Eight ranked bipolar channels reached 95% of all-channel key-press decoding gain; random subsets needed 64 in six responsive SEEG subjects

This offline analysis measured how test AUC for cued key-press versus fixation-rest decoding changed with the number of bipolar SEEG channels. It used archived high-gamma (70–150 Hz) and beta (13–30 Hz) log-power features from 8 subjects in OpenNeuro ds004473, trained shrinkage LDA on trials 1–100, and tested on trials 101 onward. For each channel count N = 1, 2, 4, 8, 16, 32, and 64, the analysis averaged 200 random subsets or selected channels ranked by a move-versus-rest high-gamma t statistic from training trials. Among the 6 subjects with all-channel AUC ≥0.75, median normalized gain for random subsets was 0.17 at N=1, 0.50 at N=8, 0.85 at N=32, and 0.96 at N=64; N95 was 64, N50 was 8, and N80 was 32. Gain did not saturate before 64 channels in any subject. Ranked channels reached N95 at 8. At N=8, ranked channels had higher AUC than random subsets in all 8 subjects (mean difference 0.144, median 0.135; exact two-sided sign-flip p=0.0078). With high-gamma features alone, random N95 was not reached at 64 (median gain 0.94), while ranked N95 remained 8. These results apply to an offline analysis of one session per subject, one simple task, and clinical electrode placements. “Random” refers to subsets of clinical channels, not random brain placement. N95 is based on 6 responsive subjects and a twofold channel-count grid, so its value lies between 33 and 64 channels.

stereo-EEG · SEEG · brain-computer interface · channel selection · key-press decoding · shrinkage LDA · AUC · high gamma · beta power