Full beta covariance gave lower mean AUC than contrast powers in eight SEEG patients with normalized power and 50% shrinkage

We tested whether beta covariance improves the separation of baseline and keypress windows in later SEEG trials. We used 64 selected trials from each of eight patients in OpenNeuro ds004473, version 1.0.2. The first 32 trials trained each patient model. The last 32 trials tested it. We used fixed Helmert contrasts, total beta power normalization, and 50% diagonal shrinkage. Full covariance gave a mean test AUC of 0.6598. Contrast powers gave 0.6863. The mean paired difference was -0.02649. Its 95% patient-bootstrap interval was -0.06348 to +0.00635. Three patients improved and five declined. The uncertainty does not establish a population benefit or loss. All five fixed sensitivity checks had negative mean differences. We checked all 1,024 public EDF ranges and reproduced the numerical analysis. The package preserves source data, source versions, retrieval code, model code, split records, scores, model matrices, figures, and repeat instructions.

stereoelectroencephalography · SEEG · beta covariance · contrast power · AUC · forward-time classification · OpenNeuro