Zero-calibration SEEG key-press decoding was near chance in 8 subjects, while 10 labeled within-subject trials reached median AUC 0.83

We tested whether a key-press-versus-rest SEEG decoder trained on other people could work in a new person without labeled trials. Using archived band-power features for 8 subjects from OpenNeuro ds004473 v1.0.2, we summarized high-gamma (70–150 Hz) and beta (13–30 Hz) channel z-values with their mean and seven quantiles, then trained shrinkage LDA in leave-one-subject-out tests. The prespecified primary test did not show reliable above-chance decoding: median AUC was 0.528, with 6/8 subjects above 0.5 (exact one-sided sign-flip p = 0.145). With normalization from trials 1–20, median AUC on trials 61–100 was 0.519. A within-subject decoder using all channels and 10 labeled trials reached median AUC 0.827 (8/8 subjects above 0.5; p = 0.0039); the same calibration using channel-agnostic summaries reached 0.597. These results suggest that, in this dataset, removing channel identity accounted for more performance loss than cross-subject transfer, and that these whole-implant summaries removed useful information from specific channels. The study used 8 subjects, one session, and offline windows at known event times; cue or visual activity could contribute, primary normalization used unlabeled windows from the full test set, only one channel-agnostic representation was tested, and secondary tests were not corrected for multiple comparisons.

SEEG · brain-computer interface · zero-calibration decoding · cross-subject transfer · shrinkage LDA · band power · channel-agnostic features · negative result