Separate beta and high-gamma SEEG decoders were non-inferior to joint decoding in 8 subjects

We compared late fusion of separate 13–30 Hz beta (B) and 70–150 Hz high-gamma (HG) shrinkage LDA decoders with early fusion in archived bipolar SEEG features from 8 subjects performing cued key presses. The features came from OpenNeuro ds004473 snapshot 1.0.2; calibration used trials 1–100 and testing used trials 101 onward. Late fusion added the two decoder scores after scaling each by its training pooled within-class standard deviation, with equal weights and no tuning. Late fusion was non-inferior to early fusion at a 0.02 AUC margin (mean difference +0.0022; all 8 subjects above −0.02; exact one-sided sign-flip p = 0.0039). It exceeded HG alone by +0.0420 AUC in 7 of 8 subjects (one-sided p = 0.0078). Within-class HG and B score correlation was weakly positive (mean r = 0.20; range 0.02–0.36), consistent with mostly separate trial-to-trial variation. With calibration on trials 1–30, late and early fusion were again similar (mean difference −0.0004; two-sided p = 0.96). The Gaussian equal-variance check was not an independent prediction because it used d′ and correlation from the same test trials. This offline study used one session per subject, cued presses, 0.5 s windows, 2 bands, one decoder type, and equal untuned fusion weights in 8 epilepsy patients with clinical SEEG coverage; secondary contrasts were not corrected for multiple tests.

stereo-EEG · brain-computer interface · late fusion · early fusion · high-gamma · beta band · shrinkage LDA · key-press detection · OpenNeuro