Improving EEG-Based Motor Imagery Classification Using Hybrid Neural Network

Abstract

Motor imagery EEG (MI-EEG) is a subjective signal generated by testers, which is collected through brain-computer interface (BCI). With the characteristics of noninvasive, inexpensive, and easily applied to human beings, MI-EEG classification is a popular research area in recent years. Due to the low signal-to-noise ratio and incomplete EEG signals, high accuracy rate classification is still a challenging problem. Most existing works of deep learning only regard EEG signals as chain-like sequences data and use single neural network for classification. To solve the above issues, we propose an improved EEG signals classification method via a hybrid neural network (HNN). In our work, we first use the origin EEG signals without removing noise and any filtering process, to ensure real-time property. Then, the EEG signals are divided into some small segments, and we arrange the data by considering the spatial position of electrodes. Finally, we propose a hybrid neural network by combing CNN, DNN, LSTM network. Experimental results for two challenging EEG signal classification benchmark datasets show that the proposed method has a good classification performance compared with several state-of-the-art EEG signal classification algorithms. After multiple sample testing, the average experiment result is 75.52%, which is 7.32% higher than the latest method.