The beta mean shift, not HG-beta noise covariance, explains the 13-30 Hz gain in SEEG key-press decoding

In archived features from 8 SEEG subjects performing a cued key-press task, we examined why 13-30 Hz beta log power adds to 70-150 Hz high-gamma (HG) log power in a shrinkage linear discriminant analysis (LDA). The decoder was trained on trials 1-100 and tested on trials 101 onward. The full HG+beta model replicated an AUC gain of +0.040 over HG alone (6 of 8 subjects, sign-flip p = 0.016). Setting the beta mean shift to zero, so beta contributed through its covariance with HG, reduced AUC by 0.033 relative to HG alone (1 of 8 subjects higher, p = 0.95; primary endpoint not supported). The beta mean shift contributed +0.073 AUC (8 of 8, p = 0.0039). In an exploratory analysis, removing cross-band covariance did not reduce AUC (+0.005; non-inferiority margin 0.02, p = 0.0039). Same-channel within-class HG-beta noise correlation was near zero (median r = 0.020 across subject medians). These results indicate that beta helped as an independent signal, not as a noise reference, in this dataset and calibration setting. The zero-mean probe is not an optimal noise-covariate decoder; its result does not show that true covariance has no value. The study was offline, used one session, fixed bands, 0.5 s windows, zero-phase filters, and 8 epilepsy patients with clinical contacts. Other bands, causal filters, and tasks were not tested.

SEEG · electrocorticography · key-press decoding · beta power · high-gamma power · linear discriminant analysis · brain-computer interface · noise covariance