Comparison of Automatic Sleep Stage Classification Methods for Clinical Use

Abstract

Sleep stage scoring is necessary for diagnosing several sleep disorders. However, it is an intensive and repetitive task and a vital automation candidate. This work seeks to evaluate different kinds of Machine Learning based classification algorithms available in the scientific literature to determine which one fits better the clinical practice requirements. The comparison is made with a predefined experimental design, using electroencephalography, electrooculography, and electromyography signals from the polysomnographic records of the Sleep-EDFx dataset. The comparison considers the accuracy and speed of algorithms based on Linear Discriminate Analysis, Support Vector Machines, Random Forests, and Artificial Neural Networks. The latter group includes the Deep Neural Networks DeapFeatureNet, based on Convolutional Neural Networks, and DeepSleepNet, additionally based on Recurrent Neural Networks. It is determined that several of the tested algorithms boast high accuracy levels (85%). From them, DeepSleepNet is chosen as the fittest due to its considerable advantage in execution time. Nevertheless, the final result should always be reviewed by the experts.