In 8 SEEG subjects, a 0.2 s window did not show non-inferiority to 0.5 s for offline key-press detection, and the best length varied by subject

We tested how the analysis window length L (0.05 to 1.0 s) and the window end time (0.3 s before to 0.2 s after the key press) change offline key-press vs rest detection in 8 SEEG subjects of OpenNeuro ds004473 v1.0.2. Features were bipolar log power at 70-150 Hz and 13-30 Hz; the decoder was shrinkage LDA with 5-fold cross-validation over 100 trials. Primary endpoint (pre-specified): AUC(0.2 s) - AUC(0.5 s) at 70-150 Hz with the window ending at the key press. The mean change was -0.021 (median -0.039, sign-flip p = 0.30); the non-inferiority test with margin 0.02 failed (p = 0.52), so the result is not decided. Mean AUC stopped increasing at about 0.3 s (0.64 at 0.05 s, 0.71 at 0.1 s, 0.78 at 0.2 s, 0.81 at 0.3 s, 0.80 at 0.5 s and 1.0 s). The best length was different for each subject (0.1 s to 1.0 s). Adding 13-30 Hz power increased AUC in all 8 subjects at 0.05 s (+0.083) and 0.1 s (+0.056), but only by +0.004 at 1.0 s. Limits: offline, zero-phase filter (small leak after the window end), 8 subjects, two subjects near chance. Code (numpy only), per-subject features, results, provenance, and an SVG figure are in the package.

SEEG · key press detection · window length · AUC · shrinkage LDA · band power · motor decoding