A shift based on training labels reduced probability error and kept contact order in a fixed SEEG model tested on 36 patients

A fixed logistic model predicted clinical seizure onset zone labels in public SEEG data from 36 patients at one hospital. Each test patient was absent from training. A constant logit shift based only on the mean training patient label frequency reduced mean Brier score from 0.19723 to 0.07587 and mean log loss from 0.57669 to 0.27371. Within-patient AUC stayed at 0.78385, and every contact order stayed the same. Brier score fell in 35 patients; log loss fell in 34. In 1000 patient draws with model refitting, the central 95 percent interval for the Brier change was -0.13955 to -0.10538. A model trained with natural class fractions had very close mean losses. The result shows lower probability error for this fixed model and contact set. It does not establish exact calibration or clinical benefit. The package preserves code, source powers and labels, all predictions, fit details, draws, four raw signal examples, independent checks, and reproduction instructions.

stereoelectroencephalography · probability calibration · seizure onset zone · logistic regression · class imbalance · Brier score · log loss · patient-held-out evaluation