Hybrid CNN-RNN in Motor Imagery Identification of Brain-Computer Interface

Abstract

Brain-computer interface (BCI) is a technology that develops human and machine interactions. BCI allows the brain to move external devices without gestures, muscles, and sounds. This technology has great benefits, such as biomedical applications, neural rehabilitation, and entertainment applications. BCI depends on the ability of intermediate devices to translate brain commands, whether consciously or not, to select the appropriate action. The instrument most often used in BCI is the Electroencephalogram (EEG), so BCI-EEG seems inseparable. BCI actions can be Motor Imagery variables, emotions, or focus. Usually, the Motor Imagery variable is carried out in a conscious state, making it easier to control. The identifying Motor Imagery variables in the EEG signal needs to be improved continuously. First, an EEG signal needs to be extracted representing the variable under consideration, the Motor Imagery. Usually, the extraction of frequencies containing Beta and Mu waves is carried out. The next problem is that the multi-channel use of EEG recording resulted in data redundancy. Research with similar data has been discussed using Independent Component Analysis (ICA) but has not paid attention to the sequence. This study proposed the Hybrid Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) methods as a multi-channel identification and handling method that considers the sequences of EEG signal data to classify BCI into four classes. Experiments using Hybrid CNN and RNN resulted in an accuracy of 98.62% and resulted in the shortest computation time compared to previous studies with similar data. We also experiment with the use of wavelets and some optimization weight models.