Known background templates increased mean AUROC from 0.665 to 0.812 for withheld SEEG shafts in five offline sessions with eight output sites

We tested whether templates from known background sites helped reject responses from a whole stimulation shaft absent from deployed fitting. The public SEEG data came from five sessions in OpenNeuro ds004457 version 1.0.2. Each of 38 shaft folds selected eight output sites from three early responses per known site. Separate fourth responses set fixed thresholds. Final four responses tested the readers. Mean AUROC increased from 0.665 to 0.812. The paired gain was 0.147, with a descriptive 95% patient bootstrap interval of 0.059 to 0.230. All five patients had a gain, but the exact two-sided sign-flip test gave p = 0.0625. Mean withheld-response rejection increased from 15.8% to 56.9%. Correct and accepted selected readings decreased from 74.1% to 72.2% of all selected attempts. A diagnostic that added withheld-site calibration templates gave AUROC 0.933 and rejection 76.9%. The deployed rule still accepted 43.1% of withheld responses. This was a retrospective test of reused data. It does not establish live BCI performance. The package includes three figures, source code, fixed models, compact voltage data, source hashes, and a raw retrieval recipe. Direct numerical checks, an exact replay, and a raw-data rebuild passed.

stereoelectroencephalography · brain-computer interface · template classification · response rejection · withheld-site validation