An onset-trained SEEG detector detected late reach activity in all 15 patients

We tested whether a detector trained on 70–150 Hz stereo-electroencephalography (SEEG) activity during the first 0.5 s of mouse movement could detect activity in the last 0.5 s before the choice click. Using author-provided window features from all 15 patients in OpenNeuro ds006253 v1.0.3, we trained shrinkage linear discriminant analysis (LDA) decoders on the first 60% of time-ordered trials and evaluated them on the last 40%. The onset-trained detector had a mean held-out area under the receiver operating characteristic curve (AUC) of 0.810 for late-phase detection and exceeded chance in all 15 patients (one-sided exact sign-flip p = 3.1 × 10⁻⁵). A late-trained detector had a mean late-phase AUC of 0.824; its advantage over the onset-trained detector was not significant (mean difference +0.014, p = 0.25; post-hoc 95% bootstrap CI −0.008 to +0.036). The onset-trained detector also showed no significant AUC change from onset to late phase (0.817 to 0.810; p = 0.76). These offline results come from one session per patient, a cued task, and reaches lasting at least 1 s. Late windows include click preparation, and non-significant differences do not establish equivalence.

stereo-electroencephalography · brain-computer interface · movement detection · high-gamma activity · linear discriminant analysis · cross-phase decoding · motor neuroscience