High-gamma variance did not outperform random channel selection in 8 new SEEG patients; supervised selection did

In eight new ds006253 SEEG patients, selecting eight channels by label-free high-gamma variance (VAR) did not outperform the mean of 200 random eight-channel sets (mean AUC difference +0.010; median -0.021; 4/8 positive; one-sided exact sign-flip p = 0.391). Supervised selection by the absolute paired t statistic for move versus rest on calibration trials outperformed VAR in all eight patients (mean difference +0.163; Holm p = 0.0078) and outperformed random selection in all eight (mean difference +0.173; Holm p = 0.0078). Mean test AUCs were 0.686 for VAR, 0.676 for random selection, 0.849 for supervised selection, and 0.877 using all channels. The same code reproduced the reported values for seven original ds006253 patients. VAR was much better than random in two patients with the strongest movement signal, but it was not reliable across patients. The new cohort had eight patients; a non-significant primary test does not prove equivalence. The features used the same author extraction pipeline as the original work, and the task was fixed-window rest-versus-move detection, not continuous decoding.

SEEG · high-gamma activity · channel selection · movement decoding · replication · AUC