In 8 SEEG patients, ERP covariance decoders had lower AUC than voltage decoders, while high-gamma covariance added no gain

We compared seven offline decoders of left versus right key press from the 0.4 s before the press in eight patients from OpenNeuro ds004473. All decoders used the same cross-validation folds; voltage t statistics selected channels within training folds. Shrinkage LDA on 0–30 Hz voltage bins from the same top eight bipolar pairs had mean AUC 0.657, compared with 0.595 for an ERP-covariance tangent-space decoder. The paired difference was −0.062 (range −0.206 to 0.000), with zero of eight patients showing a positive difference (exact two-sided sign-flip p = 0.008; Bonferroni-adjusted p = 0.048 across six contrasts). High-gamma spatial covariance did not improve on high-gamma log variance (mean AUC 0.583 versus 0.584; p = 0.875). Fifty-millisecond voltage bins did not improve on a single mean voltage. The mean-voltage baseline reproduced the prior analysis exactly (mean AUC 0.659). Results are limited to eight patients, one offline pre-press window, and one Riemannian variant per signal type. In correct trials cue side and key side are confounded, so the decoders may use visual or motor information. The live dataset mirror's exact release is uncertain; source URLs, ETags, and hashes are preserved.

stereo-EEG · SEEG · brain-computer interface · key press decoding · Riemannian geometry · tangent space · covariance · shrinkage LDA · high gamma · ROC AUC