A Hierarchical Neural Network on Riemannian Manifold and Convolutional Neural Network for Sleep Stage Classification

The automation of sleep stage classification from polysomnography (PSG) signals is a critical task in sleep research, traditionally reliant on the manual extraction of features followed by a classification algorithm. Our study presents an advanced hierarchical neural network method that refines the automatic classification process into five distinct sleep stages using multi-channel PSG signals. This model operationalizes a two-stage process: the initial stage merges Riemannian manifold features with deep network-trained features to construct an enriched feature matrix. The use of Riemannian manifold methods harnesses the geometric properties of physiological signals for a robust feature set. The second stage employs a pair of fully connected layers with a sequence correlation mechanism that delineates the temporal dependencies between sleep stages. Also, it enhances the ability of the mode to discern between transitions in sleep epochs. Our framework is rigorously evaluated on two publicly available sleep datasets, namely ISRUC-S3 and DREAMS Subject. The results exhibit promising outcomes, with an overall accuracy of 0.837 and an F1-score of 0.829 on the ISRUC-S3 dataset, as well as with an overall accuracy of 0.877 and a Kappa coefficient of 0.838 on the DREAMS Subjects dataset. These results signify substantial advancements over traditional methods and showcase our method in effectively interpreting complex, multivariate PSG data for precise sleep stage classification, holding significant potential for clinical diagnostics and sleep disorder treatment.