Deep learning scheme with attention mechanism for motor imagery task classification based on efficient channel selection and time-frequency analysis of eeg signal

Abstract

Brain-computer interface (BCI) systems can serve as a means of communication for the motor-disabled people, for example BCI-driven control devices used in wheelchairs. Motor imagery (MI) signals collected from the electroencephalogram (EEG) are gener- ally classified to control the BCI devices. One major concern here is to get higher clas- sification accuracy for different MI tasks. In this case, deep learning-based algorithms are getting popularity because of their capability to provide improved decoding accu- racy. In this thesis, convolutional neural network (CNN) based deep learning networks with different types of spatial/temporal attention mechanisms are proposed to classify MI tasks utilizing the multi-channel EEG signals and their transformed versions. First, a spatial-temporal attention based network (STAMI-Net) is proposed where a point- wise attention (PAT2D) module operated multi-channel 2D representations of the MI EEG data are applied. In the STAMI-Net, multiple spatial and temporal convolutional operations are performed on the input 2D data and then a squeeze and excitation spatial attention network (SES-net) is applied to get channel-wise attention. Another point is that the use of an efficient channel selection method on the large number of channels of multi-channel EEG data can dramatically reduce the computational problems. Apart from providing channel attention in the proposed network, a channel selection strategy is proposed based on neurophysiological aspects of different regions of the brain. Next, a deep learning network based on the combined use of five vigilance bands of MI-EEG signals (namely, ViBMINet) is proposed to investigate the classification performance. The proposed network performs spatial and temporal operations on band-limited MI- EEG signals, which not only extracts inter-channel relationship but also utilizes inter- band relationship to generate effective features. Instead of using the five bands, a dual- band CNN (DBCNN) architecture is proposed utilizing a wide-band and narrow-band MI-EEG signals. As an alternate, considering the advantages of multi-resolution time- frequency decomposition, discrete wavelet transform (DWT) of the MI-EEG data is performed and the first level DWT coefficients are applied to a deep learning network, namely STOC-Net. An extensive experimentation is performed on multiple subjects taken from publicly available BCI Competition IV 2a and 2b MI-based EEG datasets. The proposed STAMI-Net, ViBMI-Net and STOC-Net models offer classification ac- curacies of 84.88%, 84.79% and 84.65%, respectively for dataset IV 2a and 80.70%, 81.68% and 82.21% for dataset IV 2b, respectively, which are higher than those ob- tained by some state-of-the-art methods.