MSleepNet: A Semi-Supervision-Based Multiview Hybrid Neural Network for Simultaneous Sleep Arousal and Sleep Stage Detection

The complexity of sleep disorder diagnosis continuously increases the clinical requirement for simultaneous measurement of sleep arousal and sleep stage, which, however, has not received enough attention from engineering scholars. To the best of our knowledge, all previous machine learning-based detection methods use the single view mechanism to identify either sleep arousal or sleep stage. In this article, a multiview hybrid neural network with semi-supervised learning (SSL), named as MSleepNet, is proposed for simultaneous sleep arousal and sleep stage detection. In particular, the features of single-channel electroencephalography (EEG) signal from both the time domain and frequency domain are extracted by the improved residual backbone network. Then, an attention mechanism is introduced to enhance the feature recognition ability in the frequency domain. A multitask classification loss function is also designed to synchronously consider the correlation between sleep arousal and sleep stage. In the classification part of the network, the supervised loss and semi-supervised loss for each task are effectively combined to alleviate the data imbalance problem and improve the classification accuracy. Overnight polysomnographic recordings from two public datasets (sleep heart health study (SHHS), $n=200$ and Physio2018, $n=100$ ) and one dataset from a local clinic ( $n=60$ ) were applied to validate MSleepNet. Experimental results on Physio2018 dataset demonstrate that MSleepNet achieves an overall accuracy of 0.78 and $F1$ score of 0.73 for sleep staging, and area under the precision-recall curve (AUPRC) of 0.39 and an area under the receiver operating characteristic curve (AUROC) of 0.75 for sleep arousal detection. Therefore, MSleepNet framework based on multiview hybrid neural network has shown its potential in intelligent sleep monitory for synchronizing sleep arousal and sleep stage measurement with only very limited labeled data.