In 8 SEEG subjects, calibration-range clipping recovered 87% of AUC loss from simulated single-channel artifacts

We evaluated shrinkage linear discriminant analysis (LDA) for detecting cued key presses from stereo-EEG (SEEG) high-gamma (70–150 Hz) and beta (13–30 Hz) log-power features in 8 subjects. Each subject’s first 100 trials (200 rest and key-press windows) formed the calibration set; later trials formed the test set. We simulated additive shifts of k calibration standard deviations to both bands on a random bipolar channel in test windows with probability r, and compared plain LDA with a label-free guard that clips each feature to its calibration minimum and maximum. For the primary condition (r = 0.10, k = +10), artifacts lowered mean AUC by 0.0055 (95% bootstrap CI, 0.0045–0.0064; all 8 subjects; exact one-sided sign-flip p = 0.0039). Clipping recovered 0.0048 AUC, or 87% of that loss (95% CI, 0.0032–0.0066; all 8 subjects; p = 0.0039). Clean-data AUC was 0.814 for plain and clipped LDA; the difference was −0.00004 (95% CI, −0.0013 to +0.0015), meeting the prespecified 0.02 non-inferiority margin (p = 0.0039). Across tested conditions with k = 10 or 20, clipping recovered 70–94% of artifact-related loss. Artifacts were simulated as additive log-power shifts in offline test windows, and the study included one session per subject; these results do not establish performance for real artifacts or other settings.

stereo-EEG · SEEG · brain-computer interface · artifact robustness · shrinkage LDA · signal processing · feature clipping · AUC