Seizure onset zone (SoZ) identification from interictal intracranial electroencephalogram

Accurate localization of the seizure onset zone (SoZ) is critical for effective surgical treatment in drug-resistant epilepsy. Traditional SoZ biomarkers, such as interictal epileptiform discharges (IEDs), high-frequency oscillation (HFO) rate, and relative band power (RBP), often fail to capture the complex network dynamics underlying seizure generation fully. We analyzed graph-based centrality measures using the novel frequency-domain convergent cross mapping (FD-CCM) technique and found that SoZ regions exhibited significantly lower centrality values compared to the non-SoZ areas, with strong statistical significance ([Formula: see text] [Formula: see text]). This reduced centrality remains a marginally explored biomarker for SoZ localization. A multi-layer perceptron (MLP) classifier was developed to integrate graph-based centrality features with existing biomarkers in the literature. Features are ranked using the minimum Redundancy Maximum Relevance (mRMR) algorithm, and their incremental contributions to performance are evaluated using a 10-fold group cross-validation scheme with subject independence. Discriminability improved progressively from univariate SoZ features to bivariate and multivariate biomarkers, with graph-based features demonstrating the highest T-values across folds. The combined approach with 19 features achieved a mean performance of 0.86 AUC, 82% sensitivity, 90% specificity, and 89% accuracy, outperforming individual feature sets. Particularly, the combined feature set demonstrates superior sensitivity compared to all prior works using interictal recordings (outperforming prior best by 5%), emphasizing its improved capability to identify SOZs accurately. It also exhibits robust generalizability, as evidenced by its internal validation and cross-dataset transfer learning performance. These findings demonstrate the importance of graph-based centrality measures in capturing interictal dynamics of SOZ and highlight the value of integrating diverse biomarkers for improved discriminability and generalizability, paving the way for more accurate pre-surgical evaluations in epilepsy.