). “The biggest danger nationwide is that we don't know what we don't know because the number of symptomatic patients is not being able to track the exposure to potential avian flu and how many of those patients are being tested, so infections may not be detected.
Since early 2024, H5N1 has infected more than 1,075 dairy herds in 17 states, and over 175 million poultry and wild birds have tested positive for this outbreak. Identified human cases remain rare, with 70 confirmed cases and one fatality in the US by mid-2025. Centers for Disease Control and Prevention (CDC). However, there could be even more undetected infections due to a lack of extensive testing. Additionally, new strains can occur, allowing for air spread from human to human, leading to an increase in cases and a potential epidemic.
“The AI review requires only 26 minutes of human time and costs just 3 cents per patient memo, indicating high scalability and efficiency,” said the Study co-author. Anthony Harris, MD, MPHProfessor and Representative Chair of Epidemiology and Public Health at UMSOM. “This method could create a national network of clinical sentinel sites for new infectious disease surveillance to better monitor emerging epidemics.”
LLM (GPT-4 turbo) showed strong performance in identifying mentions of animal exposure. This showed a positive and 98% predicted positive and 90% negative predicted positive values before bird fluids circulate in US livestock in a sample of 10,000 historic emergency department visits from 2022 to 2023. However, this model was conservative in identifying exposures that are particularly relevant to maintaining the need for human review of flagged cases to patients who flag patients with low-risk animal contact, such as dog exposure.
As the risk of infections transmitted by animals increases, researchers suggest that large-scale language models can be used proactively to alert healthcare providers in real time. This could encourage people to be more vigilant about potential exposure to infected animals, targeted testing, and asking about infection control by isolating patients. The CDC currently relies on mandatory lab reports to track avian flu, but there is no system to assess whether clinicians are questioning or documenting relevant exposures in symptomatic patients.
Researchers would then want to test a large-scale linguistic model for future monitoring and deployment within electronic health records to speed up real-time identification of high-risk patients. As respiratory virus season returns in autumn, it is especially important to have a quick and accurate method to identify patients who need special testing for avian flu, or preventive isolation while receiving treatment.
“We are at the forefront of a disruptive yet incredibly promising revolution on big data and artificial intelligence,” Umsom Dean said. Mark T. Gladwin, MDHe is also a renowned vice president of medical care at the University of Maryland, Baltimore (UMB), John Z. and Akiko K. Bauers. “The engineers and physician researchers working at the Institute for Health Computing are: The 2 million patients we serve across Maryland and this study demonstrate will be able to use AI and big data to identify early signals of new infectious diseases, such as avian flu, to test these diseases faster and prevent them from spreading. ”
Includes co-authors of other UMSOM faculty members on paper Dr. Lawrence S. MagderProfessor of Epidemiology and Public Health at Umsom, Dr. Jonathan D. Baghdadi, MDAssociate Professor of Epidemiology and Public Health at UMSOM is also a faculty member at UM-IHC. Daniel J. Morgan, MD, MSProfessor of Epidemiology and Public Health at UMSOM.
This study would not have been possible without the contribution of the UM Health Computing Institute, which was founded two years ago in North Bethesda, Maryland. University of Maryland, College Park, University of Maryland, Baltimore, and University of Maryland Medical System. Laboratory integrates tHe studied calculation expertise, clinical expertise, biomedical innovation, health data and academic resources for the three institutions.
“As an academic health system, we are responsible for providing today's care and preparing for tomorrow's treatments, and have been a national leader in data promoting medical research and patient care for a long time.” MBA, MOHAN SUNTHA, MBAPresident and CEO of the University of Maryland Health System. “We recognize that the value of data across the system represents the diversity of the communities we can serve.”
Funding for the research was provided by the Federal Institute for Medical Research and Quality. Computing and data storage costs for LLM analytics were supported by the UM Health Computing Institute.
