A label-free detector recovers a working-memory decoder from simulated SEEG channel failures with offsets of 2 SD or more

We simulated post-training failures on archived SEEG band-power features from 15 patients performing a verbal working-memory task (set size 8 vs 4). Failed channels lost their neural signal and received offsets of 0, 1, 2, 3, or 5 training standard deviations on 10% or 25% of channels. With 25% of channels failed and a 3 SD offset, the undetected-failure decoder had mean AUC 0.645, compared with 0.703 after known-channel mean imputation (intact AUC 0.714). A label-free detector based on the band-mean z score across 3–7 unlabeled test trials detected 99–100% of failed channels for offsets of at least 2 SD, with a 1.1% false-alarm rate. Detection followed by imputation improved AUC by 0.056 over leaving failures undetected (14/15 patients; Holm-adjusted p = 0.0043) and was within 0.003 AUC of known-channel imputation. Retraining on surviving channels from stored training data added 0.006 AUC over imputation (12/15 patients; Holm-adjusted p = 0.0079). The detector found only 35% of failures at 1 SD and none at 0 SD. False alarms had a mean cost of −0.0005 AUC (p = 0.93), with larger cost in some patients. Failures were simulated; the results concern one dataset, task, session, feature set, and linear decoder, with 17–35 trials per patient. The detector requires a batch of 3–7 unlabeled trials. Real failures may have other feature patterns.

intracranial EEG · SEEG · brain-computer interfaces · channel failure · fault detection · mean imputation · linear discriminant analysis · working memory