Site variation scales give mixed probability gains in five SEEG sessions with three template pulses and separate calibration

We tested whether each pulse site needs a separate trial variation scale. We used public SEEG data from five epilepsy sessions, 183 sites, and 732 test pulses. Both readers used three early template pulses per site, a common covariance shape, and separate probability calibration on fourth pulses. For responses from 20 to 100 ms, mean test log loss was 3.358 bits per pulse with a shared scale and 3.348 with site scales. The mean gain was 0.010 bits per pulse. Three patients had a gain and two had a loss. The descriptive patient interval was -0.013 to +0.039 bits per pulse. Mean accuracy was 31.80% and 31.78%. With temperature fixed at one, site scales increased loss in all five patients. These data do not establish a consistent gain from site scales with this fitting budget. Tests were later within each site, but most occurred before fitting ended at other sites. This is an offline extension on a reused cohort. The package preserves figures, inputs, code, fitted models, all test probabilities, source hashes, and reproduction instructions. A raw-data rebuild and a second analysis run matched the preserved data and numerical results exactly.

stereoelectroencephalography · brain-computer interface · electrical stimulation · probability calibration · site variation · log loss