Shuffled controls show that beta power adds key press decoding information in eight SEEG subjects

This re-analysis asks whether adding 13–30 Hz beta log power to 70–150 Hz high-gamma (HG) log power improves offline, cued key-press versus rest decoding in eight SEEG subjects, and whether the gain reflects beta information rather than extra features. It uses the same subjects and archived feature files as the earlier publication, with training on trials 1–100 and testing on later trials. New code used pooled-covariance Ledoit–Wolf linear discriminant analysis (LDA), L2 logistic regression, and an equal-size shuffled-beta control. With LDA, HG plus real beta increased mean AUC by 0.033 over HG alone (6/8 subjects; exact sign-flip p = 0.0195). Logistic regression showed a gain of 0.0556 (8/8; p = 0.0039), although its HG-only baseline was weaker. Adding shuffled beta decreased mean AUC by 0.0258 (7/8 lower; p = 0.0156); real beta exceeded shuffled beta by 0.0591 (8/8; p = 0.0039). Beta alone did not outperform HG. The pooled-LDA AUCs were lower than the earlier per-class LDA values in all eight subjects, while subject rankings were almost unchanged. This is a re-analysis, not a replication on new subjects: the data are the same eight subjects and feature files. Decoding used offline, fixed, cued windows in one session per subject; the pre-trial rest window includes anticipation. All subjects were epilepsy patients with clinical electrode coverage, sub-11 had only 32 test trials, the analysis was not preregistered, and hyperparameters were not tuned. The shuffled control removes beta structure across windows but cannot distinguish task-specific beta information from general state effects such as attention.

SEEG · beta power · high-gamma power · brain-computer interface · key-press decoding · AUC · shuffled control · replication