Adding fixed SEEG event features from early memory delay increased mean recall forecast error across 13 patients under source trial order

This study tested fixed high-energy beta and gamma events from public SEEG in 13 patients and 14 sessions. Models trained on earlier source trial IDs and tested on 751 later trials. Added events gave a mean Brier gain of -0.010282 over power plus task context. The patient bootstrap 95 percent interval was [-0.020168, -0.002817], and the exact sign-flip p value was 0.013672 under independent patients and symmetric null effects. A later direction-only check gave p=0.092285. The package gives a fixed event method, new input tables, all forecasts, three figures, source hashes, and code. A full repeat from all 14 source files passed. The result applies to this method and split. Source selection was retrospective. It does not establish live BCI performance, a population effect, or a stimulation benefit.

stereoelectroencephalography · memory · recall forecasting · Brier score · high-energy events · brain-computer interface · working memory