In 8 SEEG subjects, 20-trial channel selection was not shown non-inferior to 100 trials, and channel-map reliability stayed below 0.9 at 100 trials

A SEEG BCI must select channels from a short calibration. We used the archived per-trial 70-150 Hz features of 8 subjects in OpenNeuro ds004473 (sci 04683a2a). We wrote and hashed the plan before the analysis. We selected the top 8 bipolar channels by movement effect size from the first n trials. We trained shrinkage LDA on trials 1-100 and tested it on trials 101 and after. Primary result: selection from 20 trials was not shown non-inferior to selection from 100 trials (mean AUC difference -0.014, 6 of 8 subjects in margin, exact sign-flip p = 0.37). The data also did not show a loss (p = 0.23). Two subjects lost 0.065 and 0.105 AUC. With 40 trials, the mean difference was -0.001 (7 of 8 in margin, p = 0.10). The full map of 100-112 channels was not reliable: the median split-half Spearman r at 100 trials per half was 0.54, and no subject reached 0.9. Spearman-Brown extrapolation gives 222-2239 trials per half for r = 0.9. The 8 channels from 100 trials shared only half of the best 8 test-set channels, but the AUCs were almost equal (0.741 vs 0.754). Thus many channels carry redundant movement information. A stable channel ranking is not necessary for good decoding. Limits: offline cued presses, one session, 8 subjects, fixed windows, 70-150 Hz only, low test power. The package contains the code, the features with SHA-256 values, the plan, the results (JSON, CSV), one figure, and reproduction steps.

SEEG · brain-computer interface · channel selection · calibration · high-gamma power · linear discriminant analysis · non-inferiority · split-half reliability