An eight-pair reader trained from SEEG ensembles reduces mean Brier error compared with half confidence in eight recordings

We tested public SEEG data from eight people in OpenNeuro ds004473 version 1.0.2. Frozen ensembles used the first 32 trials. Each ensemble had 256 logistic models and required 59–98 signal pairs. One new reader used eight fixed pairs and learned from ensemble probabilities on 52–165 middle trials per person. No middle-window class label entered its fitting loss. We tested 674 later trials. Mean Brier error decreased from 0.2299 for an eight-pair half-confidence control to 0.2147 for the new reader. The paired reduction was 0.0152 (95% person-bootstrap interval 0.0058–0.0246), with a reduction in seven people. The exact two-sided sign-flip p value was 0.0234 under independence and null sign symmetry. The large teacher had less mean error, at 0.2102. The new reader required 12–16 input contacts and one logistic evaluation per window. Broad input coverage was still required before transfer. Known events defined the windows, and source quality checks used later signal. This is an exploratory offline test on reused clinical recordings. It does not establish continuous intention detection or online BCI control. The package has code, features, frozen teachers, new fits, predictions, figures, source records, fresh raw checks, and full rerun checks.

stereoelectroencephalography · model distillation · brain-computer interface · logistic regression · Brier score