Sparse Autoencoders crack open EEG foundation models for clinical trust

A multi-institutional team led by William Lehn-Schiøler applied TopK Sparse Autoencoders (SAEs) to three leading EEG foundation models—SleepFM, REVE, and LaBraM—to shine light on their internal workings. The SAEs extract sparse feature dictionaries from model embeddings, then map these features to clinical concepts like abnormality, age, sex, and medication usage. This grounding allows researchers to benchmark monosemanticity (one feature = one concept) and entanglement across architectures. A single hyperparameter procedure, driven by an intrinsic dictionary health audit, robustly transfers across all three models, suggesting a universal approach to interpreting EEG AI.

The team introduced a 'target vs. off-target' probe area metric to quantify steering selectivity—how precisely a concept can be manipulated without affecting others. This revealed three operational regimes: selectively steerable, encoded but entangled, and non-encoded. Critically, they exposed 'wrecking-ball' interventions that collapse global model performance, and clinical entanglements such as age-pathology confounding, where suppressing one concept inevitably corrupts another. Finally, a spectral decoder translates these interventions into amplitude spectrum changes—e.g., pathological slow-wave suppression and alpha-band restoration—making latent manipulations physiologically interpretable. This framework is a major step toward trustworthy AI in neurology and psychiatry.