An onset-trained SEEG decoder detected late movement non-inferiorly in 8 new patients

This pre-registered replication tested whether a movement-onset-trained decoder detects later movement in eight new patients from OpenNeuro ds006253. It used archived author-derived 70–150 Hz high-gamma activity (HGA) window features and the same analysis code and 0.05 AUC non-inferiority margin as the earlier seven-patient study. A Ledoit–Wolf shrinkage linear discriminant decoder trained on the 0–0.4 s onset window was tested on 0.6–1.0 s movement windows against pre-stimulus rest, using a time-ordered 60/40 trial split. The onset-trained decoder detected late movement above chance (mean AUC 0.795; all 8 patients above 0.5; Holm p = 0.012). The primary non-inferiority test passed: the mean difference between onset-trained and late-trained decoder AUC was −0.011 (7/8 patients within the 0.05 margin; Holm p = 0.035). One patient was outside the margin (−0.150). Onset-trained AUC declined from 0.884 at onset to 0.795 late (Holm p = 0.016); the late-trained decoder’s late-window AUC was also 0.806. The earlier seven-patient results were reproduced exactly. This is an offline result from one task, one session per patient, one time-ordered split, and window features rather than a causal streaming decoder. Pre-stimulus rest does not represent other non-movement states, and the late window can contain response clicks in short trials.

SEEG · motor brain-computer interface · movement detection · high-gamma activity · non-inferiority · replication