Sparse Autoencoders Crack Open EEG Foundation Models
Researchers applied TopK Sparse Autoencoders (SAEs) to three EEG transformers—SleepFM, REVE, and LaBraM—to expose the internal features driving their clinical predictions.
They grounded extracted features in clinical attributes (abnormality, age, sex, medication) and benchmarked how cleanly each model separates these concepts, then introduced a steering metric to measure selectivity.
A single hyperparameter tuning procedure transfers across all three architectures, suggesting a general recipe for interpretability in EEG models—critical for clinical adoption.