Abstracts

Rationale: Brain connectivity analysis has been used as an emerging novel tool to study epilepsy (Liao, Zhang, et al. 2010). In recent years, it has been shown that brain network topology based on the graph theory can be used to identify seizure onset zone (SOZ) in patients with epilepsy (Li, Chennuri, et al. 2018, Ren, Cong, et al. 2019). In this study, we investigated the identification of the SOZ using a machine learning approach based on graph measures as input features. We used a convolutional neural network (CNN) to identify efficient input features from the graph measures and then used these input features to train and test a classifier for identifying SOZ in electrocorticographic (ECoG) recordings. Methods: We retrospectively analyzed 19 seizures in six patients (4 males; aged 19-40 years) who underwent a Phase II epilepsy surgery evaluation with subdural electrodes in our center. All patients were seizure-free after a minimum 6-month follow up. Resections were tailored individually based on visual inspection of the ECoG ictal onset in all patients. After preprocessing of ECoG data, we calculated dynamical connectivity between subdural electrodes using Granger causality based on a multivariate autoregressive model. Granger causality between each pair of electrodes was performed from 10 s before to 10 s after ictal onset using a moving window of 10 s with 9 s overlap in a 1 to 150 Hz frequency range. Then outflow and inflow graph measures were extracted from the adjacency matrix constructed from the Granger causality. Next, a 3-dimensional feature matrix was constructed by considering outflow and inflow and their dot product. After that, we used the Alex-net CNN as a backbone network for feature extraction. The fully-connected layer 7 output of the Alex-net was used as the extracted feature vector (Krizhevsky, Sutskever et al. 2012). Finally, the extracted feature vector was used in a support vector machine (SVM) with the linear kernel to classify the electrodes into two classes: 1) visually detected electrodes (VDE) as SOZ by epileptologists; and 2) non-resected electrodes, presumably electrodes outside of the SOZ (non-SOZ). We trained, cross-validated, and tested the SVM for binary classification of the subdural electrodes, and reported accuracy, sensitivity, and specificity. Results: The accuracy, sensitivity, and specificity of the SVM classifier for identifying SOZ and non-SOZ electrodes were 81.0%, 87.5%, and 79.1%, respectively. Conclusions: The results of this study show that the Alex-net CNN can efficiently extract features from graph measures, and an SVM classifier based on these features can accurately classify SOZ and non-SOZ electrodes. The developed method in the current study may be used for identification of SOZ electrodes that may result in an increased chance of successful epileptic surgeries. Funding: This study was funded by the Children’s Foundation Research Institute & The Shainberg Neuroscience Fund, Memphis, TN.