Decrease channels carry held-out movement information and improve equal-size SEEG decoding in 15 subjects

We tested whether bipolar stereo-electroencephalography (SEEG) channels with lower log 70–150 Hz power during movement carry information about movement on later trials. We used archived features from 15 subjects in two datasets (ds004473, 8 subjects; ds006253, 7 subjects), ranked channels on training trials by paired move–rest t statistics, and evaluated shrinkage linear discriminant analysis on time-ordered held-out trials. The eight channels with the largest decreases (NEG8) had mean held-out AUC 0.667; all 15 subjects had AUC above 0.5 (Holm-adjusted p = 6.1 × 10⁻⁵). Combining eight increase and eight decrease channels (MIX16) gave mean AUC 0.783, 0.068 above the 16 largest-increase channels (POS16), in 14 of 15 subjects (Holm-adjusted p = 9.2 × 10⁻⁵). NEG8 channels retained a negative test-set t statistic in 92.5% of channels; their mean test move–rest difference was −0.064 log units in all 15 subjects. MIX16 was close to using all channels (AUC 0.783 vs 0.788). These results show that a map based only on high-gamma increases can discard reproducible movement information in these data. The study used two datasets, one feature, one decoder, movement-versus-rest classification, offline evaluation, and one time-ordered split per subject; it did not test why power decreased or analyze channel anatomy.

stereo-electroencephalography · SEEG · brain-computer interface · movement decoding · high gamma · power suppression · shrinkage LDA · held-out evaluation