Two label-free SEEG channel ranks did not outperform random selection in 15 subjects across two offline datasets

In archived intracranial SEEG feature data from 15 subjects across two datasets, we tested whether two label-free channel ranks could select channels that detect movement better than random selection. MIX ranked channels by separation of a two-Gaussian mixture fit to pooled, unlabelled training windows; PC1 ranked by absolute loading on the first principal component. With shrinkage LDA on held-out windows, MIX top-8 mean AUC was 0.630 versus 0.623 for random channels (difference +0.006; 6/15 subjects positive; one-sided exact sign-flip p = 0.42). PC1 top-8 mean AUC was 0.604 (difference -0.019; Holm-adjusted p = 0.37). Neither rank beat random selection at K = 4, 8, or 16. Both were substantially below supervised top-8 selection (mean AUC 0.776; MIX and PC1 differences -0.146 and -0.172, respectively; each positive in 1/15 subjects, Holm-adjusted p < 0.001). The ranks also had little association with the supervised scores. Together with a previously reported variance rank, three simple label-free ranks did not identify movement-informative channels in these data; channel selection for this detector still required labelled calibration. This result is limited to 15 subjects, two offline datasets, log 70-150 Hz power, shrinkage LDA, movement-versus-rest task windows, one time-ordered split per subject, and the tested rank definitions. The exact OpenNeuro releases underlying the archived feature files were not recorded.

SEEG · brain-computer interface · channel selection · label-free ranking · movement decoding · high-gamma power · shrinkage LDA · OpenNeuro