In 8 SEEG subjects, removing shared HG–beta noise increased rest–key-press AUC in 7 subjects

We tested whether trial-to-trial noise shared by 70–150 Hz high-gamma (HG) and 13–30 Hz beta power at the same SEEG bipolar pair explains the gain from adding beta features to an HG movement decoder. In eight subjects from OpenNeuro ds004473, shrinkage LDA trained on trials 1–100 and tested on later trials gave a mean HG+beta AUC gain of 0.040 (6/8 subjects, p = 0.016). A within-class permutation of the complete beta feature block removed HG–beta cross-band noise correlation in expectation while retaining within-band covariance and class means. Contrary to the planned hypothesis, this removal did not lower AUC: mean AUC(real) minus AUC(removed) was −0.020 (1/8 positive, one-sided p = 0.996), and removal raised AUC in 7/8 subjects (two-sided sign-flip p = 0.016, post hoc direction). Median within-class noise correlation averaged 0.098 (6/8 positive; one-sided p = 0.059; Holm p = 0.12), with large values in only two subjects. At single pairs, signal-aligned noise correlation was positive in all eight subjects, but its effect on single-pair discriminability was near zero. The multivariate AUC increase after removal may reflect cross-pair HG–beta covariance or easier covariance estimation; this analysis does not distinguish them. Results are limited to one dataset and task, eight subjects, fixed 0.5 s windows, log band power, one train/test split, and an exchangeability assumption for trials within each class.

stereo-EEG · movement decoding · high-gamma power · beta power · noise correlation · shrinkage LDA · AUC · OpenNeuro