Time blocks reduced later beta SEEG Gaussian score loss versus Ledoit-Wolf in 21 people

This study tested how to set a covariance model from a short beta SEEG recording. The test used 35 public recordings from 21 people in OpenNeuro ds006519, version 1.0.0. All model fits and setting choices used 1.5 seconds of early baseline only. Calibration with time blocks reduced Gaussian score loss versus Ledoit-Wolf in separate later windows. The mean post-window gain was 1.018 nats per channel, with a 95% person bootstrap interval of 0.582 to 1.548. Twenty of 21 people had a positive gain. Random-row calibration kept 99% of measured correlation in all records and had higher later loss. Time blocks kept 0% to 50%. The post-window gain over fixed 10% correlation was small and uncertain: 0.0055 nats per channel, with an interval of -0.0056 to 0.0160. The result supports conservative covariance calibration for this short beta signal model. It does not establish task decoding accuracy or online BCI performance. An inherited quality mask used one flat post-window channel. The data have no independent confirmation set. The package contains source code, compact voltages, fitted models, person results, three figures, source hashes, and replay instructions. All 105 raw files passed hash checks. All 35 rebuilt signal archives matched exactly. A second full analysis reproduced the results.

SEEG · brain-computer interfaces · covariance estimation · Gaussian score · time-series validation · signal processing