Across-channel mean removal was non-inferior for offline SEEG movement decoding in 15 patients

In archived 70–150 Hz high-gamma amplitude features from 15 patients in OpenNeuro ds006253, an offline shrinkage LDA distinguished movement from rest with mean held-out AUC 0.833 using all channels, 0.824 after subtracting the across-channel mean in each window (CMR), and 0.577 using only that mean (CM). CMR met the pre-registered non-inferiority criterion versus all channels (margin 0.02 AUC; mean difference −0.009, 95% bootstrap CI −0.020 to −0.001; exact one-sided sign-flip p = 0.027), while the confidence interval indicates a small AUC cost. All channels outperformed the mean-only decoder in 14 of 15 patients (mean difference 0.256, Holm-adjusted p = 0.0002); the mean-only decoder was above chance in 13 of 15 (Holm-adjusted p = 0.0042). In an exploratory simulation, an equal ±3 dB offset to all test channels reduced the fixed-threshold balanced accuracy for the standard features from 0.759 to 0.586–0.592; CMR remained at 0.759. The result agrees with prior work in a second SEEG dataset and task. The analysis is offline, uses one session per patient and task windows aligned to known events, compares different task phases, simulates equal shifts across channels, and uses only the available 70–150 Hz features; it does not establish performance for causal online detection or channel-specific drift.

SEEG · movement decoding · high-gamma amplitude · common-mode removal · brain-computer interface · shrinkage LDA · gain robustness