SEEG spectra classify gray- and white-matter contacts in held-out patients, while shaft position does not

We tested whether short segments of inter-pulse SEEG activity can classify gray-matter (GM) versus white-matter (WM) contacts without imaging labels. In eight patients from one center, we used 30 two-second windows per patient, each 2–4 s after a clinical single pulse, and evaluated log band power and spectral slopes with leave-one-patient-out logistic regression. Mean AUCs were 0.867 for the recorded reference, 0.882 for common average reference (CAR), 0.892 for shaft Laplacian (LAP), and 0.902 using all three references. All spectral models exceeded chance in all eight held-out patients (exact sign-flip p = 0.0078); differences between references were not significant. Shaft position alone had mean AUC 0.570 (p = 0.18) and added nothing to the spectral model. With LAP, the top 20% of contacts by score were 97.6% GM on average, compared with a 64.2% GM base rate. The study used dataset tissue labels that were not checked against imaging; windows were not true rest and could retain late pulse effects; mixed GM/WM contacts were excluded. Results are limited to eight epilepsy patients from one center.

SEEG · intracranial EEG · gray matter · white matter · spectral power · contact classification · leave-one-patient-out · brain-computer interfaces