Eight SEEG channels were robust under random cross-validation but not shown non-inferior with chronological calibration

This replication examined whether eight bipolar SEEG channels selected within each training fold were non-inferior to all 100–112 channels for offline key-press versus rest decoding. The archived features from OpenNeuro ds004473 v1.0.2 cover eight subjects and the first 100 trials; log 70–150 Hz power was classified with LDA, using a non-inferiority margin of 0.02 AUC. The original LDA code reproduced the published result exactly (median dAUC −0.003, p = 0.039); new pooled Ledoit–Wolf LDA on the same folds gave median dAUC +0.001 (p = 0.035). Across 100 new random five-fold splits, eight channels met p < 0.05 in 86% of splits, and the split-averaged test gave p = 0.012. Four channels passed in 9% of splits, while 16 passed in 99%. With channels selected on trials 1–50 or 1–70 and tested on later trials, eight channels were not shown non-inferior (p = 0.24 and 0.12); 3–4 of eight subjects had a loss greater than 0.02 AUC. Sixteen channels were near the threshold (p = 0.051 and 0.055). The chronological evaluation used one split and 30–50 test trials, so it had less power than cross-validation; failure to show non-inferiority is not proof of inferiority. The evidence is limited to eight subjects, one session, offline cued key press versus pre-trial rest. The margin was chosen by the original publication, and no correction was made for multiple tests over channel counts and conditions.

stereo-EEG · brain-computer interface · channel selection · cross-validation · linear discriminant analysis · non-inferiority · key-press decoding