Fitting at two times reduces mean Brier error at an excluded middle time in later SEEG trials from eight recordings

We tested whether fitting one SEEG reader at two input times reduces probability error at a middle time excluded from fitting. We used all eight recordings and 2014 retained trials from public OpenNeuro ds004473 version 1.0.2. Each reader used eight fixed bipolar pairs. Selection and scaling used nominal training inputs only. Both fits had equal total training weight. They learned from earlier trials and were tested on 674 later trials. At the middle time, mean Brier error fell from 0.2244 to 0.2073. The paired mean decrease was 0.0171 (descriptive 95 percent person-bootstrap interval 0.0058 to 0.0294). Seven of eight people had lower error. The accuracy and AUC changes remained uncertain. These are overlapping task-state windows within reused clinical recordings. The result does not establish continuous intention reading or a stimulation effect. The package preserves new features, pinned source retrieval records, models, predictions, figures, and independent checks.

stereoelectroencephalography · SEEG · brain-computer interface · logistic regression · Brier score · time-window robustness · probability error