A z-scored interictal SEEG model transferred across awake centers with no meaningful loss, but awake-to-sleep loss remained uncertain

Across archived interictal bipolar-pair feature tables from HUP (36 awake patients), ds003876 (30 awake patients across four centers), and ds007118/ds007119/ds007120 (43 NREM patients), we evaluated transfer of L2 logistic models against leave-one-patient-out (LOPO) training. In the HUP-to-ds003876 test, mean patient-level AUC was 0.723 (95% CI 0.678–0.768) for transfer and 0.715 for LOPO; the transfer-minus-LOPO difference was +0.008 (0.005–0.012), within the prespecified ±0.05 margin. Transfer from ds003876 to HUP gave AUC 0.719 versus 0.714. Across five transfer tests, all AUCs exceeded chance (Holm-adjusted p ≤ 0.00045). HUP and ds003876 tests showed no meaningful loss under the reported criterion; for awake-to-sleep transfer, AUC was 0.659 versus 0.683 for LOPO and the loss confidence interval (−0.035 to +0.076) did not exclude a meaningful loss. Models did not clearly improve on the best single feature, delta log power, whose values were higher in SOZ pairs. Performance near AUC 0.7 is insufficient for clinical localization. The analysis used archived feature tables without new signal processing; it assumes comparable clinical SOZ labels and that 65–115 Hz and 80–300 Hz bands carry the same SOZ information. The HUP and ds003876 source snapshots are not recorded, while the sleep report records snapshot 1.0.0.

stereo-EEG · seizure onset zone · interictal · cross-center transfer · logistic regression · AUC · delta power · sleep