Move and rest scores had similar spread in an offline SEEG key-press decoder across eight subjects

This analysis asked whether move-window scores from a high-gamma plus beta shrinkage linear discriminant analysis (LDA) decoder had greater spread than rest-window scores in an offline SEEG key-press task. It used archived features from OpenNeuro ds004473 version 1.0.2 for eight subjects. The decoder was trained on trials 1–100 and evaluated on trials 101 onward (32–198 trials per subject). The preregistered primary measure was the mean across subjects of ln(SD_move / SD_rest), tested with an exact one-sided sign-flip test. Move scores did not show a significantly larger spread: mean ln ratio +0.037 (SD ratio 1.04), median +0.062, five of eight subjects positive, exact p = 0.22. The hypothesis was not supported. Robust spread, within-training-set cross-validation, and a high-gamma-only decoder gave similar group results. Two subjects had Brown–Forsythe p < 0.05, in opposite directions. The median binned z-ROC slope was 0.96. The equal-variance binormal model’s AUC-based prediction of true-positive rate had mean error +0.003 at false-positive rate (FPR) 0.05 and +0.013 at FPR 0.10; individual errors ranged from −0.17 to +0.09 at FPR 0.05, with few rest windows contributing to sampling uncertainty. These results support the equal-variance model as a first approximation for this decoder and task at FPR 0.05–0.10. They do not establish performance at lower FPR. The study was offline and limited to one cued task, eight patients with one session each, one linear decoder, and 0.5-second cue-locked windows. With eight subjects, a small true SD ratio, such as 1.05, cannot be excluded.

SEEG · brain-computer interface · linear discriminant analysis · ROC curve · AUC · score variance · false-positive rate · motor decoding