Eight bipolar signals selected on early trials gave higher mean key-press AUC than random sets on later trials in eight SEEG sessions

We tested eight bipolar SEEG signals selected from early trials for response periods versus rest in later trials. We used 1,998 trials from eight patients in OpenNeuro ds004473, version 1.0.1. Mean later-trial ROC AUC was 0.836 for the selected set, 0.713 for 100 random sets per patient, and 0.859 for all available signals. The selected-minus-random difference was 0.123, with a 95 percent patient bootstrap interval of 0.078 to 0.174. The selected-minus-all difference was -0.022, with an interval of -0.042 to -0.004 and an exact sign-flip p value of 0.07031. This does not establish equivalence. The rest-versus-rest control had mean AUC 0.505. A post hoc check removed generic C-number pairs and left the selected-set AUC unchanged. The model used only early trials for selection, scaling, and nested parameter choice. The response window included 0.5 seconds after the key press. This is an offline task-state test within one session per patient. It does not isolate motor intent or test stimulation. The package preserves features, trial times, fitted models, predictions, figures, code, source metadata, full-file hashes, and reproduction instructions.

stereo-EEG · bipolar signals · brain-computer interface · ROC AUC · nested cross-validation · signal selection · task-state decoding