|
|
|||||
|
|
|||||
|
AI Has
a Statistics Problem |
|||||
|
Note: This piece was originally published by the University of North Carolina's Gillings School of Global Public Health on September 1, 2026 in their online SPH News and is reprinted here with their full permission.
Artificial intelligence (AI) has gotten pretty good at recognizing patterns, making predictions and generating convincing answers from massive amounts of data in record time. But being able to find a pattern doesn’t always mean that the pattern is correct or should be acted upon. That distinction matters much more when AI is being used to make decisions about people’s health. AI can be confidently wrong. It can learn biases unintentionally embedded in its training data. A model that works well with one population may not work as well with another. And a prediction can show that extreme heat is linked with higher rates of illness without identifying which interventions will protect the people most at risk. These are not new problems. They are problems that biostatisticians have spent decades working to solve. In the Department of Biostatistics at the UNC-Chapel Hill Gillings School of Global Public Health, faculty are applying that expertise to AI and helping make these tools more reliable. The work builds on a foundation that dates back to the department’s founding in 1949: developing statistical methods in direct contact with real health questions, not in isolation from them.
“The Gillings School has many people working on many facets of AI and health, from solving under-the-hood technical challenges all the way to evaluating and implementing AI tools in practice, both locally and globally — and everything in between,” said Michael Kosorok, PhD, W.R. Kenan, Jr. Distinguished Professor of biostatistics and director of the UNC Gillings Center for AI and Public Health (CAIPH). “It is an exciting time for public health!” Within CAIPH, researchers develop and apply AI tools to address global health challenges by bringing together expertise in health and AI from across UNC. Looking beyond the pattern For biostatisticians, figuring out whether an AI model can be trusted starts with looking at the data behind it. Where did the data come from? Who is represented and who is missing? What influenced the decisions captured in the data? How was the outcome measured?
Hongtu Zhu, PhD, Kenan Distinguished Professor in the Department of Biostatistics and editor of Journal of the American Statistical Association (JASA): Applications and Case Studies, says AI failures often come down to “the wrong data, the wrong comparison or unwarranted certainty.” Consider an AI tool designed to help communities prepare for extreme heat. Historical data could show that one neighborhood has fewer heat-related emergency room visits than another, leading the model to identify residents there as being at lower risk. But what if residents in that neighborhood have less access to health care and are simply less likely to visit an emergency room? Without that context, the model could underestimate the community’s risk and influence where limited public health resources are sent. A biostatistical approach would account for factors that could be affecting what appears in the data before using that pattern to draw a conclusion. The same principle applies to cause and effect. More data and computing power cannot turn a flawed comparison into evidence that one thing directly caused another. That becomes especially important when AI moves from predicting what might happen to informing what should happen next. How sure are you? If an AI chatbot confidently gives you a bad dinner recommendation, the stakes are pretty low. But imagine AI helping predict where the next disease outbreak could occur, which communities are most at risk during a heat wave or where resources are needed most following a natural disaster. When the stakes are higher, “How sure are you?” becomes a question that we need AI to be able to answer honestly and confidently. “Making machines honest about what they do not know is one of the great statistical problems of this decade,” Zhu said. For public health applications, an AI system needs to do more than produce an answer. Researchers also need to understand how certain that answer is and when there isn’t enough evidence to rely on it. Biostatistics also helps researchers evaluate whether an AI tool actually works. That includes testing it across different populations and determining whether using it improves health outcomes, rather than relying only on how well the model performs in a controlled test. Building what comes next As biostatisticians work to improve AI, AI is also changing biostatistics. Routine analyses, data cleaning and coding can increasingly be automated. That gives biostatisticians more opportunity to focus on work that requires human judgment: asking the right questions, determining whether available data can answer them, figuring out whether one factor actually caused an outcome and testing whether AI tools work across different populations. It also presents a choice for the field. “If we do not evolve, we become auditors of other people’s models instead of architects of better ones,” Zhu said. At Gillings, that evolution is already affecting research and training. Students are studying traditional statistical methods alongside machine learning, causal inference and computing, and working on AI-for-health projects involving imaging genetics, electronic health records, causal machine learning and large language model evaluation For Gillings biostatisticians, the goal isn’t to compete with AI or simply check its work. It’s to be involved in how these tools are developed and evaluated from the beginning, bringing statistical rigor to technology that is increasingly being used across public health. ■ |
|||||
|
|
|||||
