A nonnegative contact variance fit gave higher mean correlation error than zero correlation in paired 30–80 Hz SEEG records from 36 patients

We tested shared-contact bipolar correlation prediction between two SEEG records per patient. We used public OpenNeuro ds004100 version 1.1.3 and all 36 paired patients in a checked AutoSci archive. The primary band was 30–80 Hz. A convex nonnegative fit used only the training covariance. Each direction used 3,160 shared-contact links. Scores gave each patient equal weight. Mean squared correlation error was 0.181015 for the fit, 0.281098 for equal contact variances, 0.113506 for zero correlation, and 0.008842 for measured training correlation. The fit improved on equal variances in all patients. Its mean gain was 0.100082, with a 95 percent patient bootstrap interval of 0.091342 to 0.110077. Zero correlation gave less error in 31 patients. Training correlation gave less error in all patients. All six fixed checks kept the same order of mean errors. Synthetic checks and an exact offline repeat passed. Fresh EDF retrieval reproduced all 72 covariance inputs exactly. The package preserves inputs, code, source pins, fitted parameters, every prediction, and bootstrap draws. This is a new measurement analysis of a reused clinical cohort. It does not establish neural independence or clinical benefit.

SEEG · stereoelectroencephalography · bipolar signals · contact variance · correlation prediction · measurement model · covariance · OpenNeuro