Privilege Is Not Enough: Evaluating Generalized Distillation for Normative Brain Age Modeling in Depression, Anxiety and Alcohol Use Disorder

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Models trained on rich multimodal data may depend on modalities unavailable at deployment. In generalized distillation, a teacher trained on such privileged data guides a student using only deployable features. We ask whether students inherit clinically meaningful properties beyond predictions in brain-age modeling, where brain-age gaps index deviations from normative aging and brain health. In simulations, distillation recovered a gap-severity association which an age-trained baseline missed. In UK Biobank analyses of depression, anxiety, and alcohol use disorder, students using structural MRI, functional MRI or biosocial data produced gaps closer to their teacher's than matched baselines in 10 of 12 comparisons. Where teachers most exceeded baselines (anxiety in biosocial, alcohol use in functional MRI students), gap-severity associations rose toward their teacher's. Our findings characterize conditions for successful distillation, i.e. clinical signal the teacher holds, the baseline misses and deployable features contain, and demonstrate their relevance in a real-life psychiatry application.

Article activity feed