A Comparison of LSTM and CNN Performance in EEG Motor Imagery with Application to Edge Computing Non-invasive Brain-computer Interface Possibilities

A Brain-computer interfaces (BCIs) can translate bioelectric signals in the human brain into instructions for decision-making or controlling the environment around them. As such, BCIs can be applied to assist in the medical care of people with limited mobility. However, BCIs are usually bulky due to their low level of integration. In this work, we use electroencephalography (EEG) datasets from motor imagery experiments captured by a non-invasive EEG device to compare the ability of Convolutional Neural Networks (CNNs) and Deep Learning Long Short-Term Memory Classifiers (LSTMs) to detect the recognition of both left-handed and right-handed EEG motor imagery and to discuss their feasibility for Brain-Computer Interface applications. Our results show a validation accuracy of 0.622 for correct classification. In addition, this paper proposes a possible architecture for edge computing brain-computer interfaces that helps to improve usability and mobility. These results illustrate that the combination of deep learning, brain-computer interfaces and edge computing can help to achieve the generalization of non-invasive brain-computer interfaces.