Division of evidence by four reduced mean log loss with a fixed Gaussian reader in five offline SEEG sessions

We tested probability error for repeated electrical pulse responses in public SEEG data. We used OpenNeuro ds004457 version 1.0.2, five patient sessions, 183 sites, and 15 recording pairs. A fixed reader used three early pulses per site for its template and one more pulse for its probability scale. The final four pulses formed one test episode per site. We compared the sum of response evidence with the same sum divided by four. Both rules gave the same site decisions and 48.3% mean accuracy. Mean log loss was 2.891 bits for sum evidence and 2.620 bits for mean evidence. The reduction was 0.271 bits and was positive in all five patients. The descriptive patient bootstrap interval was 0.116 to 0.427 bits. The exact two-sided patient test gave p = 0.0625. Mean evidence did not improve mean Brier error and had confidence below accuracy. The result supports a test of probability scale when a reader combines repeated responses. It does not establish a population gain, an optimal scale, or online brain control. The package contains two figures, source code, voltage inputs, model arrays, pulse records, and source retrieval instructions.

SEEG · electrical stimulation · brain interfaces · probability calibration · log loss · repeated responses · Gaussian classifier