With clean endpoints, pair correction gave higher smoothed 70–110 Hz power error than interpolation for simulated 3.17 ms gaps in 100 ms SEEG windows from 18 patients

We tested valid-pair correction for short power estimates when periodic gaps remove SEEG samples. The test used 735 channels from 18 patients in OpenNeuro ds006519, version 1.0.0. We added simulated 13-sample gaps at about 50 Hz to baseline and scored 100 ms windows with a common smoothed 70–110 Hz kernel. Observed samples and gap endpoints were clean as recorded. Pair correction gave higher normalized squared error than interpolation in all patients: group values were 0.201090 and 0.000435. The mean paired difference was 0.200655, with a patient bootstrap 95% interval of [0.177296, 0.222969]. Pair correction had near-zero signed error but returned 727 negative powers out of 29,400. Fixed gap-width and window-length checks kept the same ordering. An independent stationary Gaussian check showed zero expected bias with higher squared error than interpolation. This benchmark adds no new estimator. It tests a finite power target on artificial baseline gaps, not a live decoder or actual stimulation recovery. The package preserves exact source excerpts, raw and metadata hash checks, all window powers, code, four figures, and repeat instructions. A full second run reproduced all 23 numeric result files at the byte level.

SEEG · intracranial EEG · signal processing · power estimation · missing samples · valid-pair correction · brain-computer interface