Automatic sleep stage classification: From classical machine learning methods to deep learning

Abstract

Background and objectives: The classification of sleep stages is a preliminary exam that contributes to the diagnosis of possible sleep disorders. However, it is a tedious and time-consuming task when conducted manually by experts. Many studies explored ways of automating polysomnogram signals analysis. They are based on two main strategies: conventional machine learning and deep learning methods. The objective of this work is to carry out a comparative study on these two classes of models. Method: A primary comparison of performance of these classifiers is carried out using eight conventional machine learning algorithms and a feed-forward neural networks to assess whether this latter method have definitely supplanted the first. As sleep epochs show inter-epochs correlation, a study of the distinctive influence of this temporal dependence on the classifiers performance is then conducted introducing for this purpose (uni- and bi-directional) long short-term memory networks. In a context of generalization of the use of wearable devices, a comparison of the classification methods examined is also carried out in their accuracy when dealing with a reduced number of channels. Finally, the robustness of the results obtained to the choice of features selection algorithms is discussed. Results and conclusion: Our results show that support vector machine with radial basis function and random forest are just as valid for predicting sleep stages classification as feature-based neural networks with performance closed to the state of the art. This conclusion remains valid even after the introduction of inter-epochs temporal dependence, reduction of the number of channels or change in features selection method.