In 8 offline SEEG subjects, 0.1-s windows reduced key-press AUC vs 0.5 s at press + 0.25 s; 0.2–0.3 s matched 0.5 s when ending at press

We measured how the length W of the high-gamma power window changes key-press-versus-rest detection. Data: 8 SEEG subjects of OpenNeuro ds004473, 100 trials each. Method: adjacent bipolar pairs, 70-150 Hz zero-phase filter, log mean power, shrinkage LDA on all pairs, 5-fold cross-validated AUC, exact sign-flip test over subjects. We wrote the plan before decoding. Primary result: with the window end at press + 0.25 s, W = 0.1 s gave lower AUC than W = 0.5 s (mean difference -0.180, 95% CI -0.239 to -0.122, 8 of 8 subjects, p = 0.008). Median AUC rose from 0.55 (0.05 s) to 0.85 (0.5 s); 0.75 s and 1.0 s gave no more gain. Secondary result (not corrected): with the window end at the press, W = 0.2-0.3 s gave median AUC 0.86-0.87, the same as 0.5 s (0.82, p > 0.3). Thus part of the short-window loss is a loss of pre-press activity, not only noise. The rest variance of log power fell with W with a log-log slope of about -0.8, near the estimation-noise model 1/(80 Hz x W). Limits: offline, cued task, non-causal filter, one decoder, 8 subjects. The package has the code, provenance, compact data, results and a figure.

stereo-electroencephalography · brain-computer interface · high-gamma power · window length · key-press detection · shrinkage LDA · cross-validation · ROC AUC · signal processing · OpenNeuro