A Comparative Analysis of CNNs and RNNs for EEG-based Motor Imagery Classification in BCIs

Abstract

EEG-based motor imagery (MI) classification plays a vital role in brain-computer interface systems (BCIs) to enable the control of external devices with the human brain. However, there is currently limited research focusing on the comparison between different machine learning models for this task. This research paper aims to present a comprehensive comparative analysis of two popular deep learning architectures, Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) for MI recognition with EEG data. The experiments utilised the EEG Motor Movement/Imagery Dataset v1.0.0 from PhysioNet, which contains EEG signals recorded during a variety of motor imagery tasks. The respective performances of the CNN and RNN architectures were subsequently evaluated and compared based on classification accuracy and computational efficiency. Various metrics and statistics, namely accuracy, precision, training speed, memory usage, etc., were used for assessment. The results revealed that CNN outperforms RNN in terms of accuracy, while RNN demonstrates superior computational efficiency. These findings potentially serve as a valuable guideline for researchers and practitioners in the field of BCIs, aiding them in selecting the most suitable neural network architecture for performing MI related tasks.