In eight SEEG subjects, beta power did not lose more key-press detection AUC than high-gamma power with shorter causal windows

This offline analysis tested whether 13–30 Hz beta power loses more key-press detection AUC than 70–150 Hz high-gamma power as causal analysis windows shorten. We analyzed 100 cued key-press trials in each of eight SEEG subjects from OpenNeuro ds004473 v1.0.2, using causal fourth-order Butterworth filters, windows of 0.05–0.8 s, shrinkage LDA, five-fold trial-level cross-validation, and exact sign-flip tests. The preregistered one-sided comparison from 0.5 s to 0.1 s was not supported: the mean difference in AUC loss was −0.014, with a positive difference in 3/8 subjects (p = 0.63). Mean high-gamma AUC rose from 0.65 at 0.05 s to 0.78 at 0.3 s, then was 0.77 at 0.5 s and 0.74 at 0.8 s. At 0.05 s, high-gamma lost more than beta in 7/8 subjects, descriptively. Adding beta to high-gamma raised AUC by 0.063 at 0.05 s (7/8 subjects, uncorrected p = 0.023); this did not remain significant after Holm correction. The median subject retained at least 90% of 0.5-s high-gamma performance at 0.2 s. This was an offline cued task: the movement window ended at key press, and the rest window was in the pre-cue fixation period. The study included eight subjects, two near chance, and did not test an asynchronous real-time detector. Beta and high-gamma filters had different group delays, which also affects the window-length comparison. All reported results use the corrected v2 analysis; the archived v1 output contains a filter start-up artifact and is explicitly invalid.

stereo-EEG · SEEG · brain-computer interface · key-press detection · causal filtering · high gamma · beta power · window length · ROC AUC · shrinkage LDA