Removing future samples from a 5,000 microvolt screen reduced mean balanced accuracy on later trials by 0.11 percentage points for eight fixed SEEG readers

We tested whether future voltage samples change the test set for a fixed SEEG reader. Public data came from eight people in OpenNeuro ds004473 version 1.0.2. We reused eight saved state readers. We recovered all 18 trials rejected for peak-to-peak voltage above 5,000 microvolts. Every raw range matched its archived hash. A screen that used only available windows restored 12 examples among 2,032 synchronized trials. Four examples were in the later test: three rest windows and one window before response. Equal-person mean balanced accuracy changed from 70.686% to 70.574%. The change was -0.112 percentage points, with a descriptive 95 percent person bootstrap interval of [-0.277, 0.000]. Class-balanced Brier error changed from 0.190139 to 0.191278. Six records had no score change. Added-window counts bounded the possible mean accuracy change between -0.245 and +0.097 percentage points when added outputs were unknown. The package contains recovered raw ranges, power features, a reusable screen function, fixed models, test outputs, three figures, code, and checks. Window endpoints still use known task events. This reused offline audit does not validate a live BCI or test stimulation.

stereoelectroencephalography · brain-computer interfaces · signal quality screening · future-sample leakage · balanced accuracy · Brier error · fixed classifiers