A frozen SEEG gray/white-matter classifier transferred across two centers, but its threshold did not

A frozen L2-logistic classifier trained on background spectra from SEEG recordings at Utrecht (OpenNeuro ds005448) ranked gray- versus white-matter bipolar pairs at Bucharest (ds006519) without refitting. On six subjects with at least three pairs of each class, mean per-subject AUC was 0.845 (6/6 above 0.5; sign-flip p=0.031); with at least one pair of each class, it was 0.799 (10/11; p=0.002). Pooled AUC was 0.796 (95% subject-bootstrap CI 0.736–0.871). A within-Bucharest leave-one-subject-out model had mean AUC 0.802 in the six-subject analysis. In reverse transfer, a Bucharest-fit model had mean AUC 0.898 on eight Utrecht subjects. However, the frozen probability threshold of 0.5 had balanced accuracy 0.72, so the threshold did not transfer. The results are limited by a different montage between centers, few white-matter pairs (77 total; six subjects with at least three), dataset-derived labels, and five-second background segments recorded during clinical stimulation sessions. The complete original code, derived features, scores, summaries, sensitivity results, figure, archived training inputs, and source-file provenance are included.

SEEG · intracranial EEG · gray matter · white matter · spectral features · logistic regression · cross-center transfer · AUC · epilepsy