A 60-s SEEG spectrum separates atlas-labeled white- and gray-matter pairs across held-out patients (mean AUC 0.747)

Question: can one 60-s bipolar SEEG spectrum tell a white-matter pair (WW, both contacts in WM) from a gray-matter pair (GG) in a patient that the classifier did not see? Data: archived 1-45 Hz spectra of RAM FR1/catFR1 (AutoSci c3188d78; OpenNeuro ds004789, ds004809) and wb.region atlas labels (rule of e17a2508, as in 4fd526f0 and b4e46e3c). 126 patients, 3994 WW and 3079 GG depth pairs. Method: the plan was hashed before analysis. 9 features per pair (broadband level, aperiodic exponent, 5 relative band powers, 4-13 Hz peak), normalized inside each patient without labels. L2 logistic regression, leave-one-patient-out, AUC per held-out patient, sign-flip tests with the patient as the unit, Holm correction. Results: mean AUC 0.747 (95 % CI 0.722-0.771; 115/126 patients above 0.5; 40 % of patients at 0.8 or more; balanced accuracy 0.69). Broadband level alone gave 0.722. The full model was better by 0.025 (CI 0.006-0.044, Holm p 0.010). Shape features without level gave 0.618. WW pairs had less delta and gamma and more beta and alpha. Checks: same shaft 0.770; train FR1, test catFR1 0.763; raw features 0.744; without hippocampus/amygdala GG pairs 0.712; mixed pairs scored between WW and GG. Conclusion: the spectrum can flag likely WM pairs when imaging labels are not available, but about 1 in 3 pairs is wrong at the 0.5 threshold, so it does not replace imaging. Limits: atlas labels in MNI space, 60 s task data, epilepsy patients, bipolar montage, 1-45 Hz, one classifier with a fixed penalty, only pure WW and GG pairs scored.

stereo-electroencephalography · white matter · gray matter · electrode localization · spectral power · patient-level validation