An SEEG model trained on fixation and early delay gave 51.9 percent mean balanced accuracy on fixation and late delay in 13 held-out patients with baseline calibration under source trial order

We tested whether SEEG power from other patients can separate fixation from late memory delay in a held-out patient. The fixed model used eight regional power inputs. It trained on fixation and early delay. An earlier fixation reference from each session set the scale. We assumed that source trial IDs gave task order. The test used public OpenNeuro ds006136 version 1.0.1, with 13 patients, 14 sessions, and 751 later test trials. Mean patient balanced accuracy was 51.9 percent. The descriptive patient resampling interval was 48.4 to 55.7 percent. A conditional whole-patient label-swap reference gave p = 0.157 after 999 joint refits. The test did not establish performance above that reference. The release was already processed and selected. The result does not isolate memory or test brain stimulation. The package preserves the plan, source hashes, source metadata, matched power tables, all models and predictions, three figures, and reproduction code. A full source rebuild gave identical input tables and numerical results.

stereoelectroencephalography · working memory · phase classification · brain-computer interface · cross-patient transfer · balanced accuracy