Abstracts

Rationale: Pediatric epilepsy surgery is an important therapeutic option in the treatment of intractable seizures, yet preoperative assessments cannot guarantee seizure-free outcomes. Recently, techniques for extracting macroscopic connectivity patterns from preoperative intracranial EEG recordings have been developed to improve evaluations of surgical candidacy and delineate the epileptogenic zone with greater sensitivity and accuracy than human reviewers. Using this approach, distributed cortical networks extending beyond the neurologist-defined seizure onset zone (SOZ) have been identified and hypothesized as potential surgical targets. In this study, we used a graph theoretic approach to determine whether local and global parameters of interictal functional connectivity networks predict clinical variables such as MRI presentation and postsurgical outcome.Methods: This study is IRB approved. Eighteen pediatric patients (mean age = 10.3, range = 3-17 years) with neocortical epilepsies were studied. For each patient, 20 minutes of randomly-selected, preoperative intracranial EEG (1-second epochs, n = 1200) were analyzed. Functional connectivity measures were obtained by calculating the pairwise Spearman correlation coefficient, and the resultant networks were characterized across a wide range of threshold densities using graph theory. Local synchrony was quantified by computing the channel-wise root mean square (RMS) amplitude within each epoch. Nonparametric permutation tests and support vector machine (SVM) learning assessed whether univariate amplitude measures and multivariate network indices varied across clinical groups.Results: Three main findings emerged from this study: (i) increased network resilience and local information transfer efficiency (global clustering coefficient, transitivity) characterized patients with postsurgical seizure recurrence; (ii) increased node partitioning and community structure (modularity coefficient, node diversity, degree variance) were observed in patients with MRI-visible cortical lesions relative to nonlesional patients; and (iii) SOZ electrodes exhibited increased network influence (degree centrality, eigenvector centrality) and univariate signal complexity compared to non-SOZ electrodes. These measures were collapsed using principle components analysis (PCA) and submitted to a linear-kernel SVM as predictors. The learning algorithm accurately classified patients by post-surgical outcome (70.4%) and presence/absence of MRI-visible lesions (74.5%).Conclusions: We conclude that univariate and multivariate signal analyses provide efficient characterization of the epileptogenic zone. Further, these measures may represent clinically-valid biomarkers allowing neurologists to make more quantitative surgical candidacy assessments. This graph theoretic approach to network delineation is an important advance in presurgical planning and in understanding the structure and function of the neuronal networks that are abnormal in a patient with intractable epilepsy.