Movement-detection channels did not improve direction decoding over random channels in 15 SEEG patients

This offline analysis tested whether stereo-electroencephalography (SEEG) channels selected to detect movement also decode left/right direction. We analyzed author-provided 70–150 Hz high-gamma activity window means for 15 patients from the OpenNeuro ds006253 motion-discrimination task. We included correct trials without a change of mind, used stimulus direction as the label, measured log10 high-gamma activity from 0–0.5 s after movement onset, and evaluated five contiguous time-ordered folds with channel selection on training data only. Eight movement-detection channels gave a mean direction AUC of 0.543; eight random channels gave 0.529 (difference +0.013, 95% bootstrap CI −0.023 to +0.050; exact sign-flip p = 0.50; Holm-adjusted p = 0.75). Eight direction-selected channels gave AUC 0.584, a difference from random of +0.054 (95% CI +0.026 to +0.086; Holm-adjusted p = 0.0097). Detection and direction channel sets overlapped by a mean of 0.84 of 8 channels, versus 0.76 expected by chance (p = 0.62). Mean channel-rank correlation was Spearman rho = 0.028 (p = 0.32). Direction decoding remained weak in every set (mean AUC at or below 0.584); all-channel AUC was 0.571, consistent with an earlier result of 0.567. These results provide no evidence that movement-detection channels improve direction decoding over random channels in these clinical implants. The analysis was offline, included only correct trials without a change of mind, used one post-movement window and k = 8, and cannot separate motor direction from stimulus direction. The direction-selected set used the decoder label, so its improvement over random channels was expected.

stereo-electroencephalography · brain-computer interface · movement detection · direction decoding · high-gamma activity · channel selection · linear discriminant analysis