What Raw Models See That Experts Miss
Clinical experts routinely convert raw inputs—imaging scans, patient histories, dermoscopic images—into structured feature sets before rendering a decision. A persistent question in clinical AI is whether black‑box models trained on raw data capture signals that expert‑feature models overlook. Standard residualization or concept‑erasure techniques can strip away variance, but they do not distinguish between noise and genuinely complementary predictive information.
Shared Canonical Component Removal (SCCR) addresses this by selectively removing only those directions strongly shared between learned raw embeddings and expert embeddings. The framework does not simply subtract the expert view; it isolates the residual space where the raw model exerts predictive power independent of expert features.
Across four evaluation domains—synthetic data, coronary artery disease detection, antidepressant treatment dropout prediction, and skin lesion classification using SkinCon annotations—SCCR consistently identifies where expert views fall short. In each case, the method recovers residual predictive signal while minimizing redundancy with the expert-provided features. The synthetic benchmarks demonstrate the framework’s ability to isolate non‑overlapping components; the clinical datasets show that raw models exploit structured signal beyond what expert features capture, and that this signal is reproducible and clinically relevant.
The authors do not claim that raw models always outperform expert models, nor that the recovered signal is universally beneficial. Rather, SCCR provides an audit tool: it quantifies the gap, surfaces the “blind spots” of expert feature engineering, and offers a way to incorporate complementary raw‑model knowledge without re‑introducing the noise or bias that expert curation is designed to filter.
For researchers and engineers working on clinical decision support, the paper offers a concrete method for diagnosing model behavior relative to expert knowledge. The framework could inform hybrid pipelines that retain expert interpretability while surfacing additional signals from raw data. Whether SCCR see s broader adoption beyond medical AI remains to be seen, but its principled approach to component removal sets a new standard for model auditing in high‑stakes domains.