Thirty calibration trials reached near-final SEEG decoding AUC; eight selected channels did not improve 10-trial performance

In an offline analysis of cued key presses from 8 SEEG subjects, we measured how calibration length and calibration-only channel selection affected move-versus-rest decoding on later trials. With archived high-gamma log-power features, shrinkage LDA using all channels reached within 0.02 of its 60-trial median test AUC by 30 calibration trials (median AUC 0.829 at 30 and 0.847 at 60). At 10 trials, selecting 8 channels did not improve AUC over all channels (mean difference -0.010; 4 of 8 subjects positive; exact two-sided sign-flip p = 0.35). The best single channel performed 0.06–0.10 AUC below all channels. Adding beta-band power increased all-channel AUC in a post hoc comparison. The analysis used one session per subject and a fixed later test block; secondary results were not corrected for multiple tests. These findings concern offline cued decoding and do not establish performance for real-time, self-paced BCI use or across days.

SEEG · brain-computer interface · calibration · channel selection · shrinkage LDA · ROC AUC · motor decoding · OpenNeuro ds004473