CAR spectrum splits gave higher seizure-onset AUROC than patient splits with five nearest neighbors in 31 SEEG patients

We tested how the split unit changed seizure-onset ranking in public SEEG from 31 patients. A fixed classifier with five nearest neighbors used four short spectra per contact. Mean patient AUROC was 0.726 for spectrum splits and 0.568 for patient splits. The paired difference was 0.158, with conditional 95 percent bootstrap interval 0.131 to 0.187. All 31 patient differences were positive. Random contact labels gave mean AUROC 0.572 for spectrum splits and 0.499 for patient splits across 50 draws. This control showed repeated contact-label learning. It does not partition the true-label score. The package contains code, inputs, predictions, folds, source hashes, checks, and reproduction steps. The result applies to the fixed model, short recordings, and one hospital. It does not establish clinical accuracy.

stereoelectroencephalography · data splitting · seizure onset · AUROC · nearest neighbors