Abstract
Shared Canonical Component Removal (SCCR) is an auditing framework for identifying predictive information in raw-input models that is not captured by expert-defined features. Across synthetic and clinical applications in coronary artery disease, antidepressant treatment dropout, and skin lesion classification, SCCR removes shared representation directions while retaining residual signal that exposes limitations in expert views.
- Type
- Preprint
- Publication
- arXiv
- DOI
- 10.48550/arXiv.2610.13819
Jose Roberto Tello Ayala
PhD Candidate in Applied Mathematics
I develop interpretable machine-learning methods for clinical decision-making, with applications in cardiovascular imaging and precision psychiatry.