A fixed high-gamma–beta ratio does not improve SEEG key-press detection and underperforms free two-band weights

In offline, cued key-press detection from stereo-EEG (SEEG), a fixed spectral-tilt feature (70–150 Hz high-gamma log power minus 13–30 Hz beta log power) did not improve transfer AUC over high-gamma alone in eight subjects (mean difference −0.004; 4/8 positive; exact one-sided sign-flip p=0.57). Tilt underperformed a decoder with free high-gamma and beta weights (mean difference −0.044; 6/8 lower; two-sided p=0.031), with the same direction in five-fold cross-validation (mean −0.045; 7/8 lower; p=0.016). In the two-band decoder, high-gamma and beta weights had opposite signs in 49% of channels, consistent with a fixed +1:−1 ratio being unsuitable for many channels. The preregistered tests used archived features from OpenNeuro ds004473 version 1.0.2 and shrinkage LDA, trained on trials 1–100 and tested on trials 101 and later. This is an offline analysis of one task and dataset with eight subjects; it does not test closed-loop detection or other tilt definitions. The comparison with the earlier two-band result uses the same features and code and is not an independent replication.

stereo-EEG · SEEG · key-press detection · high-gamma · beta band · spectral tilt · shrinkage LDA · ROC AUC