Invited Talk at the ML4H Clinical AI Seminar Series, Broad Institute of MIT and Harvard

The seminar is held online on 4 November 2026, 12:00–13:00 EST (18:00–19:00 CET). More on the series: broadinstitute.org/ml4h.

Slides: interactive deck (press → to advance; some slides have live toggles).

From Shortcuts to Shared Structure: Learning Objectives for AI That Works Where Care Happens

Medical foundation models are increasingly reported to “generalise”, yet much of that evidence comes from pooled benchmarks that can hide what the models have actually learned. In this talk I show that frozen foundation-model features often encode where an image was acquired more strongly than what it shows. In mammography, a linear probe recovers the source hospital almost perfectly, and aggregating multi-institutional data can mask this shortcut. I then ask what the objective function should look like when the goal is AI that works beyond the hospital it was trained in. I present four answers from our lab:

• Group-robust training and shortcut diagnostics based on mutual information, for auditing representations across sites and populations.
• Federated, label-free canonical priors that remove the hospital fingerprint while keeping clinically meaningful axes.
• K-MaT, a prompt-learning method that transfers diagnostic structure from high-end to low-end modalities, without low-end training images, through knowledge anchoring and fused Gromov–Wasserstein alignment.
• PromptGate, a vision–language gate that saves expert labels in open-set federated active learning.

Finally, I present MRIQT, a physics-aware 3D diffusion model that brings portable 0.064 T neonatal MRI towards high-field quality, validated in a clinical reader study. Together, these works argue for objectives that build in robustness, physics and clinical knowledge, so that models serve the patients and settings where care actually happens.