In a second SEEG dataset, 8 ranked channels retained 95% of movement detection gain, while random subsets required 64

Question: does the SEEG channel-count curve of sci d44dfe93 (ds004473, 8 subjects) replicate in a different dataset? We used the archived 70-150 Hz features of 15 ds006253 patients (sci ceba42cd, 50-148 bipolar channels) and the unchanged ceba42cd detector: shrinkage LDA, MOVE vs REST windows, first 60 % of trials for calibration and last 40 % for test. For N = 1 to 64 channels, we used 200 random subsets per N, or the N channels with the largest Welch |t| on calibration trials. We wrote and hashed the plan before any decoding. Gain G(N) = (AUC(N)-0.5)/(AUC(all)-0.5). In the 13 patients with all-channel AUC 0.75 or more, the median random G was 0.13 at N=1, 0.48 at N=8, 0.81 at N=32 and 0.96 at N=64 (d44dfe93: 0.17, 0.50, 0.85, 0.96). Both pre-registered predictions passed (G(32) < 0.95; G(8) in 0.35-0.65). The AUC increased by a median of 0.042 per doubling of N. Ranked channels gave G = 0.67 at N=1 and 0.948 at N=8. At N=8, ranked channels were better than random channels in 15 of 15 patients (mean +0.159 AUC, exact p = 3.1e-5). Mean all-channel AUC was 0.833. Thus, in two SEEG datasets, randomly placed contacts need about 64 channels, but about 8 selected contacts keep about 95 % of the detection margin. Unlike d44dfe93, high gamma alone reached 95 % at 64 random channels. Limits: offline, one session, task-locked windows, clinical placement, one ranking rule, a 2x grid of N. The package has the plan, code, features with hashes and provenance, results and a figure.

stereo-EEG · brain-computer interface · movement detection · channel selection · high-gamma activity · linear discriminant analysis · replication