In a second SEEG dataset, 8 random bipolar channels reached 48% of all-channel AUC gain; ranked channels reached 93%

This offline replication tested whether the SEEG channel-count curve reported for OpenNeuro ds004473 also appeared in ds006253, a second center and task. We used archived 70–150 Hz features from 8 ds004473 subjects (key press versus rest) and 7 ds006253 patients (movement versus pre-stimulus rest), with one time-ordered train/test split per subject and shrinkage linear discriminant analysis. For channel counts of 1, 2, 4, 8, 16, 32, and 64, we evaluated 200 random bipolar subsets per count and channels ranked by the absolute paired t statistic from training trials. Normalized gain was (AUC(N) − 0.5)/(AUC(all) − 0.5), summarized as the median among subjects with all-channel AUC ≥ 0.75 (5 per dataset). In ds006253, random channels reached median gains of 0.48 at 8 channels and 0.81 at 32; ds004473 values were 0.46 and 0.83. The preregistered replication rule (gain ≤ 0.70 at 8 and < 0.95 at 32 in ds006253) was met. Ranked channels reached 0.93 at 8 in ds006253. Ranked 8-channel AUC exceeded random 8-channel AUC in 7/7 ds006253 patients (mean difference +0.140, exact sign-flip p = 0.0078) and all 15 subjects (+0.152, p = 3e-5). Median random-channel AUC gain per doubling from 1 to 32 channels was 0.041 for ds004473 and 0.037 for ds006253 (permutation p = 0.83). Random subsets reached gain ≥ 0.95 at 64 channels for 2/5 included subjects in each dataset, or did not reach it. This is a small offline analysis of archived features, with one split per subject; ds006253 sub-04 had 49 channels, so the 64-channel result included four patients. The task comparison was movement versus rest, and channel ranking used training labels.

SEEG · stereo-EEG · bipolar channels · neural decoding · channel selection · high gamma · linear discriminant analysis · AUC · movement detection · replication · OpenNeuro