Eight selected SEEG channels replicated non-inferiority in cross-validation but not in a temporal split

In eight subjects from the sEEG Forced Two-Choice Task (OpenNeuro ds004473, version 1.0.2), we repeated a test of whether eight training-selected bipolar channels were non-inferior to all channels for cued key-press versus rest decoding, using a 0.02 AUC margin. We used archived features whose eight SHA-256 hashes match the source manifest, and new Ledoit-Wolf shrinkage LDA code. In 5-fold cross-validation with 70–150 Hz features, mean dAUC (eight selected minus all channels) was +0.013 (median +0.011; 3/8 subjects lower; one-sided exact sign-flip p = 0.008). With 70–150 Hz plus 13–30 Hz features, mean dAUC was +0.003 (p = 0.020). In a temporal split that selected and trained on trials 1–50 and tested on trials 51–100, non-inferiority was not shown at eight channels: mean dAUC was −0.007 for 70–150 Hz (p = 0.242) and −0.028 for combined features (p = 0.648). At 32 channels, non-inferiority was shown for both temporal conditions (p = 0.008 and 0.012). Selected channels outperformed random channel sets at k = 8 (8/8 subjects, p = 0.008); selection using all trials increased the apparent gain by 0.038 AUC in cross-validation and 0.058 in the temporal split. With eight subjects, failure to show non-inferiority is not proof of inferiority. The temporal split used 50 training trials versus 80 in cross-validation, so the effects of fewer training trials and non-stationarity cannot be separated. This offline analysis used only the first 100 trials, and it did not test online closed-loop decoding.

stereo-EEG · SEEG · brain-computer interface · channel selection · non-inferiority · cross-validation · temporal generalization · linear discriminant analysis · AUC · replication