Added noise up to 40 nV/√Hz kept high-gamma decoding within a 0.02 AUC margin in 8 SEEG subjects

We added independent white Gaussian input noise to SEEG contacts from eight subjects performing a cued key press, then decoded movement versus pre-cue rest offline from bipolar 70–150 Hz power. A time-ordered 60/40 train-test split and shrinkage LDA gave a baseline mean test AUC of 0.789. With a non-inferiority margin of 0.02, added noise up to 40 nV/√Hz per contact was non-inferior (mean ΔAUC −0.004; one-sided 95% bootstrap lower bound −0.013; exact sign-flip p=0.016); at 80 nV/√Hz, mean ΔAUC was −0.023. Beta-band decoding remained non-inferior up to 80 nV/√Hz. Noise was added on top of the clinical recording noise. The result applies to this offline, cued task in eight epilepsy patients at one center; only one independent white-noise draw per subject and level was tested. The clinical amplifier noise is unknown, and 1/f, common-mode, and correlated noise were not tested.

SEEG · intracranial EEG · amplifier noise · high-gamma · movement decoding · brain-computer interface · signal-to-noise ratio