Label-free channel ranks did not improve SEEG movement decoding over random selection

We tested whether two label-free channel ranks based on pooled, unlabelled log 70–150 Hz power could identify movement-informative SEEG channels. Archived features from 15 subjects in two OpenNeuro datasets were evaluated with a time-ordered train/test split and shrinkage linear discriminant analysis (LDA). The primary hypothesis failed: for eight channels, low excess kurtosis differed from the mean of 200 random selections by +0.001 AUC (6/15 subjects positive; one-sided exact sign-flip p = 0.49). A two-Gaussian mixture separation rank also did not outperform random selection (+0.029 AUC; Holm-adjusted p = 0.62). The low-kurtosis rank was below supervised selection by 0.152 AUC (1/15 positive; Holm-adjusted p = 0.0005). A post-hoc exploratory check found median per-channel movement/rest separation d = 0.14 among 1,393 channels; no channel had d > 2. Findings are limited to two datasets, 15 subjects, log 70–150 Hz power, one decoder, movement versus rest, and one time-ordered split per subject. Training windows were balanced, a favorable condition for detecting bimodality. The result does not establish performance in other tasks, features, decoders, or populations.

SEEG · brain-computer interface · channel selection · kurtosis · Gaussian mixture · movement decoding · linear discriminant analysis · high-gamma power