I’m thrilled to sit down with Ivan Kairatov, a renowned biopharma expert with a wealth of experience in research, development, and the integration of cutting-edge technology in healthcare. Today, we’re diving into a groundbreaking application of generative AI in public health—specifically, its role in identifying high-risk avian influenza exposures through emergency department records. Our conversation explores the inspiration behind this innovative approach, the mechanics of how AI sifts through vast amounts of data to flag potential cases, the challenges of undetected infections, and the transformative potential of scalable AI tools in disease surveillance. Let’s get started.
What sparked the idea to harness generative AI for scanning emergency department notes to detect bird flu risks?
The motivation came from a clear need to bridge gaps in our public health surveillance systems. With avian influenza, or H5N1, circulating among animals in the U.S., we noticed that many potential human exposures were slipping through the cracks. Traditional methods rely heavily on clinicians recognizing and documenting specific risks, which doesn’t always happen in busy emergency settings. We saw generative AI as a way to systematically analyze unstructured data—like free-text notes in medical records—to uncover hidden patterns of risk, such as occupational exposures. It started as a hypothesis: could we teach an AI model to spot mentions of high-risk activities or environments that might indicate a bird flu exposure, even if the clinician didn’t explicitly suspect it?
How does the AI model actually work when it’s analyzing thousands of medical records for potential exposures?
The model, built on a large language model framework, is trained to parse through the often messy, narrative-style text in emergency department notes. It looks for specific keywords and contextual clues related to bird flu risk—think mentions of working on a farm, handling livestock, or being a butcher. Out of over 13,000 patient visits with respiratory or conjunctivitis symptoms, it flagged 76 cases where such high-risk exposures were mentioned, often just in passing as part of a patient’s background. The AI doesn’t diagnose; it’s more like a highly efficient filter that highlights cases for human review. It’s about narrowing down a huge haystack to find those critical needles.
Can you share more about the human review process that confirmed relevant animal exposures in 14 of those flagged patients?
Absolutely. After the AI flagged 76 cases, our team manually reviewed the notes to assess the relevance of the exposures to H5N1 risk. We looked for direct contact with animals known to carry the virus—poultry, wild birds, or livestock like dairy cows. Of those, 14 patients had clear, recent exposures that aligned with known transmission pathways. For instance, some were farmworkers handling chickens, while others had contact with potentially infected cattle. Our criteria were based on CDC guidelines and current outbreak data, ensuring we prioritized cases with the highest likelihood of risk for further follow-up.
Given that these patients weren’t tested for H5N1, how do you gauge the reliability of the AI in identifying true potential cases?
It’s a fair concern. We can’t say definitively that these patients had bird flu without testing, but the AI’s reliability comes from its high predictive values—90% positive and 98% negative in historical data validation. That means it’s very good at catching relevant mentions of exposure and rarely misses them. Still, it’s conservative, sometimes flagging low-risk contacts like pet dogs, which is why human review is critical. Moving forward, we’re working on integrating prospective surveillance to trigger real-time alerts for testing. The lack of testing isn’t just our challenge—it’s a nationwide issue, and we’re hoping this tool can push for more targeted diagnostics.
There’s a concern that many bird flu cases might be going undetected. Can you elaborate on why this is such a pressing issue?
The phrase “we don’t know what we don’t know” really captures the problem. H5N1 infections in humans are rare, with only a handful confirmed in the U.S., but without systematic screening or widespread testing, we’re likely missing cases. Many symptomatic patients might not disclose animal exposures unless asked directly, and clinicians, especially during flu season, may not think to probe for bird flu. If undetected cases are circulating, there’s a small but real risk of viral mutations enabling human-to-human spread. That’s the nightmare scenario—an outbreak we didn’t see coming because we weren’t looking hard enough.
The efficiency of this AI tool—taking just 26 minutes of human time and costing 3 cents per note—is striking. How do you envision this scalability impacting public health efforts?
The scalability is a game-changer. At such a low cost and time investment, we can imagine deploying this across diverse healthcare settings, from large urban hospitals to small rural clinics with limited resources. It could form the backbone of a national surveillance network for emerging infectious diseases, not just bird flu. Real-time monitoring of patient records could alert providers to ask the right questions or initiate testing sooner. We’re already exploring ways to embed this into electronic health record systems for seamless integration, ensuring that even understaffed facilities can benefit without overhauling their workflows.
Looking ahead, what’s your forecast for the role of AI in infectious disease surveillance over the next decade?
I’m incredibly optimistic. Over the next ten years, I see AI becoming a cornerstone of public health, not just for surveillance but for predictive modeling and rapid response. We’ll likely see AI tools evolve to anticipate outbreaks before they escalate, using data from medical records, social media, and environmental sensors. For infectious diseases like bird flu, AI could help us stay ahead of zoonotic threats by identifying transmission patterns in real time. The challenge will be ensuring equitable access to these technologies and addressing privacy concerns, but if we get that right, AI has the potential to transform how we prevent and manage epidemics on a global scale.
