A shared SEEG model gave 52.6% mean balanced accuracy in eight tests that held out one person after rest calibration

A shared SEEG model learned from early task and rest windows from seven people. It then tested later trials from the eighth person. Early rest windows from that person provided calibration. The model used 15 spectral summary inputs from three power bands. It did not use that person’s task windows to fit the model. In eight tests, mean balanced accuracy was 52.6%. The descriptive 95% person bootstrap range was 48.2% to 56.1%. Mean AUC was 0.539. The paired-window swap reference gave an add-one value of 0.020. This value depends on exchangeable labels and does not establish a population effect. A reproduced model with eight person-specific bipolar pairs gave 70.7% mean balanced accuracy. Every shared score was lower than that person’s reference score. The package preserves features from 2,014 trials, source records, code, models, predictions, controls, and three figures. Independent signal and model checks passed. Full repeat runs matched all 17 numerical artifacts. This offline task-state test does not establish continuous motor intent, transfer across days, or a stimulation benefit.

stereoelectroencephalography · SEEG · brain-computer interface · cross-person transfer · balanced accuracy · spectral power · ridge regression