Motor Imagery EEG Signals Decoding by Multivariate Empirical Wavelet Transform-Based Framework for Robust Brain–Computer Interfaces

Muhammad Tariq Sadiq,Xiaojun Yu,Zhaohui Yuan,Fan Zeming,Ateeq Ur Rehman,Inam Ullah,Guoqi Li,Gaoxi Xiao

DOI: https://doi.org/10.1109/access.2019.2956018

IF: 3.9

2019-01-01

IEEE Access

Abstract:The robustness and computational load are the key challenges in motor imagery (MI) based on electroencephalography (EEG) signals to decode for the development of practical brain-computer interface (BCI) systems. In this study, we propose a robust and simple automated multivariate empirical wavelet transform (MEWT) algorithm for the decoding of different MI tasks. The main contributions of this study are four-fold. First, the multiscale principal component analysis method is utilized in the preprocessing module to obtain robustness against noise. Second, a novel automated channel selection strategy is proposed and then is further verified with comprehensive comparisons among three different strategies for decoding channel combination selection. Third, a sub-band alignment method by utilizing MEWT is adopted to obtain joint instantaneous amplitude and frequency components for the first time in MI applications. Four, a robust correlation-based feature selection strategy is applied to largely reduce the system complexity and computational load. Extensive experiments for subject-specific and subject independent cases are conducted with the three-benchmark datasets from BCI competition III to evaluate the performances of the proposed method by employing typical machine-learning classifiers. For subject-specific case, experimental results show that an average sensitivity, specificity and classification accuracy of 98% was achieved by employing multilayer perceptron neural networks, logistic model tree and least-square support vector machine (LS-SVM) classifiers, respectively for three datasets, resulting in an improvement of upto 23.50% in classification accuracy as compared with other existing method. While an average sensitivity, specificity and classification accuracy of 93%, 92.1% and 91.4% was achieved for subject independent case by employing LS-SVM classifier for all datasets with an increase of up to 18.14% relative to other existing methods. Results also show that our proposed algorithm provides a classification accuracy of 100% for subjects with small training size in subject-specific case, and for subject independent case by employing a single source subject. Such satisfactory results demonstrate the great potential of the proposed MEWT algorithm for practical MI EEG signals classification.