Row splits gave higher mean SOZ AUC than patient splits for five nearest neighbors in 26 SEEG patients

We tested seizure onset zone contact labels in public SEEG. A model with five nearest neighbors gave mean patient AUCs of 0.698 with row splits and 0.633 with patient splits. The paired mean difference was 0.065. Its descriptive patient resampling interval was 0.036 to 0.095. The study used four fixed power features from two 30-second public SEEG clips in each of 26 patients. The source was OpenNeuro ds004100, version 1.1.3. We compared row, contact, shaft, and patient test groups. Each test used five folds and ten repeats. Logistic regression changed much less: its row-minus-patient difference was 0.005. Across 100 random contact-label assignments, the five-neighbor mean AUC was 0.525 for rows and 0.496 for patients. Intervals use fixed saved predictions and exclude full model fitting uncertainty. This is a model-dependent test-design result, without evidence of clinical use. The package preserves code, features, folds, predictions, source versions, retrieval recipes, and checks of all 52 source recordings.

SEEG · seizure onset zone · patient-level validation · data leakage · AUC · power features · nearest neighbors · logistic regression