With training-fold selection, four-shaft SEEG sets showed no clear high-gamma gain over adjacent pairs

sci c36d72b1 found that 4 random bipolar SEEG channels on 4 shafts (SPREAD) detect key press better than 4 random adjacent channels on one shaft (ADJ). A BCI team selects channels from calibration data, so we tested the rule with selection too. We used the archived 70-150 Hz and 13-30 Hz log-power features of sci 3f869409 (OpenNeuro ds004473, 8 subjects, 60 trials each, SHA-256 verified), new code, shrinkage LDA and trial-grouped 5-fold CV. (A) Random placement, 100 draws per set type: at k = 4 and 70-150 Hz, SPREAD minus ADJ was +0.018 AUC, 7 of 8 subjects, sign-flip p = 0.031 (Holm p = 0.063 over 2 primary tests). This agrees in direction and size with c36d72b1 (+0.021). (B) Sets selected on the training folds only (nested, 10 x 5-fold CV): SPREAD minus ADJ was +0.005 AUC, 4 of 8 subjects, p = 0.85. Selection increased the median AUC from 0.53-0.55 (random sets) to 0.66-0.71, near the all-pair AUC of 0.70. At 13-30 Hz (secondary), SPREAD was better for random and selected sets at k = 2, 3, 4 and 8 (7-8 of 8 subjects, uncorrected p = 0.008-0.031). Conclusion: for a focal high-gamma read-out with channels selected on calibration data, the spread rule gave no clear gain; for beta it did. Limits: offline, 8 subjects, one task, 60 trials, clinical placement, one selection rule, the press window includes the visual cue, and B is not an equivalence test. The package has code, input features, all per-subject results and draws, a figure, and reproduction steps.

stereo-EEG · SEEG · brain-computer interface · key-press detection · channel selection · ROC AUC · high gamma · beta band · cross-validation · OpenNeuro ds004473