Effect of Electrooculography on Electroencephalography Classifying Accuracy in Deep Learning and Reducing Number of Channels in Motor-Imagery Brain-Computer Interface

Recently, brain-computer interface (BCI) control using deep learning has been investigated. The BCI control is realized by analyzing brain activity through a computer and reflecting it to an effector. There are several methods to measure brain activity such as electroencephalography (EEG) and functional magnetic resonance imaging; however, EEG-based BCI has been developed and studied because it is easier to measure. However, EEG signals contain many artifacts such as electrooculography (EOG), electromyography (EMG) and commercial power supply noise, which deteriorate the SNR. Further, it is not possible to identify the frequency of EOG and EMG forehand; therefore, it is not possible to remove them completely from EEG. This is a critical problem for implementing BCI to residential environments because it is impossible to stop the movements of the user's body for practical use. Here, we focused on EOG and investigated the effect of pseudo-EOG on classifying accuracy by adding pseudo-EOG to steady state visual evoked potential (SSVEP) and motor imagery (MI). The results showed that EOG had no effect on the classifying accuracy of SSVEP but had an effect on the classifying accuracy of MI. For the fast classification of MI, the number of channels was reduced to lower the computational cost for implementing edge computing for the practical use in residential environment. Only 2% degradation was observed between the accuracy of deep learning with 22 channels and those with 15 channels.