Design of BCI motor imagery classification using WPT-CSP and CNN

Abstract

This study proposes a new method in an electroencephalograph (EEG)-based Brain-Computer Interface (BCI) that can directly utilize brain signals to control external devices. The motor Imagery (MI) signal, which contains an image of a certain limb movement, is generally used in BCI. It does not need direct movements. The implementation of MI-EEG signal into BCI still experiences major issues because the patterns obtained for each recording can vary from one another even though they have the same type of motion. In this study, we utilized the Wavelet Packet Transform (WPT) method to decompose the EEG signal into specific sub-band frequencies and Common Spatial Pattern (CSP) as a spatial filter to increase the spatial resolution of the EEG signal. The Convolutional Neural Network (CNN) was subsequently selected for training from the classifier. Next, the results of the training were used to classify the movements of the given MI-EEG. We evaluated the model using dataset 2a from Brain-Computer Interface Competition (BCIC) IV. The results of this method showed the increase in the accuracy of 32% and in Kappa up to 0.42 and the decrease in Root Mean Square Error (RMSE) up to 1.21, compared with only using CNN as the classifier. These results showed fairly good performance compared to other methods used previously in dataset 2a from BCIC IV.