Multimodal Sleep Signals-Based Automated Sleep Arousal Detection

Abstract

Excessive sleep arousal affects one's sleep quality that would induce disease. Polysomnography is a powerful tool for sleep related monitoring. Clinically, there are being two kinds of causes on sleep arousal. One is apnea and hypopnea related arousal and the other is non-apnea and non-hypopnea arousal. The latter is relatively hidden and is difficult to determine in clinical. We aim to classify the sleep arousal caused by non-apnea and non-hypopnea from apnea and hypopnea related arousal. We propose an improved ensemble deep learning architecture that use a positional embedding based multi-head attention method to keep temporal relations of multimodal physiological signals. The experimental datasets are based on an open access dataset from the public cardiology challenge 2018. We conduct several groups of comparison experiments among our proposed convolutional-residual network with positional embedding and multi-head attention (CRPEMA) method and other methods that includes methods presented on the cardiology challenge 2018. The results show that CRPEMA has high efficiency and accuracy. When the parameters decrease by more than 50%, the accuracy is keeping improved. Experiment results reflect that CRPEMA outperforms others and obtains the Area Under the Precision-Recall curve (AUPRC) of 0.391.