Deep Learning with Convolutional Neural Networks for Sleep Arousal Detection

Jia Dongya1, shengfeng yu2, Cong Yan3, Wei Zhao4, Jing Hu3, Hongmei Wang5, Tianyuan You1
1Guangzhou Shiyuan Electronics co., ltd, 2South China University of Technology, 3Guangzhou Shiyuan Electronics Technology Co., Ltd., 4Guangzhou shiyuan electronics co.,Ltd; Guangzhou Xicoo Medical Technology Co.,Ltd, 5Guangzhou Shiyuan Electronics Technology Co., Ltd


Abstract

Aims: Sleep arousal is one dominating reason for poor sleep quality. And the diagnosis of sleep arousal is the precondition for the further treatment and research of sleep disorder. The Physionet Challenge 2018 aims to correctly classify various arousal regions across time based on multiple physiological signals including electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), electrocardiograph (EKG) and oxygen saturation (SaO2). Therefore, we developed a method to detect the sleep arousal regions based on these multiple physiological signals.

Methods: Our algorithm was composed of two steps: feature extraction and classification. In the step of feature extraction, we extracted several kinds of features from the physiological signals. On the EEG, the features based on the energy ratio of specific frequency band, the scaling exponents, energy spectrum and entropy were extracted. The features based on the amplitude and Kolmogorov complexity were calculated on the EMG. Besides, the variance of the SaO2 and the heart rate measured on the EKG were adopted. After the feature extraction, the support vector machine with the kernel of the radial basis function was used as the classifier.

Results: The proposed model was trained and test on the training set and test set provided by Physionet Challenge 2018, which contained 994 and 989 records respectively. Our method achieved an AUROC of 0.5 and AUPRC of 0.027.

Conclusion: In this study, we proposed a multiple physiological signal based classification algorithm to recognize arousal region via several kinds of hand-crafted features. Experimental results showed that our method would be effective.