Recorded Talks

Loading...

Loading description...

MechSci: Accelerating Clinical Science through Mechanistic Interpretability of Medical Foundation Models

Robert Holland, PhD — Postdoctoral Scholar, AIMI Center; Department of Radiology, Stanford University

Medical foundation models can identify prognostic signals in patient data at scale, yet their black-box nature prevents translation into testable scientific hypotheses. This talk demonstrates how to transform these models into engines for scientific discovery using mechanistic interpretability — specifically Sparse Autoencoders (SAEs) combined with LLMs — to decompose abstract features from 3D CT scans and lab values into understandable clinical concepts.

Loading...

Loading description...