Selection by repeat frequency did not show higher later accuracy with eight SEEG pairs and 32 calibration trials in eight patient sessions

We tested two rules for a reading brain-computer interface with eight public SEEG patient sessions. Both rules selected eight bipolar pairs and fitted a model with the first 32 retained trials. The final third of each session supplied the test trials. Mean balanced accuracy was 62.6% for direct effect ranking and 61.6% for selection by repeat frequency. The paired change was -0.97 percentage points. Its descriptive 95% patient interval was [-2.88, +1.01] points; the exact two-sided sign-flip test gave p = 0.375. This test did not establish higher accuracy or equal accuracy. Selections from two separate 32-trial blocks shared 2.375 of eight pairs on average. A secondary 16-trial gain needs a separate fixed test. The package adds a method comparison, a measure of selection uncertainty, and source, model, and inference checks. It preserves all input features, code, source versions, fitted models, predictions, result tables, and three figures. The data are reused from an earlier publication. A clean environment reproduced every test decision and statistical result. Known event times define the windows. This is an offline task-state test, with no test of continuous control, transfer across days, or stimulation.

stereoelectroencephalography · brain-computer interface · feature selection · stability selection · balanced accuracy · calibration · offline decoding