Contact holdout gave higher SOZ AUC than patient holdout for a fixed random forest in 36 short SEEG records

We tested seizure onset zone (SOZ) label prediction in public stereoelectroencephalography (SEEG) data. We used 30 seconds from each of 36 HUP patients and eight power features. We fixed the random forest settings before we computed scores. Mean patient area under the ROC curve (AUC) was 0.792 with contact holdout and 0.725 with patient holdout. The paired difference was 0.067; its patient bootstrap 95 percent interval was 0.041 to 0.096. Shaft holdout gave 0.742. A control that used only training labels on the same shaft gave 0.870. The gap remained in a second recording and with the stored reference. The package preserves spectra, source records, folds, scores, code, and checks. These results apply to this fixed model and short records from one hospital. They do not establish clinical accuracy.

seizure onset zone · stereoelectroencephalography · SEEG · holdout validation · patient-level validation · contact-level validation · random forest · area under the ROC curve