Designing Spectral and Spatial Filters Simultaneously for Brain-Computer Interface Based on Mutual Information and Particle Swarm Optimization

Extracting discriminative features is a crucial step in brain-computer interfaces (BCIs) that could affect directly on the classification performance. Common spatial patterns (CSP) is a commonly used algorithm for such propose in motor imagery based BCI systems. CPS tries to extract the most appropriate spatial patterns in the electroencephalogram (EEG) signals to discriminate different motor imagery classes. Before applying CSP, Usually EEG signals are filtered out in 8-30 Hz to capture event related desynchronization (ERD) specific frequency rhythms called mu and beta bands. However, this frequency band could be highly subject specific. Therefore, optimizing spectral and spatial filters jointly could improve the classification accuracy. In this paper, we proposed a novel learning algorithm to derive spatial and spectral filters simultaneously using an evolutionary learning algorithm called particle swarm optimization (PSO). Furthermore, we utilized mutual information between extracted features and class labels as a cost function in the learning algorithm. Our simulations on BCI competition IV, dataset 1 reveals that the proposed method significantly outperforms the conventional CSP and filter bank CSP (FBCSP) with two different filter bank architectures.