Abstract

Artificial Intelligence is reshaping every layer of modern healthcare, from disease prediction and medical imaging to personalized medicine and drug discovery. Yet the path from a model that performs well in the lab to one that delivers value at the bedside is rarely straightforward, giving rise to a translational gap that we set out to examine in this seminar. A central challenge is that patient populations, imaging modalities, and disease prevalence vary across sites and over time, quietly breaking the assumptions behind standard model transfer. Compounding this, data at deployment sites are often scarce and unlabeled, making training at the target domain particularly difficult. We show how unsupervised domain adaptation can recover lost performance, and present results in medical imaging, specifically multi-site prostate MRI and multi-modality whole-heart segmentation. Yet once the clinician enters the AI loop, accuracy alone is not enough: a single best-guess prediction conceals the model’s doubt where stakes are highest, leaving clinicians without a basis for trust. To address this, we introduce conformal prediction as a distribution-free safety wrapper that delivers patient-level prediction sets with mathematically certified coverage. We show that we can efficiently control set size and coverage, effectively managing the tradeoff between model usability and trust. Still, even an accurate and trusted model, delivers little value if it cannot navigate the operational complexity of a hospital. We close by describing a surgical case-sequencing system we have deployed at the Duke Ambulatory Surgery Center, which combines risk-aware optimization, digital twins, and online scheduling to turn predictions into decisions that improve OR utilization, reduce cancellations, and smooth patient flow. Taken together, these threads point to a single conclusion: real impact in Health AI is achieved only when robustness, trust, and actionability are engineered into the system from the outset, rather than appended as an afterthought.

Short Biography

Michael M. Zavlanos is a Professor in the Department of Mechanical Engineering and Materials Science at Duke University. He also holds a secondary appointment in the Department of Electrical and Computer Engineering and the Department of Computer Science. He currently serves as the Director of the Healthcare Systems Optimization program with Duke AI Health and is also an Amazon Scholar with Amazon Robotics. His research focuses on control theory, optimization, machine learning, and AI with applications in autonomous systems and healthcare/medicine. He is a recipient of various awards including the Office of Naval Research Young Investigator Program (YIP) Award and the National Science Foundation Faculty Early Career Development (CAREER) Award. Michael M. Zavlanos received the Diploma in mechanical engineering from the National Technical University of Athens (NTUA) in Greece, in 2002, and the M.S.E. and Ph.D. degrees in electrical and systems engineering from the University of Pennsylvania, in 2005 and 2008, respectively.

________________________________________________________________________________

Microsoft Teams meeting

Join: https://teams.microsoft.com/meet/36764456066195?p=xF3xQiTiIdz4zMT9Ao

Meeting ID: 367 644 560 661 95

Passcode: Kc6fy2Kz