Automated sleep stage classification in sleep apnoea using convolutional neural networks. (2021)

• Record Type: Journal Article
• Authors: Sundar, G. Naveen; Narmadha, D.; Anton Jone, A. Amir; Sagayam, K. Martin; Dang, Hien; Pomplun, Marc
• Is Part Of: Informatics in medicine unlocked. Volume 26(2022)
• Publication Date: 2021
• Subjects: Polysomnography -- Electroencephalogram -- Electrooculogram -- Electromyogram analysis -- Sleep stage classification -- Convolutional neural networks; Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/j.imu.2021.100724
• Languages: English
• ISSNs: 2352-9148

Abstract

Abstract: A sleep disorder is a condition that adversely impacts one's ability to sleep well on a regular schedule. It also occurs as a consequence of numerous neurological sicknesses. These types of disorders can be investigated using laboratory-based polysomnography (PSG) signals. The detection of neurological disorders is exact and efficient thanks to the automated monitoring of sleep relegation stages. This automation method publicly presents a flexible deep learning model and machine learning approach utilizing raw electroencephalogram (EEG) signals. The deep learning model is a Deep Convolutional Neural Network (CNN) that analyses invariant time capacities and frequency actualities and collects assessment adaptations. It also captures the inviolate and long brief length setting conditions between the epochs and the degree of sleep stage relegation. This method uses an innovative function to calculate data loss and misclassified errors found while training the network for the sleep stage, considering the restrictions found in the publicly available sleep datasets. It is used in conjunction with machine learning techniques to forecast the best approach for the process. Its effectiveness is determined by using two open-source, public databases available from PhysioNet: two recordings with 5402 epoch counts. The technique used in this approach achieves an accuracy of 90.70%, precision of 90.50%, recall of 92.70%, and F-measure of 90.60%. The proposed method is more significant than existing models like AlexNet, ResNet, VGGNet, and LeNet. The comparative study of the models could be adopted for clinical use and modified based on the requirements.