A baseline covariance filter did not show lower mean squared signal change in 20 short SEEG tests

We tested a baseline covariance rule for recording during brain stimulation. We used one short SEEG train from each of 20 people in OpenNeuro ds005169 version 1.0.0. The first 24 pulses fitted one spatial direction. The first 24 baseline epochs fitted covariance with fixed shrinkage 0.10. The next 24 baseline epochs tested squared signal change. The next 24 pulses tested repeatable class waveforms. Both filters removed the same fitted direction exactly. The orthogonal filter changed 1.240 percent of test baseline energy on average. The covariance filter changed 1.343 percent. The planned mean advantage was -0.103 percentage points. Its descriptive 95 percent person bootstrap interval was -0.666 to 0.282 points. The exact sign-flip probability was 0.865715. The test did not establish a mean benefit. The covariance filter also left a larger later pulse class measure in 19 people. An exact proof explains why a fitted covariance optimum need not transfer to test data. All 140 source files passed fresh hash checks. Fresh extraction matched all 20 epoch files. A repeat matched all 24 numeric files. The package preserves data, models, code, three figures, and reproduction instructions. This is an exploratory extension in a reused epilepsy cohort. It does not establish neural recovery or live interface benefit.

intracranial EEG · SEEG · spatial filtering · covariance · brain-computer interfaces · electrical stimulation