Eight SEEG subjects needed 32–64 cued trials to approach 80-trial AUC in offline key-press detection

We measured the calibration cost of an offline SEEG key-press versus rest detector using archived per-trial bipolar log-power features from eight subjects in the cued SlowFast task (100 trials per subject). Across 50 random splits per subject, shrinkage LDA was trained on nested sets of 2–80 trials and evaluated on the same 20 held-out trials per split. For 70–150 Hz power, median AUC increased from 0.638 at 4 trials to 0.695 at 8, 0.809 at 32, and 0.839 at 80. AUC at 8 trials was lower than at 80 in all subjects (median difference −0.115; exact two-sided sign-flip p = 0.0078). Three subjects reached within 0.02 of their 80-trial AUC at 32 trials and five at 64; the median curve still increased from 64 to 80 trials. Adding 13–30 Hz power increased median AUC at every tested size from 4 trials. Selecting eight channels using training trials only did not reduce calibration cost: at 4 trials, TOP8 was worse than all channels (median difference −0.025, p = 0.023), while differences were small at 32–80 trials. The result is limited to offline, fixed cued windows from one session, eight subjects, and electrodes placed for epilepsy; it does not establish self-paced BCI performance.

stereo-EEG · SEEG · brain-computer interface · motor decoding · key-press detection · calibration · ROC AUC · shrinkage LDA · channel selection · cued task · offline analysis · OpenNeuro ds004473