In 8 SEEG subjects, most changes in top-8 high-gamma channels reflected sampling noise, with no reliable within-session drift

Question: a BCI that reads few channels usually selects them one time, from the first trials. Do the best SEEG channels drift during a session? Data: OpenNeuro ds004473 v1.0.2, 8 SEEG subjects, cued key press, archived features of sci 07478163 (same SHA-256). Rest vs move, log 70-150 Hz power (primary) and 13-30 Hz power, adjacent bipolar channels (100-112 per subject). Method (plan hashed before analysis): rank channels by |Welch t| on trials 1-50 (EARLY) and 51-100 (LATE). Compare the top-8 overlap with 2000 random 50/50 trial splits. Train shrinkage LDA on EARLY, test on LATE. Exact sign-flip tests over 8 subjects. Results: top-8 high-gamma overlap was 0.41 for the time split and 0.46 for random splits (chance 0.08). The primary test gave p = 0.039 (6/8 subjects lower). But the effect was small (about 0.4 of 8 channels), and it did not occur at k = 4 or 16, in the rank correlation, in beta, or in any single subject. 8 channels selected on EARLY gave AUC 0.782 on LATE, against 0.798 for all channels and 0.617 for 8 random channels. Non-inferiority to all channels at a 0.02 margin was not shown (p = 0.38; losses of 0.06-0.10 in 2 of 8 subjects). Channels selected on LATE gave +0.044 AUC, but an exploratory control gave the same gap for random splits (+0.043). Thus this gap is selection bias, not drift. Conclusion: within one session of 100 trials, we found no reliable drift of the best channels. With 50 trials, about half of the top-8 channels change between any two halves of the data. Thus a short calibration, not time, limits a fixed reduced channel set. Limits: 8 subjects, one session, offline, rest vs move windows, no test across days.

stereo-EEG · brain-computer interface · channel selection · sampling noise · high-gamma power · shrinkage LDA · key-press detection · within-session drift