Across 5 SEEG centers, a 10-feature model trained at other centers ranked clinical SOZ pairs better than HFA

Many studies show that 1 interictal feature ranks seizure-onset-zone (SOZ) SEEG pairs above chance. A clinical tool must also work at a new center. We tested whether a 10-feature model, fitted only on patients from other centers, ranks SOZ pairs better than high-frequency activity (HFA, 80–150 Hz) alone. We used archived per-pair tables, with no new signal processing, for 60 patients from 5 centers (HUP 31, NIH 11, PY 8, rns 9, umf 1; 5,820 pairs, 559 SOZ). Features were z-scored within each patient. We used L2 logistic regression with leave-one-center-out (LOCO) fitting, per-patient rank AUC, a sign-flip test, a patient bootstrap, and Holm correction. The plan was hashed before analysis. Mean LOCO AUC was 0.755 and HFA AUC was 0.665 (difference +0.090, 95% CI 0.041–0.141, Holm p=0.0008; higher in 43 of 60 patients). LOCO was better than HFA at all 5 centers. Leave-one-patient-out fitting, which also used the same center, gave no measurable gain (+0.004, p=0.20). Pooled z-scores instead of within-patient z-scores reduced AUC by 0.035 (p=0.0017). Spike rate and 1–40 Hz power had the largest stable weights. A post hoc test found the model also outperformed the best single feature, 1–40 Hz power (+0.058, p=0.0015). Limits: clinical labels without surgical outcome; 120-s clips; 4 of 5 centers came from one dataset, all features came from one pipeline, and umf had 1 patient.

SEEG · seizure-onset zone · interictal · epilepsy · leave-one-center-out · logistic regression · rank AUC · high-frequency activity · spike rate