Common-mode removal was non-inferior for offline SEEG key-press AUC and invariant to simulated uniform gain in eight patients

In eight patients performing a cued key-press task, we tested whether subtracting the across-channel mean of high-gamma (70–150 Hz) and beta (13–30 Hz) log power in each window preserved offline SEEG decoding and protected a fixed threshold from uniform gain changes. A shrinkage linear discriminant was trained on trials 1–100 and tested on later trials using archived features from OpenNeuro ds004473 v1.0.2. Mean AUC was 0.828 with standard features and 0.824 after common-mode removal. The mean AUC difference (common-mode minus standard) was −0.004 (95% t-CI −0.017 to +0.009; 3 of 8 subjects improved; sign-flip p = 0.64), meeting the prespecified non-inferiority margin of −0.02. The two across-channel means alone had mean AUC 0.597 (7 of 8 subjects above 0.5; one-sided sign-flip p = 0.008). With simulated uniform amplitude gain applied to test channels, balanced accuracy at the trained threshold changed from 0.758 at ×1.0 to 0.681 at ×0.8 and 0.673 at ×1.25 for standard features. With common-mode removal it was 0.755 at each gain, invariant by construction. The ×0.8 comparison was not significant (p = 0.22), while the ×1.25 comparison was significant (p = 0.031). These offline results suggest that the detector relied mainly on cross-channel patterns and that common-mode removal can protect its fixed threshold from uniform gain changes. The study used eight patients, one session, 0.5 s windows, one linear decoder, and simulated uniform gain; it does not establish robustness to real or channel-specific drift.

SEEG · common-mode removal · key-press detection · band power · shrinkage LDA · uniform gain · brain-computer interface