Increase performance of four-class classification for motor-imagery based brain-computer interface

Brain computer interface (BCI) is a system that provide a direct communication between human brain and external devices. BCIs which based on mental tasks of users are widely used for disabled or paralyzed patients, in order to help their mobility. Preprocessing techniques have been extensively developed to increase the signal-to-noise ratio and spatial distribution of the signals. Common Spatial Pattern (CSP) has shown to be a robust and effective method for processing Electroencephalogram (EEG) data. However, the results of CSP filter are still far from being completely explored. CSP was originally designed for two-class problem despite the fact that a practical application of Motor-imagery (MI) based BCI contains numbers of activities. It is necessary to design the classification algorithm which applicable to more than two-class problem. In this paper we investigate the performance of CSP by selecting optimal time slice and components for training CSP filters in four-class BCI by separating the four-class problem into multiple binary classifications. Our method is verified in the testing phase with four different types of classification approaches which are linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), linear support vector machines (LSVM), and support vector machines with radial basis function kernel (RBF-SVM). The result showed that, under the optimal time slice and components, the classification accuracy reach 78.82% for the best untrained subject in this dataset.