AI is reshaping medicine, helping researchers and practitioners make better predictions, understand diseases in new ways, and personalize treatments. For instance, Google’s DeepMind won the 2024 Nobel Prize for creating AlphaFold, an AI system that advanced biomedical research and drug discovery by accurately predicting 3D protein structures from amino acid sequences, a previously slow and labor-intensive process.
Ben Lengerich, an assistant professor in the Department of Statistics at UW–Madison, is on the forefront of this field, exploring and advancing the power of AI to predict and explain health outcomes. His work bridges the gap between data-driven insights and medical interventions by building models designed to be transparent for doctors and researchers and to account for the specific health context of each patient. Lengerich is one of many faculty members across the School of Computer, Data & Information Sciences (CDIS) studying ways to harness AI for the public good; for instance, researchers in Computer Sciences and the Information School (iSchool) are examining how to improve the safety and security of AI technologies.
Lengerich, who also holds affiliate appointments in the Departments of Computer Sciences and Biostatistics & Medical Informatics (BMI), recently discussed his career path, his research on interpretable AI, and the future of AI in medicine. This conversation has been condensed and edited.
Tell us about your academic background. How did you get involved in your research field, and what attracted you to UW–Madison?
As I entered my undergraduate years at Penn State, I was interested in chemistry, but I found I enjoyed using computers to simulate and predict chemical processes more than memorizing chemical properties. So I decided to major in computer science and math, which led me to machine learning, where I could develop models that help make sense of complex data.
I completed my Ph.D. in Computer Science and an M.S. in Machine Learning at Carnegie Mellon University, where I focused on making AI more interpretable, especially in biomedical applications. After that, I was a postdoctoral associate at MIT’s Computer Science and Artificial Intelligence Lab (CSAIL) and the Broad Institute of MIT and Harvard, where I continued to work on AI for healthcare applications. After my postdoctoral work, I was able to secure a tenure-track role here in Madison in 2024.
I was attracted to UW–Madison because of its strengths in statistics, computer science, and biomedical research. There is also a rich history of impactful research and education, from the development of Warfarin, the revolutionary blood thinner, to the historical impact of George Box, the founder of the Department of Statistics here. Beyond that, the people I have met so far, from students to faculty and senior administrators, are very approachable, helpful and collaborative.
You mentioned your research focuses on making AI models interpretable. What does that mean, and why is it important in medicine?
Interpretability in AI means two main things. First it means we can understand what a model is doing, or how it makes its predictions. Second—and this is especially important in biomedicine—interpretability means the model helps us learn something new about the world. In medicine, we don’t just want a system that predicts patient outcomes; we need one that explains why those outcomes occur so we can identify opportunities to improve treatments.
For example, in pregnancy research, we want models that don’t just forecast risks but also reveal which factors—whether environmental, genetic, or behavioral—are driving those risks. An AI that predicts risk without explaining why misses the chance to guide better questions, tests, and interventions. My work focuses on designing AI systems that don’t just predict outcomes but also reveal how they come to those predictions, improving how human-AI teams acquire data and make decisions and enabling researchers to develop better interventions.
Interpretability matters because medicine isn’t just about reacting to information—it’s also about knowing what to ask. An AI system that predicts risk without explaining why misses the chance to guide better questions, tests, and interventions. My work focuses on designing AI systems that don’t just predict outcomes but also reveal why, improving how human-AI teams acquire data and make decisions.
How do these AI models adapt to different patients and situations?
Every patient is different, so context-adaptive models, like the ones I help design, recognize that what works for one person might not work for another. Instead of applying the same formula to everyone, these models adjust based on an individual’s medical history, genetics, and other factors. An example of this type of work is Personalized Regression, a method I helped develop to analyze complex biomedical data that identifies the most important factors for a specific person, rather than assuming the same risks apply to everyone. Methods like this can help move medical professionals toward more personalized treatment recommendations.
Can you discuss your recent research in the contexts of pregnancy and COVID-19?
Pregnancy is a unique challenge because there are two patients—the mother and the baby—and providers need to protect the health of both. Additionally, unlike randomized drug trials, where researchers test treatments in controlled settings, pregnancy data mostly comes from real-world cases. Models I’ve helped create in this area aim to help medical professionals understand which risk factors matter most, even when we can’t run controlled experiments.
COVID-19 was interesting from an experimental design perspective, too. By unfortunate accident, we had somewhat randomized trials going on naturally as the pandemic unfolded, helping us evaluate evolving risk factors and treatments. We built context-adaptive AI models to predict who was at the highest risk of severe illness, which was a real challenge because those risk factors kept changing as new variants emerged and patients’ medical histories evolved.
Where do you see AI making significant impacts in healthcare in the future?
We are finding that there’s a lot of low-hanging fruit, or opportunities to apply machine learning models to understand real-world outcomes better. For instance, in pregnancy, we’re finding that sophisticated analysis of real-world data can show some significant impacts of certain factors, such as maternal height, socioeconomic status, or chronic high blood pressure, on adverse outcomes like stillbirths. Patients and doctors can then adapt to those risk factors to improve health outcomes.
While using real-world data can be useful, it can only tell us so much. That’s why one of the biggest opportunities for AI in medicine is in personalizing treatment. By building AI models that reveal more personalized risk factors, we might find that we’re missing potential health interventions for subsets of people that population-level data fails to capture. This way, we can determine new interventions in a bottom-up way rather than a one-size-fits-all, top-down approach.
How do large AI models (also known as “foundation models”) fit into your research?
I’m very excited about the combination of foundation AI models and statistical inference. Large AI models, which can perform a variety of tasks—think text, image and audio generation—are powerful. But, they often act as black boxes, making their decisions hard to interpret. On the other hand, we have a big toolbox of statistical methods and models that have been developed over years of theory and practice. Those key concepts are only getting more powerful and important in the era of foundation AI models. My research will continue to explore ways we can apply our statistical toolbox to build better AI models, particularly for biomedical applications.
You work across multiple fields—computer science, biomedical informatics, and statistics. Why is this kind of collaboration important?
It is vital to involve researchers from across disciplines to continue making advances in this area. We need medical researchers to understand diseases, providers to understand treatments, statisticians to measure uncertainty and make predictions, and computer scientists to build scalable AI models. The potential for interdisciplinary research across departments at UW–Madison—both within the CDIS and beyond—is truly exciting.
In that spirit, I am currently welcoming graduate students to apply to join our lab! For students in CDIS or BMI who are interested in building AI and machine learning models to better understand precision medicine, this would be a great opportunity to dive into cutting-edge research in this area and publish in top venues for AI and biomedical informatics. I would also note that many other researchers in CDIS are looking at AI from various angles, such as my colleague Kris Sankaran, whose research investigates the intersection of AI and climate change adaptation. There are also faculty members in Computer Sciences and the iSchool working on various aspects of AI, so students interested in this research have a range of options for groups to join.
What do you enjoy doing outside of research and teaching?
Right now, I don’t have much free time because I have two young kids! In addition to spending time with family, I love Madison’s outdoor culture; it’s a great place to get outside. Also, I’m looking forward to going to some Badger hockey games!