Keynote at the BIGS CPS Retreat 2026: When AI Learns the Hospital, Not the Patient

On Monday, 16 November 2026, 11:00–12:00 CET, I give the keynote “When AI Learns the Hospital, Not the Patient: What Population Science Can Teach Medical AI” at the annual retreat of the Bonn International Graduate School of Clinical and Population Science (BIGS CPS), in the Festsaal of the University of Bonn’s main building (Am Hof 1, Bonn).

Abstract

AI is often promised as the engine of the hospital of the future: predicting disease at home, triaging patients at the front door, and supporting clinicians in the smart hospital. Yet behind impressive benchmark scores, many models learn the hospital rather than the patient. They pick up scanners, markers and case mix instead of disease, and they fail when moved to a new site or a different population.

In this keynote, I argue that the problems standing in the way are ones population science has studied for decades: confounding, selection, access to data, and measurement. Drawing on work from our lab at the University Hospital Bonn, I will show how these lenses help build AI that works where care happens. Examples include shortcut learning in breast cancer screening, under-represented populations in global health, and federated learning across hospitals that never share patient data. I will also show measurement done more cheaply: imaging-derived phenotypes from routine scans, portable MRI for newborns, inherited anaemias from a simple density gradient, and virtual staining that knows when not to trust itself.

I will close with concrete ideas for what AI could do with the cohorts and studies of BIGS CPS, and how doctoral researchers can work with us.