Covariance recovery with 0.5 shrinkage gave higher mean sound target error after known simulated contact loss in ten offline SEEG readers

We tested a fixed recovery rule for a sound reader after known simulated loss of one SEEG contact. The public data were OpenNeuro ds004703 version 1.1.0. The fixed sample had ten people, 95 passages, and 129 distinct contact loss cases. Missing inputs were estimated from surviving inputs with early training covariance and 0.5 diagonal shrinkage. Every selected bipolar pair that used the missing contact was removed. Mean normalized target MSE was 0.9598 with training mean replacement and 0.9723 with recovery. Each person had equal weight. The recovery benefit was -1.25 baseline percentage points. Its descriptive 95% person bootstrap interval was -2.88 to +0.16 points. The exact sign-flip p value was 0.1836. Target error rose in eight people. Recovery reduced deviation from the intact output in eight people, but six of these had higher target error. An exact error identity explains why output preservation alone does not ensure target accuracy. All 95 passages rebuilt from pinned public ranges. All arrays in 229 compact files matched. A full offline replay matched all saved results. The package supplies compact inputs, source recipes, tested readout code, fitted states, predictions, complete tables, and three figures. This is a reused clinical sample and an offline unavailable-input simulation. It does not establish a wider benefit or harmful effect, intended speech readout, neural signal origin, or live device performance.

stereoelectroencephalography · SEEG · speech envelope · brain-computer interface · covariance recovery · contact loss · sound amplitude · fault tolerance