In 8 SEEG subjects, label-free channel ranking did not beat random and underperformed label-ranked channels for offline key-press decoding

We reused the archived ds004473 SEEG features of SCI publication 07478163 (SHA-256 checked) to test whether a label-free rule can select channels for a small-channel key-press detector. Three rules ranked bipolar channels within each training fold without labels: variance of log 70-150 Hz power over all windows (UVAR), the same variance over rest windows only (RVAR), and rest 13-30 Hz power (RBETA). The primary endpoint, set before decoding, was AUC(UVAR, k = 8) minus AUC(8 random channels). The mean difference was +0.021 AUC in 6 of 8 subjects, with exact one-sided sign-flip p = 0.25. RVAR and RBETA also did not outperform random channels at k = 4, 8, or 16. Label-ranked channels (|Welch t|) outperformed UVAR in all 8 subjects at each k; at k = 8, median AUC was 0.824 for label-ranked channels, 0.623 for UVAR, and 0.840 for all channels. The UVAR and label-ranked top-8 sets overlapped less than chance (median 0.2 vs 0.62 channels). In this offline task, these three simple label-free rules did not replace labelled or cued calibration for channel selection. Limits include 8 subjects, one session, 100 trials, and only three simple label-free scores.

stereo-EEG · SEEG · brain-computer interface · channel selection · label-free ranking · high-gamma power · motor decoding · linear discriminant analysis · ROC AUC · OpenNeuro ds004473