Nearest training labels gave mean AUC 0.934 within test folds for 34 SEEG patients

We tested clinical seizure-onset label prediction without test signals in OpenNeuro ds004100 version 1.1.3. The fixed sample had 34 patients and 4,031 numbered SEEG contacts. Each test score used the nearest training labels in the same patient and name prefix. Test-fold labels were excluded. Equal-patient mean AUC within test folds was 0.9340. A uniform shuffle that kept each prefix's positive count gave mean AUC 0.7197. None of 9,999 joint draws reached the observed mean; the add-one p value was 0.0001 under that reference model. Mean AUC fell to 0.7906 when training contacts with number difference at most one were excluded. The result describes label completion with known training labels. It does not establish localization in an unlabeled implant. The package preserves the source chain, code, spectra, labels, masks, scores, reference draws, checks, and reproduction instructions.

stereo-EEG · seizure onset · clinical labels · label completion · contact-level evaluation · AUC · data leakage