Multiobjective support vector machines optimization for epileptic seizure prediction

Epileptic seizure prediction, with at least some minutes in advance, would improve substantially the quality of life of patients with refractory epilepsy. This is addressed as a classification problem were the brain state is classified using a number of features extracted from the EEG. Methods based on computational intelligence, like support vector machines (SVM), are applied to build up classifiers that will make the prediction by identifying the pre-seizure phase. A problem that arises is the selection of the highly relevant inputs and classifier parameters that lead to the most appropriate classification. When all the channels and a high number of features are considered, the search space becomes high dimensional, leading to computational difficulties. In this work, SVMs were optimized by a multiobjective genetic algorithm, to improve simultaneously sensitivity, specificity and complexity (computed as the number of inputs). The Elitist Non-dominated Sorting Genetic Algorithm (NSGA-II) was parameterized to search for the appropriate inputs and also for the SVM parameters (Cost and Gamma). The obtained results in a group of five patients showed that the applied methodology could reach appropriate sensitivity and specificity values while maintaining complexity low.