Log band contrasts before averaging canceled separate positive scalar window gains; Brier error did not clearly decrease versus pooling in 15 SEEG tests at C = 0.1 under assumed timing

We tested three-band log power features for session transfer in public SEEG. The data were OpenNeuro ds004752 version 1.0.1, with 15 people, 607 training trials, and 630 test trials. The first numbered session trained each reader. The last numbered session tested it. At C = 0.1, pooled band contrasts had mean Brier error 0.2313 versus 0.2418 for full log power. The paired difference was -0.0105, with a 95 percent person bootstrap interval of [-0.0312, 0.0080]. Error rose in eight people. The constant training class rate had lower mean error, 0.2135. A later proof and test showed that taking log contrasts within each window, then averaging, cancels separate positive scalar window gains. Pooling power first does not give this invariance when delay gains differ. A twofold gain in one delay window changed pooled-reader probabilities by up to 0.1795. Window contrasts stayed invariant to numerical precision. Their Brier difference from pooled contrasts was -0.0008, with interval [-0.0035, 0.0020]. This later score test was exploratory. The cohort was reused and phase timing was assumed. The package has code, data, three figures, fitted models, trial scores, source hashes, and reproduction steps. Fresh source checks covered all 1467 trials in 30 sessions. Independent checks passed for all 210 reader fits.

SEEG · brain-computer interface · band power · session transfer · Brier error · gain invariance · verbal working memory