Beta power increased full AUC but did not significantly improve low-FPR SEEG key-press detection in eight subjects

This offline analysis tested whether adding 13–30 Hz beta log power to 70–150 Hz high-gamma (HG) log power improves SEEG key-press detection at false-positive rates (FPRs) of 0–0.10. We used unchanged archived features from eight subjects and shrinkage linear discriminant analysis trained on trials 1–100, with trials 101 onward held out. The pre-specified primary endpoint was McClish-standardized partial AUC over FPR 0–0.10 (sPAUC10; 0.5 = chance). Adding beta increased full AUC by a mean of 0.040 in six of eight subjects (exact two-sided sign-flip p = 0.031), while the mean sPAUC10 gain was 0.026 in six of eight (p = 0.086). The primary hypothesis was therefore not supported at alpha 0.05. The low-FPR gain was not significantly smaller than the full-AUC gain (p = 0.41). At FPR 0.05, the TPR gain was 0.021 (p = 0.48). With a threshold set at the 95th percentile of out-of-fold calibration rest scores, mean test TPR increased from 0.336 to 0.424 (p = 0.070), while mean test FPR changed from 0.044 to 0.040. Effects varied across subjects: the low-FPR gain was negative in subjects 1 and 11, and most of the full-AUC gain occurred above FPR 0.10 in subjects 10 and 12. The study used eight subjects, one session, one decoder and one train/test split. It was an offline analysis of cued key presses, not a closed-loop test; at FPR 0.05 each subject contributed approximately 2–10 rest windows.

stereo-EEG · key-press detection · false-positive rate · partial AUC · high-gamma · beta power · linear discriminant analysis · brain-computer interface