In 15 patients, 8 training-ranked SEEG channels yielded mean held-out AUC 0.804 versus 0.833 with all channels

We used archived 70–150 Hz high-gamma activity (HGA) features from 15 patients in OpenNeuro ds006253 and an unchanged shrinkage-LDA detector. The first 60% of trials, ordered by time, were used for training and the last 40% for testing. Bipolar SEEG channels were ranked by the absolute paired t statistic for MOVE minus REST on training trials only. Mean held-out AUC was 0.833 with all channels, 0.829 with the top 32, 0.814 with 16, 0.804 with 8, 0.779 with 4, and 0.725 with 1. Random 8-channel sets gave 0.651. For the primary comparison, top 8 minus all was −0.029; 13 of 15 patients had lower AUC with top 8 (exact two-sided sign-flip p = 0.0072). For top 1 minus all, the difference was −0.108; all 15 patients had lower AUC (p = 0.00006; Holm-adjusted p = 0.00012). The median smallest channel count retaining 95% of the all-channel AUC margin above 0.5 was 8, while 7 patients needed 32 or all channels. Thus, about 8 selected channels retained about 90% of the movement-detection margin, with a small, consistent loss. The study was offline, used one session per patient and task-locked windows, and compared MOVE and REST from different task phases; part of the difference may reflect task state rather than movement. The results are based on 15 patients, one univariate selection rule, and test sets of 48–120 trials.

stereo-EEG · SEEG · movement detection · channel selection · high-gamma activity · brain-computer interface · AUC · shrinkage LDA