A HUP-trained interictal SEEG classifier ranked SOZ channels across four centers but did not outperform spike rate

We asked whether a classifier trained on interictal stereo-electroencephalography (SEEG) features at one center could rank seizure-onset-zone (SOZ) channels at other centers, and whether it outperformed the best single training-center feature. We used existing per-channel feature tables derived from OpenNeuro ds004100 (HUP; 31 patients) and ds003876 (NIH, JHU, MIAMI, UMF; 29 analyzed patients). A nine-feature L2 logistic regression trained on HUP ranked SOZ channels in the external patients with mean per-patient AUC 0.705 (95% CI 0.638–0.769; 25/29 above 0.5; Holm-adjusted p = 2×10⁻⁵). It was not significantly better than HUP-selected spike rate (AUC 0.672; difference +0.033, 95% CI −0.005 to +0.074; Holm-adjusted p = 0.12). Reverse transfer gave AUC 0.726 and no significant gain over broadband power. Across 60 pooled patients, leave-one-site-out AUC was 0.720 versus 0.746 for leave-one-patient-out (difference 0.026, 95% CI 0.013–0.039; p = 0.0003). Raw features outperformed the planned within-patient ranks in the primary direction (AUC 0.742). Post hoc top-five precision was 0.32 versus 0.11 chance (p = 0.0001), with no significant difference from the single feature. The results support cross-center ranking but do not show that the multi-feature model improves on a single feature. AUC near 0.7 and top-five precision 0.32 are insufficient for clinical SOZ localization alone. Limitations include one 120-s clip per patient, uncontrolled sleep/wake state, center-dependent clinical labels, no anatomical control, small patient counts, one UMF patient, and fixed linear models. The analysis reused archived feature tables; it did not recompute them from raw recordings.

stereo-electroencephalography · SEEG · epilepsy · seizure-onset zone · SOZ · cross-center transfer · channel classifier · interictal features · AUC · clinical labels