Recent-trial recalibration did not improve SEEG key-press decoding, while cumulative training improved AUC in 8 subjects

We tested whether a key-press versus rest detector trained on the first 100 trials lost accuracy during one SEEG session, and whether recalibration improved it. We used archived log band-power features (70–150 Hz high gamma; 13–30 Hz beta) for 8 subjects from OpenNeuro ds004473 v1.0.2. A shrinkage linear discriminant analysis (LDA) was tested on trials 101 onward in blocks of 25. The primary comparison used high-gamma within-block AUC. Retraining on the latest 100 trials (RECENT) rather than retaining the initial model (FIXED) changed mean AUC by +0.009 (4 of 8 subjects above zero; exact two-sided sign-flip p = 0.219), not a significant difference. Training on all earlier trials (CUMULATIVE) changed mean AUC by +0.020 (7 of 8 above zero, 1 tie; p = 0.016; secondary, without correction). FIXED AUC had no significant decline over blocks (slope −0.0066 per block; 7 subjects; p = 0.312). Feature means moved farther from the training mean in 6 of 7 subjects (p = 0.031), without a corresponding AUC decrease. Label-free recentering changed balanced accuracy by +0.0044 (p = 0.250). The findings apply to offline analysis of one session in 8 subjects, up to 26.8 minutes after calibration. Power was low for small effects; CUMULATIVE used more training trials than FIXED, so its gain reflects data quantity and does not establish drift correction. Subject 12 was at chance (AUC 0.50); subject 11 had one test block. Task-event labels supplied response times that a real BCI would not know.

stereo-EEG · brain-computer interface · key-press decoding · linear discriminant analysis · neural signal drift · recalibration · high-gamma power · OpenNeuro ds004473