EEG classification for motor imagery BCI using phase-only features extracted by independent component analysis

The accurate classification of the electroencephalography (EEG) signals is the most important task towards the development of a reliable motor imagery brain-computer interface (MI-BCI) system. In this study, we utilized a publically available BCI Competition-IV 2008 dataset IIa. This study address to the binary classification problem of the motor imagery EEG data by using a sigmoid activation function-based extreme learning machines (ELM). We proposed a novel method of extracting the features from the EEG signals by first applying the independent component analysis (ICA) on the time series data and transforming the ICA time series data into Fourier domain and then extract the phase information from the Fourier spectrum. This phase information was further used to calculate the maximized cross-correlation connectivity matrix. The upper diagonal of this matrix was then vectorized and it serves as the basic feature for the ELM classification framework. By using the phase-only features we achieved 97.80% (p < 0.0022) nested cross-validated classification accuracy. In addition, this process is relatively computationally inexpensive. Thus, it can be an excellent candidate for the motor imagery BCI applications.