SEEG key-press decoding in eight subjects used channel patterns; removing the channel mean preserved AUC within 0.02

We tested whether SEEG key-press decoding uses a global change in band power or the spatial pattern across channels. We used archived bipolar log band-power features for eight subjects from OpenNeuro ds004473 v1.0.2 and trained shrinkage LDA on trials 1–100, testing on later trials. For high gamma (70–150 Hz), median test AUC was 0.538 for the channel mean, 0.851 for all channels and 0.847 after removing the per-window channel mean. The all-channel decoder outperformed the global-mean decoder in all eight subjects for high gamma, beta and both bands (exact sign-flip p = 0.0039). Removing the channel mean was non-inferior within a margin of 0.02 AUC (p = 0.0039 for high gamma; p = 0.0195 for beta and both bands). About half of channels increased and half decreased in high-gamma power during the key press, so their changes canceled in the mean. For this task and bipolar SEEG, removing the common-mode component of log band power did not reduce transfer AUC. Scope is limited to one task, one dataset, eight subjects, bipolar features and one train/test split; timing and visual-cue effects may contribute.

stereo-EEG · SEEG · brain-computer interface · key-press decoding · band power · common-mode removal · shrinkage LDA · session transfer · non-inferiority