Parametrização de sinais de eletroencefalograma para classificação de padrões via matrizes de kernel

Abstract

Electroencephalograms (EEGs) belong to a category of biopotential signals widely used in healthcare for the diagnosis of neurological disorders. However, these signals have complex characteristics that commonly lead to misdiagnosis. Hence, EEG signals are the object of study in researches that seek a better understanding of those characteristics and their correlations with certain pathologies and mental states. That said, the present master thesis addresses the process of extracting attributes from EEG signals for classification purposes. This task, also known as signal parameterization, is widely used in machine learning and has several classical methods to extract useful information from the signal. In this context, the present work introduces new EEG signal parameterization strategies, from the use of kernel matrices to build labeled attribute vectors. For that, sparsification methods are used in order to reduce the dimension of the kernel matrices. The construction of the attribute vector follows two different strategies: (i) interchannel mode, in which the EEG signals of the various channels are processed together; and (ii) intrachannel mode, in which channels are processed individually. The methods proposed in this master thesis are compared with two classical linear methods, namely: vector of covariances and vector of linear predictive coding coefficients (LPCC). Only linear discriminants are used for the classification task, since the objective is to assess the quality of the extraction methods and not to evaluate classifiers per se. A comprehensive comparative study evaluating the performance of the proposed methods from the variation of their hyperparameters is carried out. For this purpose, two benchmarking data sets are used to generate the results of the present work. From the analysis and discussion of the results obtained, it was possible to infer that the proposed methods are promising, with performance equivalent or superior to those generated by the classical methods of attribute extraction of EEG signals.