Abstracts

Rationale: The success of epilepsy surgery depends on precise localization and resection of the epileptogenic zone [1]. Aim of this study was to develop and evaluate a machine learning approach to accurately localize the seizure onset zone (SOZ) in the electrocorticographic (ECoG) recordings. We hypothesized that the synchronization between low (4-30 Hz) and high (80-150 Hz) frequencies in ECoG recordings can be used to automatically identify the SOZ. Methods: We retrospectively analyzed 23 seizure episodes in 10 patients (7 males; aged 23.0 9.0 (mean SD) years), who underwent a Phase II epilepsy surgery evaluation with intracranial electrodes. Resections were tailored individually based on visual inspection of the ECoG ictal onset in all patients. Patients 1 to 6 were seizure-free after surgery, but not Patients 7 to 10 (although these patients were significantly improved). After preprocessing of ECoG data, phase locking value (PLV) between the phase of low frequency (4-30 Hz) and phase of the Hilbert transform of high frequency (80-150 Hz) was calculated. Five following features were extracted from PLV signal within a 10-30 sec time window before seizure onset in each electrode. 1- PLV positive: This feature was assigned to "1" if the PLV would exceed a threshold of "_b+?-6 ?-s?-_b", where _b and s_b are the mean and standard deviation of PLV in a 60-second baseline, respectively. 2- Duration of PLV positive: Duration of PLV signal exceeding a threshold at _b+?-6 ?-s?-_b 3- Peak: The maximum value of PLV 4- Average: Mean of PLV 5- Energy: The energy of PLV signal from seizure onset till the end of seizure Finally, the logistic regression (LR) and support vector machine (SVM, with radial basis function kernel) classifiers were trained using the above features in seizure-free patients (Patients 1-6), and tested to predict SOZ electrodes in non-seizure-free patients (Patients 7-10). Results: All electrodes identified as SOZ by two classifiers were within the resected area (RA) in all seizure-free patients (Table 1). In non-seizure-free patients, two classifiers identified some SOZ electrodes within and also outside of the resected area. SVM classifier outperformed the logistic regression classifier in that it identified more SOZ electrodes outside of the resected area in two out of four non-seizure-free patients (Patients 7 and 8). Visually identified SOZ electrodes in all patients were identified by two classifiers. Conclusions: We compared performance of two classifiers (i.e. logistic regression and SVM) to identify the SOZ using features extracted based on the PLV between the phases of the low and high frequencies in ECoG recordings. Our results show that logistic regression and SVM classifiers can identify the SOZ. Identification of the SOZ based on a machine learning approach may enhance the standard clinical procedure based on the visual inspection, and thus this approach should be included in decisions on surgical treatment. References 1. Brain 2001;124; 1683-1700. 2. Neurology 2015; 84;2320-2328. 3. Neuroscience 2008;174;50-61 Funding: This study was funded by Herrf College of Engineering of University of Memphis and the Children's Foundation Research Institute & The Shainberg Neuroscience Fund, Memphis, TN.