A novel multi-branch hybrid neural network for motor imagery EEG signal classification

Abstract

As a typical spontaneous brain-computer interface system, motor imagery has been widely used in areas such as robot control and stroke rehabilitation. Recently, researchers have started to study and propose various Convolutional Neural Network (CNN) structures based on motor imagery signals and have obtained better decoding accuracy compared with traditional machine learning algorithms. In this paper, we focus on a four-class motor imagery task. Therefore, we propose an end-to-end novel multi-branch hybrid neural network for motor imagery EEG signal classification. Notably, we divide the input signal into four frequency bands associated with the motor imagery signal. Meanwhile, we also introduced a Bidirectional Gated Recurrent Unit (BGRU) to identify EEG features. Nevertheless, it is extremely difficult to collect high-quality EEG data and the classification accuracy is degraded due to the strict requirements of the subjects and the experimental environment. To address this issue, we propose a novel data augmentation approach for frequency domain segmentation swap (Seg-Swap) for improving EEG motor imagery signal classification accuracy. We make use of the publicly available BCI IV 2a dataset and High Gamma dataset to evaluate the decoding performance of the model. The proposed multi-branch hybrid neural network (MBHNN) achieves 86.15% and 95.04% on the both datasets. Compared with several state-of-the-art (SOA) algorithms, our proposed model has a slight improvement in classification accuracy. Experimental results show that our proposed MBHNN has higher decoding classification and stronger robustness.