In 24 SEEG patients, a 3–5 min causal baseline reduced late high-gamma offset by 31%; drift was sublinear over 1–10 min

We studied whether SEEG band-power drift over minutes limits a calibration baseline, and whether a causal, label-free exponential moving average (EMA) reduces later offset. We analyzed archived 4-s log10 Welch band-power features, sampled once per minute, from 24 subjects in one session of the HUP Color Change Detection Task (4–8, 8–13, 13–30, and 70–150 Hz). For each pair, we scaled differences by a robust 1-min difference standard deviation. Across lags of 1–10 min, all subjects and bands had variogram slopes below the random-walk value of 1 (sign-flip p = 5e-5); median variogram growth was about 20% from 1 to 3 min and nearly flat thereafter. In leave-one-subject-out selection, a causal EMA with a 3–5 min half-life was the best grid point in all bands. For high gamma (70–150 Hz), the late-window held-out EMA-to-fixed-calibration median z² ratio was 0.69 (19/24 subjects, p = 0.0001), a 31% lower offset. Across bands, late ratios were 0.68–0.78 (20–22/24 subjects). A 3-window running mean was not reliably better than fixed calibration in high gamma (17/24, p = 0.50). The study covers one task and session per subject, 13–55 min sessions, and lags up to 10 min; the score measures offset, not decoding accuracy. An EMA can also remove slow task-related signals. The analyzed features trace to OpenNeuro ds005624; acquisition records point to live S3 files and do not establish an exact dataset release, so no version-specific dataset DOI is assigned.

SEEG · band power · baseline drift · exponential moving average · calibration · high gamma · signal processing · brain-computer interface