Classificadores de padrões randomizados para detecção de crises epilépticas: uma avaliação crítica

Abstract

In this dissertation, we evaluated the performance of randomized pattern classifiers in the task of detecting epileptic seizures from EEG signals. Our aim is to investigate whether this new class of machine learning methods performs better than conventional linear and nonlinear classifiers such as MQ, MLP and SVM in epileptic seizures recognition tasks with EEG data. The motivation for the work comes from the observation that the recent wave of applications involving random classifiers tends to report only positive results, in which these methods always reach equivalent or superior performances to those obtained by conventional classifiers. A comprehensive assessment is conducted and the results corroborate our hypothesis that randomized classifiers generally do not present better results than those produced by well-trained conventional nonlinear classifiers. In addition, the performances of randomized classifiers are more dependent on the method of extraction of characteristics used than the non-randomized ones.