Random and forward splits gave mean choice AUCs of 0.564 and 0.559 after movement in 19 offline SEEG tests

We compared random and forward trial splits in public MetaRDK SEEG data. The primary reader used eight bipolar pairs and a fixed logistic penalty of 0.1. The signal window was 50 to 450 ms after mouse movement onset. All four stimulus direction by choice cells had exactly matched fit and test counts. We averaged 50 random splits for each of 19 people. Mean choice AUC was 0.564 for random splits and 0.559 for the forward split. The paired difference was 0.005, with a 95% person bootstrap interval from -0.021 to 0.033. The exact two-sided sign-flip p value was 0.727. This result gives no clear mean increase for random splits under these rules. It does not establish equivalence. The package preserves 21 compact source files, all models, 513,162 predictions, three figures, code, and source retrieval instructions. A fresh source recording rebuild and a full analysis replay passed. The source uses offline processing, so these scores do not establish live command timing. This study does not test brain stimulation.

SEEG · brain-computer interface · choice decoding · trial splitting · cross-validation · AUC · model validation