Extração e seleção de características de sinais de eletroencefalograma para diagnóstico de crises epilépticas

Abstract

Epilepsy is one of the most common diseases and affects approximately 50 million people worldwide. Its diagnosis requires analyzing an electroencephalogram (EEG), which measures brain electrical activity represented by time series. The analysis of EEG segments depends on human interpretation, which may lead to divergent results, and be tedious, imprecise, and error-prone. Moreover, one estimates more than 80% of the epilepsy exams return no anomalies at all, wasting the effort of analysis. This paper proposes an automatic way to analyze EEG segments. The approach is based on the combination of multispectral analysis, feature selection and machine learning (ML) methods for dimensionality reduction in EEG classification. As our main contribution, we proposed a methodology to minimize the number of features required for classifier training, minimizing the interference of irrelevant measures. The selected features were used as input of five ML algorithms to classify EEG segments. A minimum set of features is chosen for each selection method, and the results are compared select the best feature subset. In this work, features based on statistics, energy, entropy and specific measures were extracted from the four representations of the EEG. These features were selected by feature selection techniques and used as input in ML algorithms. In the experimental evaluation, results statistically prove that this approach is as efficient as using the complete feature dataset for classifier building. The proposed classifiers improve the related literature, reaching an accuracy between 87.2 and 90.99%, and considerably reduce the number of features, from 285 to 30, under the same accuracy scores.