New method reveals what raw data models miss compared to expert features
Shared Canonical Component Removal Beyond Expert Views
Machine Learning
Summary
When doctors use computers to help diagnose diseases, they often start with expert-made features from raw data. Sometimes, AI models trained on raw data do better than those using these expert features, but it's unclear which extra signals they use. The authors created a tool called Shared Canonical Component Removal (SCCR) that carefully removes overlapping information between expert features and raw data models. This helps find what extra clues the raw data models pick up that experts might miss, improving understanding and predictions in medical tasks.
What this means in practice
- •For clinical data scientists: Improve disease prediction models by identifying missing signals beyond expert features in medical data.
- •For medical imaging developers: Enhance image classification tools by removing redundant expert view information to focus on novel predictive signals.
Authors
Jose Roberto Tello Ayala, Akl C. Fahed, Roy Perlis, Finale Doshi-Velez
Abstract
In clinical practice, experts often begin by converting raw inputs into a set of expert features before making a decision. While black-box models trained on raw input can outperform models trained on expert views, it remains unclear what predictive signal the raw model exploits beyond what the expert view captures. We introduce an audit framework to quantify this gap using Shared Canonical Component Removal (SCCR). Unlike standard residualization or concept erasure, SCCR selectively removes only those directions strongly shared between learned raw and expert embeddings. Evaluations on synthetic data, coronary artery disease (CAD) detection, antidepressant treatment dropout prediction, and skin lesion classification using SkinCon annotations show that SCCR identifies the ways in which expert views are insufficient and recovers residual predictive signal while minimizing redundancy with the expert view.