In eight SEEG subjects, HG+beta decoder gains varied beyond test-trial noise and required a median of 351 trials for ±0.02 AUC precision

In eight subjects performing a cued key-press versus rest task, we measured whether test-trial sampling noise explained variation in the area under the ROC curve (AUC) gain from adding beta-band power (13–30 Hz) to high-gamma power (70–150 Hz). Archived OpenNeuro ds004473 features were used with a fixed shrinkage linear discriminant decoder trained on trials 1–100 and evaluated on trials 101 onward. Paired bootstrap resampling of whole test trials (2,000 resamples) gave Cochran Q = 35.2 (7 df, p = 1.0 × 10⁻⁵), I² = 0.80, and between-subject tau = 0.024 AUC for the HG+beta minus HG contrast; the observed between-subject SD was 0.040, versus RMS sampling SE 0.021. The random-effects mean gain was +0.032 (SE 0.011), while the precision-weighted mean was +0.012. Five of eight subjects had z > 1.96 and none had z < −1.96. A median of 351 test trials per subject (range 47–1,432) was estimated for a 95% CI half-width of 0.02 AUC. At the observed heterogeneity, power for a true mean gain of 0.02 was 0.48 at observed test sizes and 0.58 with 400 trials per subject. All 32 recomputed AUCs matched the archived results. Beta versus HG and top-eight versus all-contact HG contrasts also varied between subjects. The exploratory association between lower HG AUC and larger gain (Spearman rho = −0.86, exact permutation p = 0.011) may reflect ceiling effects; the gain contrast also contains −AUC(HG). This is one dataset with eight subjects and one session each. The decoder was fixed, so calibration-set noise was not included; the trial bootstrap assumes independent trials despite possible within-session drift. Consequently, I² may overstate spread beyond sampling noise, and tau is imprecise.

stereo-EEG · brain-computer interface · high-gamma power · beta power · decoder performance · area under the ROC curve · subject heterogeneity · trial bootstrap · statistical power