Random window splits gave 57.3 percent mean balanced accuracy for random trial labels in eight SEEG recordings with four overlapping windows

We tested four overlapping SEEG windows from each of 2014 retained trials in eight people. The public source was OpenNeuro ds004473 version 1.0.2. We assigned one random binary label to each trial and copied it to its four windows. All pair selection, scaling, and model fitting used training rows only. Over 100 label draws per person, random window splits gave 57.3 percent mean balanced accuracy. Random complete-trial splits gave 50.0 percent, and later-trial splits gave 49.9 percent. The window-minus-trial difference was 7.2 points, with a descriptive 95 percent person bootstrap interval of 6.5 to 8.1 points. All eight person differences were positive. Independent random labels for each window gave scores near 50 percent. The result shows false evidence for new-trial decoding under this window split. It does not isolate raw sample overlap from other shared trial information. The package preserves exact input features, labels, splits, model states, predictions, three figures, code, source hashes, and raw retrieval instructions.

stereoelectroencephalography · brain-computer interfaces · data leakage · cross-validation · overlapping windows · balanced accuracy · trial-level splits · SEEG