Burst suppression is an EEG pattern where bursts of activity alternate with near-silence, and clinicians watch for it to judge how deeply a critically ill patient is sedated, especially during an induced coma. Detecting it automatically has stayed difficult for two stubborn reasons: the pattern looks meaningfully different from one patient to the next, and annotated recordings are scarce.

Scarce labels are exactly the situation foundation models are supposed to help with, since the point of pretraining is to arrive already knowing something general. EEG foundation models have shown promise on a range of downstream tasks, but nobody had checked whether that promise carries over to burst detection.

This is the first evaluation of them for that job, and it uses reduced-montage ICU recordings rather than clean laboratory data. That choice matters. Fewer electrodes and real intensive care conditions are what the method would actually face, and it is where a model that looked strong on tidy benchmarks may turn out not to be.