Rest-only calibration did not reliably detect SEEG key presses in eight subjects

We tested whether a detector calibrated only on rest data could identify cued key presses from stereo-electroencephalography (SEEG) band-power features. Using archived features from eight subjects in OpenNeuro ds004473 version 1.0.2, we calibrated a squared Mahalanobis-distance novelty detector on rest windows from trials 1–100 and evaluated rest versus movement windows on trials 101 onward. The primary result did not support above-chance detection: subject AUCs ranged from 0.433 to 0.736, the median was 0.532, four of eight subjects were above 0.5, and the exact one-sided sign-flip test gave mean AUC above chance of 0.050 (p = 0.109). Shrinkage LDA using five labeled trials had higher AUC than the rest-only detector in seven of eight subjects (mean paired difference R − LDA5 = −0.114, p = 0.016). Diagonal, high-gamma-only, and 20- or 50-rest-window variants also did not show reliable above-chance performance. This offline, cued-task result is limited to eight patients, one session, one novelty score, and no adaptive update; the movement window followed a visual cue, so task changes as well as movement could be detected. For this type of SEEG motor switch, the findings do not support replacing labeled calibration with rest-only novelty detection.

stereo-electroencephalography · SEEG · brain-computer interface · motor decoding · novelty detection · Mahalanobis distance · shrinkage LDA · calibration · OpenNeuro ds004473