Analysis and classification of electroencephalography signal using machine learning algorithms

Brain-Computer Interface research, based on electroencephalography (EEG) sings, aims to achieve higher classification performance and faster systems than existing studies. In this thesis study, the classification of EEG data for Brain-Computer Interface systems has been performed. Standard deviation normalization has been applied to the data in order to eliminate the noise-related defects within the EEG signs and to standardize the whole signs. In addition, the signs are divided into lower frequency bands so that the information in each frequency band can be obtained separately. Subsequently, different feature groups were applied to the signs, and the feature groups showing the highest classification success were selected. Sequential Forward Generation Algorithm is used to remove insufficient features within the feature matrix. In the study, two different methods are proposed to compare classifier performances and to achieve the highest classification performance. In the first proposed method, the classification was carried out with k-Nearest Neighborhood, Support Vector and Linear Discriminant Analysis algorithms, which are among the machine learning based classifiers. In the second method, the classification was carried out using Deep Neural Networks, one of the deep learning based classifiers. In Deep Neural Networks, two, four, eight and sixteen layer deep network models were created and classification successes were analyzed. As a result of the thesis study, a classification success of 89.4% with k-Nearest Neighborhood, 88.7% with Support Vector Machines, 88.3% with Linear Discrimination Analysis, 88.07% with two-layer basic neural network model, 92.5% with four-layer deep neural network model, 96.82% with eight-layer deep neural network model and 94.67% with sixteen-layer deep neural network model. The results obtained support the view that deep learning based classifiers give higher classification success in EEG based Brain-Computer Interface systems than machine learning based classifiers.