Fixed combined models had mean SOZ AUROC 0.756 under matched and changed test references in 31 SEEG patients, with 79 percent top contact list overlap

We tested whether a change of voltage reference changes contact ranks from a fixed SEEG model. We used public OpenNeuro ds004100 version 1.1.3 and archived models for 31 patients, 2,879 contacts, and 271 seizure onset zone (SOZ) labels. The combined models used contact numbers and five band power inputs. Each model excluded its test patient from training and penalty selection. We scored the same test recording with a common average reference and a local neighbor reference. The combined models had mean AUROC 0.756480 with the training reference and 0.756316 with the other reference. The mean was first calculated across the two training directions within each patient. Opposite changes often canceled. The two test conditions shared 78.8 percent of the highest-scored 10 percent of contacts. Individual matched-minus-changed AUROC values ranged from -0.132 to 0.123. A similar mean AUROC did not give the same contact selection. All source hashes, fixed model scores, and independent metric checks passed. The package has code, frozen models, spectra, metadata, contact scores, and pinned raw retrieval instructions. The result applies to this selected cohort and model. It does not establish reference equivalence or treatment benefit.

stereo-EEG · SEEG · epilepsy · seizure onset zone · voltage reference · AUROC · contact ranking · reference stability