Internal and external validation of comprehensive high-frequency activity biomarkers for epilepsy surgery

Objective To validate models incorporating intracranial EEG-derived high-frequency activity (HFA) biomarkers for predicting the seizure onset zone (SOZ) and postsurgical seizure outcomes across independent cohorts and etiologies. Methods We quantified interictal HFA occurrence rates together with a comprehensive set of morphological features in people with drug-resistant epilepsy in a derivation cohort (N = 142), a temporal external cohort (N = 35), and two geographical external cohorts (N = 26 and N = 30). We developed models for SOZ localization using data from the 79 derivation cohort individuals who achieved postoperative seizure freedom. Model performance for SOZ localization and postoperative seizure outcome prediction was subsequently evaluated in the derivation, temporal external, and geographical external cohorts. Results HFA rate, spectral entropy, and power emerged as the most influential features for accurate SOZ classification. The model identified SOZ sites with areas under the curve (AUCs) up to 0.85, 0.86, and 0.75 in the derivation, temporal external, and geographical external cohorts, respectively. The model predicted postoperative seizure freedom in the derivation cohort (AUC up to 0.70) but failed to predict outcomes reliably in external cohorts. Among individuals in the external cohorts with MRI-nonlesional epilepsy, postoperative seizure freedom was predicted with an AUC of up to 0.73, whereas performance declined to 0.46 or lower among individuals with encephalomalacia. Conclusions Integration of HFA rates with morphological features yields an SOZ-localization biomarker with cross-center generalizability, whereas postoperative outcome prediction remains dependent on underlying etiology. Significance A surgical strategy prioritizing resection of HFA-involved areas may be ineffective in individuals with encephalomalacia, underscoring the need for etiology-specific interpretation of HFA biomarkers.