EEG-based 5- and 2-class CNN for Sleep Stage Classification

Abstract

This paper investigates the performance of automatic sleep stage classification with automatically selected features from electroencephalographic (EEG) signals using a Convolutional Neural Network (CNN) based on 5- and 2-class models. We defined two ways for 2-class stratification, to classify the sleep stages and compare its performance with predictions made using a 5-class model. All the models were created using a CNN called EEGNet, and the experiments were carried out with the ISRUC-Sleep public dataset, which consist of 100 subjects and 6 EEG channels. With a single 5-class model for the entire dataset, we have obtained an average area under the receiver operating characteristic (AUROC) of 0.948. In the best case, we have obtained an average AUROC of 0.964, 0.967, 0.982 and 0.929 for the stratified 2-class models: Awake vs Sleep (Rapid Eye Movement (REM) + Non-REM [N1+N2+N3]), REM vs Non-REM (N1+N2+N3), (N1+N2) vs N3, and finally N1 vs N2. We have shown that in the four 2-class stratification-based models in a row, we can achieve an average AUROC of 0.97. The results obtained are promising and can lead to possible combinations of the 5- and 2-class models to improve the automatic sleep stage classification.