Cue plus median reaction time reduced movement detection AUC by 0.021 in 15 SEEG patients, a lower-bound offline estimate

A brain-computer interface (BCI) calibration for a paralyzed user cannot measure movement onset, so training labels must be set from cue timing. We compared onset-locked and cue-locked labels for a 70–150 Hz high-gamma activity (HGA) movement-versus-rest detector using archived author features from 15 SEEG patients in OpenNeuro ds006253. The cue-locked label began at cue plus the median reaction time of training trials, rounded to 0.2 s. Both shrinkage LDA decoders used the same training trials and were tested on the same held-out trials, using the true 0–0.4 s movement window against pre-stimulus rest. In a pre-registered time-ordered 60/40 split, mean held-out AUC was 0.835 with onset labels and 0.814 with cue labels. The mean cost (onset minus cue) was 0.021 (bootstrap 95% CI 0.008–0.037); cue labels lowered AUC in 13 of 15 patients (exact one-sided sign-flip p = 0.0012). Cue-locked AUC exceeded 0.5 in all 15 patients (p = 3.1e-5). Exploratory analysis found that greater spread in label shift tended to accompany greater cost (Spearman rho = 0.54, p = 0.040, uncorrected). This cost is probably a lower bound: features covered only −1 to +1 s around movement onset, and an average 25% of training trials with the largest reaction-time errors were removed from both decoders. The study is an offline simulation in able-bodied epilepsy patients making real movements; attempted movement in paralysis may have different signals and timing. The rest class was task-engaged pre-stimulus rest, and the test window used true movement onset.

brain-computer interface · movement detection · SEEG · high-gamma activity · calibration · reaction time · label noise · linear discriminant analysis · ROC AUC