Convolutional neural network for classification of multiple sleep stages from dual-channel EEG signals

Backgrounds and Objectives: Sleep-related disorders are critical diseases and they need to be proper diagnosed as early as possible. The major difficulty is that fewer medical experiments are available in remote locations to diagnose different types of sleep disorders. This research work presents an automated classification of sleep stages using deep learning techniques to diagnosis multiple sleep diseases from single-channel and different combinations of channels of EEG signals. Methods: In this proposed work, we involve three different forms of signal inputs for the automatic classification of sleep stages from EEG signals. The proposed research work presents a one-dimensional convolution neural network (1D-CNN) model for multiple sleep stages because of its high robustness for automatically classifying the sleep stages from brain signals without involving any types of feature extraction/selection, which is one of the challenging processes in the earlier literature. The proposed model contained seven convolution layers followed by two fully connected layers. The main objective for designing such a custom deep neural network is to improve classification accuracy performance with less number of learnable parameters. Results: The proposed model has used two subgroups of the ISRUC-Sleep dataset. We also obtain a k-cross-fold validation approach over the subjects, which ensure that there is no possible contamination in between training and testing. The experimental results for this proposed model for classification of five classes of sleep stages (wake, non-rapid eye movement N1-N3, and rapid eye movement). The proposed model was evaluated by classification accuracy, precision, sensitivity, F1score, and Cohen's Kappa score. The proposed 1D-CNN architecture achieved the highest classification accuracy of 95.85%(C3-A2),94.11%(O1-A2), and 97.22%(C3-A2+O1-A2) using the SG-I dataset, similarly, the same model reported accuracy using the SG-II dataset of 95.73%(C3-A2),94.02%(O1-A2), and 95.06% (C3-A2+O1-A2). Conclusions: Our proposed methodology is efficient and effective for multiple sleep staging. The proposed 1D-CNN model is ready for clinical purposes and can be managed with a huge number of polysomnography data.