Correction from up to 20 past trials increased mean SEEG recall forecast error in 13 patients under assumed source trial order

We tested whether feedback from completed trials improves recall error forecasts from a fixed SEEG reader. The source was OpenNeuro ds006136 version 1.0.1, with 13 patients and 14 sessions. We assumed that source trial IDs gave task order. Earlier data trained the readers. The test used 751 later trials. A correction fitted one probability offset with up to 20 available test outcomes and penalty 2. It kept signal weights fixed. Each outcome arrived before the next matched trial. Mean patient Brier loss increased from 0.179663 to 0.184837. The increase was 0.005173, with a 95% patient bootstrap interval of [0.002305, 0.008018]. Loss increased in 11 of 13 patients. A reader based on task history and channel availability showed a similar increase. The corrected SEEG reader did not establish a gain over corrected task history. This is an exploratory extension of a known group and known fixed forecasts. Source selection remains offline. No stimulation or live interface was tested. The package preserves figures, features, forecasts, code, source hashes, and replay instructions. This version corrects the archive file hash list. The numerical results are unchanged.

stereoelectroencephalography · working memory · recall forecasting · probability calibration · Brier loss · outcome feedback · exploratory analysis