Research Of Sleep Stage Classification Algorithm

The division of sleep into different sleep stages is a vital part of sleep monitoring and sleep research.Manual sleep staging process is a time-consuming and tedious task with low inter-rater reliability,and automatic sleep staging based on sleep electrophysiological signal is widely used to enable sleep scoring and consistent sleep staging results.Portable sleep monitoring requires fewer signal electrodes than sleep research to meet portability and comfort demand,there seems to be a trade-off at sleep staging accuracy.In this study,three sleep staging machine learning models using Sleep-EDFx database of the Biomedical Signal Research Resource website PhysioNet(www.physionet.org)were trained,and a sleep staging model with temporal dependency of sleep stages transition proposed in this paper is able to achieve state-of-the-art performance.Paper research content is as listed:1,According to sleep expert knowledge,11 feature values were obtained,including three time domain features,four time-frequency features,three nonlinear features and one complexity feature of single-channel EEG signal.A sleep staging support vector machine was trained on the feature space,and obtained a gentle performance(accuracy: 74.49%,F1 score: 61.99%).2,According to low-scale decomposition of the 30-second sleep period idea,an automatic feature extraction sleep staging model based on convolutional neural network with three-layer convolutional layer was trained.The trained model was used to compare the effects of different sleep signal combinations on sleep staging performance.The result showed that the sleep staging model trained by the combination of EEG,EMG,and EOG performed best(accuracy: 79.14%,F1 score: 70.13%).In addition,a convolutional neural network based on single-channel EEG signal got a better performance(accuracy: 76.53%,F1 score: 66.34%)than the sleep staging support vector machine model,indicating that automatic feature extraction on sleep data was possible and learned features similar to the ones describled in the sleep scoring manual of the American Academy of Sleep Medicine.3,This study combined a long short-term memory recurrent neural network with the features mapped by convolutional neural network to introduce the temporal dependency of sleep stages transition,and an automatic sleep staging model with sleep temporal dependence was trained.A model with temporal dependency trained using EEG,EOG and EMG got state-of-the-art performance(accuracy: 85.56%,F1 score: 77.61%),and A model with temporal dependency trained using single-channel EEG signal also got a good performance(accuracy: 84.59%,F1 score: 75.09%).Both of the two model's performance reached the human sleep scoring criteria.The model based on single-channel EEG signal can be well applied to portable sleep monitoring.