With a common distance metric, a site mean projection kept labels but changed response rejection in five SEEG sessions

We tested a site mean projection on public SEEG responses from five sessions in OpenNeuro ds004457 v1.0.2. All candidates used a shared fitted distance metric. In the 20 to 500 ms window, the projection kept all nearest-site labels. It changed 165 of 5,576 repeated test rejection decisions from 732 distinct pulses. Adding one scalar for discarded signal energy restored all full distance scores and rejection decisions within numerical precision. The vector plus scalar reduced output value count by 83.5% on average across people. The encoder still needed the full signal. Mean unknown-shaft AUROC was 0.5647 for full distance and 0.5698 for the small map. The descriptive 95% person bootstrap interval for this change was -0.0048 to 0.0159. Detection changes were small and uncertain. The test was retrospective and used a previously studied cohort. The package includes verified raw provenance, code, models, compact packets, figures, and exact replay checks, and a separate environment replay. These results do not establish live BCI performance.

stereoelectroencephalography · brain-computer interface · response rejection · distance metric · site-mean projection · single-pulse electrical stimulation · unknown-shaft detection · signal compression