Track genomics, neuroscience and molecular biology.
Search papers, labs, and topics across Lattice.
This paper introduces Shared Canonical Component Removal (SCCR), a novel audit framework to quantify the predictive signal in raw model inputs that is not captured by expert-defined features. SCCR selectively removes shared canonical components between raw and expert embeddings, allowing for the isolation and recovery of residual predictive information. Evaluations across synthetic and clinical datasets demonstrate SCCR's ability to identify expert view limitations and recover valuable predictive signal with minimal redundancy.
Experts miss significant predictive signal in raw data that can be recovered by removing only the *shared* canonical components between raw and expert embeddings.
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.