In 11 ds006253 patients, movement decoding reached mean AUC 0.825; a pre-movement control reached 0.696

We applied an unchanged movement detector to one previously unused patient (sub-21) from OpenNeuro ds006253 and combined this result with 10 earlier patient results. The detector used author-provided 70–150 Hz high-gamma activity from stereoelectroencephalography contacts, Ledoit–Wolf shrinkage linear discriminant analysis, and a time-ordered 60/40 train/test split. For sub-21, held-out movement-versus-rest AUC was 0.953 (within-trial label-swap permutation p = 0.0001). Across 11 patients, mean AUC was 0.825 (95% patient-bootstrap CI 0.758–0.885; median 0.836; range 0.569–0.965); all 11 passed the per-patient criterion, and the exact sign-flip p value for mean AUC above 0.5 was 0.00049. File size and AUC had Spearman rho 0.57 (two-sided permutation p = 0.072), so the preregistered no-relation hypothesis was not rejected. The positive trend and selection of the largest files for earlier work mean that the cohort mean may be optimistic for smaller files. In an exploratory comparison using the 0.5 s before movement as the negative window, mean AUC was 0.696 (median 0.669; all 11 above 0.5), 0.129 below movement-versus-rest. This suggests that task-phase differences may account for part of the movement-versus-rest result. The analysis was offline, used one split per patient and author HGA only, had no closed-loop test, and did not assess 10 of 22 patients because of a time limit. The analyzed source files were fetched from an unversioned OpenNeuro S3 URL and are identified by hashes in the accompanying provenance files; their match to any version-specific dataset snapshot is unknown.

stereoelectroencephalography · high-gamma activity · movement decoding · brain-computer interface · linear discriminant analysis · held-out AUC · selection bias · time-ordered split