A Hybrid Deep Learning Approach for Classifying EEG Time-Frequency Representation based on Data Augmentation and Enhancement

Abstract

EEG (Electroencephalography) is a signal that records the electrical activity of the brain. This experimental signal is used to evaluate the electrical activity of the brain, which is used to detect possible problems in communication between brain cells. In addition, EEG can be used for Human-Computer Interaction (HCI), medical diagnostics, and psychological research. Despite its important applications, EEG signals are difficult to process and classify due to their non-stationary nature, high noise, and high variability in recording conditions. In the meantime, EEG Time-Frequency Representations have many advantages for better representation of the signal as input to a classification system based on deep learning models. However, the gap in these methods is the lack of sufficient data for training and also the lack of attention to various features of the EEG signal. To fill these gaps, a hybrid deep learning method is proposed in this research for classifying EEG time-frequency representations based on data augmentation and data improvement. The first innovation of this research is the use of a CNN+LSTM hybrid network. The use of a hybrid network allows for the simultaneous use of both sequence and spatial features for classification. The second innovation is the use of data enhancement techniques using the Efficient Sub-Pixel Convolutional Neural Network (ESPCN) method. The third innovation is the use of data augmentation techniques. Simulation results on the BCI Competition IV dataset show that the three innovations of the proposed method lead to an improvement in the classification process of EEG time-frequency representations, achieving an accuracy of 88.23%.