Abstracts

Rationale: Stereo-electroencephalography (SEEG) is a promising pre-surgical diagnostic tool enabling precise recordings from deep cortical structures while avoiding large craniotomies. Despite its high temporal resolution, SEEG can be limited by poor spatial resolution and is blind to electrical activity beyond the targeted cerebral structures. This is particularly problematic when the falsely localized epileptogenic focus lies within hippocampus or other functionally important region. Resection of the otherwise normally functioning hippocampus can lead to unnecessary comorbidities such as memory deficit and can significantly impact patient care. The goal of this study is to characterize hippocampal onset seizures (HOS) and hippocampal spread seizures (HSS) using signal analysis of SEEG recordings and build a classifier to distinguish between these two types of seizures.Methods: Seizures from 25 patients (12 hippocampal onset and 13 extra-hippocampal onset) who underwent pre-surgical evaluation and had successful surgical outcome were used for the study. The hippocampal group includes patients who underwent temporal lobectomy whereas in the extra-hippocampal group, the hippocampus was spared from resection. A total of 101 (60 HOS and 41 HSS) seizures were selected and the onset and termination of each seizure was marked (Table 1). In case of HSS's, seizure onset within the hippocampus was identified and marked. A single contact within the hippocampus which had the earliest seizure associated changes was used for subsequent analysis. Seizures were divided into 6 segments using an in-house developed seizure segmentation algorithm. Normalized power in the delta, theta, alpha, beta, gamma, high gamma, and high frequency (100>f>150) were estimated per segment. Thus a total of 42 features characterized a seizure. Features were standardized and ranked using a MSVM-RFE algorithm. A 5-fold cross-validation procedure was used to optimize a linear Support Vector Machine (SVM) classifier. To evaluate the performance of the classifier on unknown data set, a Leave-One-Out Cross validation strategy was followed. The performance of the classifier was evaluated by estimating the Sensitivity, Specificity and Area under the Curve (AUC).Results: The top seven features selected are shown in Fig 1A. In particular it was noticed that an increased expression of Beta and Gamma frequency in the onset is indicative of a hippocampal onset seizure. Overall, the classifier achieved an AUC score of 0.895 which indicate a highly discriminative classifier. For a detection threshold of 0.50 the classifier had a sensitivity of 0.80 and a specificity of 0.88 (Fig 1B). Only 3 extra-hippocampal patients were misclassified as hippocampal patients, as shown by the average probability score in Fig 1C.Conclusions: The pattern recognition algorithm presented in this study provides an objective measure to distinguish between hippocampal onset seizures and hippocampal spread seizure. Validation of the methodology on a larger data set can provide clinicians with a useful tool for improving surgery outcome and patient care.