A donor ridge prior changed mean Brier error from 0.297 to 0.299 after 16-trial calibration in eight SEEG recordings

We tested whether coefficients learned from other people reduce later SEEG prediction error after 16 local labeled trials. Public OpenNeuro ds004473 v1.0.2 supplied eight recordings and 674 later test trials. Both ridge fits used the same power summaries, donor input scale, labels and penalty 0.1. Only the coefficient center changed. Mean Brier error was 0.296869 for the local center and 0.298508 for the donor center. The change was +0.001639. Lower is better. The descriptive 95% interval from 4,000 full-refit person bootstrap draws was -0.000374 to +0.003926. It includes zero. Only two people had lower error with transfer. Both observed mean errors exceeded the constant-probability error of 0.25. This result applies to one fixed method in eight reused records. It does not establish equal performance. The package preserves exact inputs, source versions, raw-range checks, code, fitted states, predictions, bootstrap draws, figures and instructions. Independent spectral and solver checks passed. A full repeat gave identical numerical files.

stereoelectroencephalography · brain-computer interface · ridge regression · transfer learning · Brier score · calibration · power spectra · person-level bootstrap