High-gamma-decrease channels decoded key press and improved 16-channel decoding in eight SEEG subjects

This offline analysis asked whether SEEG channels with decreased high-gamma (HG, 70–150 Hz) power during a cued key press carried movement-state information beyond channels with increased HG power. Using archived features from eight subjects (100–112 bipolar channels each), channels were ranked by the calibration-trial paired t statistic for HG(move) − HG(rest) on trials 1–100; shrinkage LDA was trained on those trials and evaluated on trials 101 onward. The eight channels with the largest decreases (DOWN-8) had test AUC above 0.5 in all eight subjects (median 0.65, range 0.55–0.82; exact one-sided sign-flip p = 0.0039). Combining UP-8 and DOWN-8 outperformed UP-16 at the same 16-channel budget in all subjects (mean AUC difference +0.064, range +0.003 to +0.20; p = 0.0039). A sign-only score using negated mean HG over DOWN-8 also exceeded 0.5 in all subjects (median AUC 0.64; p = 0.0039). Removing all channels with calibration t < 0 from an all-channel decoder did not cause a significantly greater AUC loss than random removal (mean 0.063 vs 0.037; p = 0.10). The sign-only score is not fully label-free because channel signs were selected using labeled calibration trials. The rest interval was during fixation and the movement interval included the cue and visual input, so decreases may reflect attention, vision, or task state rather than motor control. The analysis was offline, used one session per subject in eight epilepsy patients, tested movement versus rest rather than left versus right, and used the preplanned channel count K = 8 only; subject 11 had 32 test trials.

SEEG · high-gamma power · brain-computer interface · key-press decoding · channel selection · shrinkage LDA · AUC · intracranial EEG