The Identity Trap in EEG Foundation Models: A Diagnostic Audit
Researchers have identified a significant issue in EEG foundation models, termed the "Identity Trap," where models achieve high accuracy by learning subject-specific features rather than genuine clinical biomarkers. A new diagnostic tool, FMScope, was developed to identify this shortcut learning at the representation level. The tool's application revealed that subject-variance is a dominant factor in model performance, and erasing these identity features can improve the decoding of actual clinical markers. AI
IMPACT Highlights a critical shortcut learning issue in foundation models, necessitating new evaluation methods to ensure genuine clinical biomarker discovery.