Modelo Híbrido Inception-XGB para Discriminação de Padrões Artefatuais em Eletroencefalografia de Pacientes com Epilepsia

Abstract

The aim of this study is to compare the performance of different machine learning algorithms for the task of discriminating four artifactual patterns of common occurrence in electroencephalography (EEG) recordings. These patterns were processed by an Inception Convolutional Neural Network (CNN) from EEG recorded from epilepsy patients. The EEG was part of a public available dataset. A total of 200 patients participated in the study, 100 men and 100 women, with varying ages and associated comorbidities. For feature engineering we used the Hilbert-Huang Transform (HHT) to obtain instantaneous power and phase from EEG samples comprising one second of recording each. They contained artifacts distributed in four classes: a) eye movement, b) muscle movement, c) electrostatic noise, and d) shivering. The two-dimensional samples composed of power and phase were processed by the Inception architecture where they were reduced to a one-dimensional vector and then fed to four different machine learning algorithms, namely: a) Extreme Gradient Boosting (XGB), b) Random Forest Classifier (RFC), c) K-Neighbors Classifier (k-NN), and d) Logistic Regression (LR). The average accuracy of each model, per training instance, was compared using ANOVA in order to verify the existence of statistical difference among models. The results demonstrated that the XGB algorithm had the best performance, reaching an accuracy score of 94.82%, outperforming the baseline model, which reached an accuracy score of 83.97% when trained with raw EEG data. This established an unprecedented Inception-XGB hybrid model together with an efficient feature extraction process.