A sleep SEEG seizure-onset-zone score transfers across centers, but sigma power alone performs as well

We tested whether a seizure-onset-zone (SOZ) score fitted on one cohort transfers to another without refitting, and whether its four-feature model outperforms sigma (12–16 Hz) power alone. We used archived pair features from 43 NREM SEEG patients (728 bipolar pairs, 160 SOZ pairs) and 18 UCLA pediatric SEEG patients (1,285 pairs, 181 SOZ pairs). Features were log HFA power, log sigma power, and slow-oscillation coupling z-scores for sigma and HFA. Features were z-scored within each patient. An L2 class-balanced logistic model was fitted on one cohort and frozen for testing in the other; outcomes were per-patient AUCs. NREM-to-UCLA transfer gave mean AUC 0.710 (95% CI 0.643–0.766; 17/18 above 0.5; Holm p = 0.00015). UCLA-to-NREM transfer gave 0.716 (95% CI 0.621–0.803; 35/43 above 0.5; Holm p = 0.0002). Within-cohort leave-one-patient-out AUCs were 0.710 and 0.731. In UCLA, the model did not outperform sigma power alone (difference −0.016, p = 0.38). The model weights and single-feature AUCs also show that sigma power had the largest weight in both directions. These results indicate cross-center transfer in these cohorts, while coupling features added no portable SOZ information beyond sigma power. Windows were short (120–180 s), sleep stage was unknown for UCLA, cohorts differed in age, clinical SOZ labels are not ground truth, and within-patient z-scoring supports pair ranking only, not an absolute threshold.

stereo-EEG · seizure-onset zone · sleep · sigma power · high-frequency activity · cross-center transfer · AUC · epilepsy