An Automatic Sleep Staging Method Using a Multi-head and Sequence Network

Sleep plays an important role in human life and sleep disorders are becoming a major public health issue. Identifying the sleep stages correctly with the polysomnogram, which is a collection of relevant physiological signals recorded during sleep, helps clinicians diagnose and treat sleep disorders. Nevertheless, the manual process of performing such a sleep staging task is very time consuming and prone to technicians' labeling errors. To tackle these issues, there have been many efforts in designing various machine learning/ deep learning-based automated methods for sleep staging over the past years, this paper proposed an innovative automatic sleep staging method based on a hybrid multi-head and sequence network architecture. Different from previous works, this method extracted multi-modality features from EEG and EOG with the multi-head CNNs and then integrated them with heuristic weights obtained through experiments. Moreover, the joint loss function based on cross entropy loss function was also applied, wherein multiple embedded losses were added simultaneously and propagated backward. We tested our method using the popular public dataset "Sleep-EDF[Expanded]" from the PhysioNet website. The best result showed an accuracy of 91.76% and Cohen's kappa value over 0.836 compared with human experts, which had competitive advantages over most of previous research results using automated methods on the same dataset.