Deep Convolutional Neural Network Based Eye States Classification Using Ear-EEG

Electroencephalography measured around the ear (ear-EEG) has been considered as an effective measurement for the development of practical EEG-based applications because it has convenience compared to the conventional scalp-EEGs in terms of EEG measurement. However, ear-EEG-based applications have presented the classification accuracy lower than those of scalp-EEG-based applications. In this study, we introduced deep convolutional neural networks (CNNs) to improve the overall performance of the ear-EEG-based application. Ear- and scalp-EEGs were simultaneously taken while 30 participants performed an experiment for eye-state identification (eyes-open and eyes-closed) for two different days. The classification of eyes-open and eyes-closed states can be used to develop various real-life applications. The cross-validated (CV) and test–retest (TR) accuracies of a conventional machine learning algorithm with the best classification performance were first obtained for the ear- and scalp-EEG. We then estimated classification accuracies using three different CNN models (EEGNet, deep ConvNet, and shallow ConvNet). The shallow ConvNet showed the best classification performance compared to other CNN models and significantly outperformed the classification accuracy of the conventional algorithm using ear-EEG. Furthermore, the classification performance of the shallow ConvNet using ear-EEG was mostly the same as that of the conventional algorithm using scalp-EEG. The shallow ConvNet based on ear-EEG also exhibited very reliable eye-state identification in a pseudo-online simulation, with a true positive rate of 93%, a false positive rate of 0.29 FPs/min, an eye-state detection speed of 2.35 sec, and an information transfer rate of 21.86 bits/min. These experimental results validated that the CNN models can be effectively employed to improve the performance of ear-EEG-based applications.