Disjoint SEEG pairs gave 53.1% mean minimum four-site accuracy, compared with 53.8%, in eight STReEF patients. Tests used four pairs, 20–100 ms features, and up to one known simulated contact loss.

We compared two rules for four SEEG recording pairs in eight public STReEF patients. One rule prevented shared contacts. The other chose the four largest fit separation scores. Each patient had four stimulation sites and 20 separate test pulses. Tests used 20–100 ms response features. Each test removed up to one known recording contact and kept surviving measurements unchanged. The mean minimum accuracy was 53.125% with disjoint pairs and 53.750% with the largest-score rule. The paired mean difference was −0.625 percentage points. A descriptive patient bootstrap interval was −5 to +5 points. Two patients improved, four worsened, and two had no change. Disjoint pairs guaranteed three surviving pairs. This count guarantee did not improve the observed mean minimum accuracy. They used eight physical contacts; the comparison used 7.125 on average. The package preserves compact inputs, source versions, code, tables, three figures, and reproduction instructions. Independent checks confirmed eight selection optima and all 20,240 predictions. A fresh replay rebuilt all 33 input files from 320 pinned raw epochs. These reused clinical records and simulated omissions do not establish live BCI performance or useful effects of stimulation.

stereoelectroencephalography · brain-computer interface · bipolar recording pairs · contact loss · robustness · site classification