EEG Foundation Models vs. Hand-Crafted Features: What Do They Learn?

A new study on arXiv looks into whether EEG foundation models capture the same brain signal patterns as traditional hand-crafted features. The research focuses on three key areas: what the model actually learns, how it’s applied, and how explainable it is. Using methods like layer-wise ridge probing and LEACE-style cross-covariance subspace erasure, the study evaluates three models (CSBrain, CBraMod, LaBraM) across five clinical tasks (MDD, Stress, ISRUC-Sleep, TUSL, Siena), analyzing 63 features in six categories. Out of 945 combinations of model, task, and feature, only 6% show a meaningful connection, suggesting that these foundation models tend to learn different representations compared to traditional approaches.