Multivariate Input based Convolutional Neural Network for Motor Imagery Classification

Abstract

Convolutional neural networks (CNNs) recently have been successfully applied to electroencephalography (EEG) based brain-computer interfaces (BCIs), especially for motor imagery (MI) classification. However, the design of input representation for the CNNs model remains challenging, considering the high dimensionality and subject-specific feature space of EEG signals. It is urgent to construct an appropriate representation to enhance the decoding performance of CNN methods. To this end, we propose a multivariate input-based CNN model in this study. First, feature representations of high dimensionality are constructed to preserve subject-dependent information. In detail, multivariate features consist of multiple covariance matrices, each of which contains the spatial information of MI-based EEG signals under a subject-optimized band. Second, a CNN model is then designed and optimized according to the input of the multivariate representations, for efficiently learning required for classification. The proposed method achieves an average accuracy of 87.55% $(\pm 6.99)$ on the data of BCI competition IV 2a, exhibiting reliable classification performance and great potential use in BCIs.