Intelligent Screening Research Of Obstructive Sleep Apnea Syndrome Based On Deep Learning

Sleep is an indispensable part of people’s lives and is an important guarantee to regulate the regularity of human life.With the accelerated pace of people’s lives and increased work and psychological stress,more and more people suffer from sleep-disordered breathing disease.Obstructive Sleep Apnea Syndrome(OSAS)is a common sleep-disordered breathing disease in which patients suffer from frequent sleep apnea at night,disrupting the sleep structure and leading to sleep structure disorders,making them prone to daytime sleepiness and memory loss.As the number of potential OSAS patients is increasing,longterm low-quality sleep seriously affects their life pattern and work and rest,and in serious cases,even life threatening,so early screening of OSAS is becoming more and more important.Polysomnography is the gold standard for clinical diagnosis of OSAS,but the procedure is complicated,time-consuming and expensive,causing great inconvenience to patients and physicians.However,other portable screening methods such as the Epworth Sleepiness Scale are too single and have poor screening effects.Therefore,this thesis proposed a multifactorial integrated of OSAS classification approach,which integrated sleep apnea situation,sleep structure disorder degree and other factors,and it is applied to OSAS intelligent screening system.Firstly,this thesis proposed an automatic sleep staging method with a multi-scale attention mechanism model.For ECG signal data,the method improved the feature information of each ECG segment by fusing the heart rate variability features at different size feature window scales;combined with the attention mechanism to fully extracted the dependency and attention relationship among the units of the feature sequence,it compensated for the limitation problem of feature window length;further,an improved cross-entropy loss function is proposed to solve the analogical imbalance problem existing in the sleep staging task.The method not only enhanced the model feature learning ability,but also alleviated the problems of category skewing and misclassification,which effectively improved the classification effect of each sleep stage in sleep staging.Secondly,this thesis designed a Spatio-temporal Convolutional Neural Network(StCNN),then proposed an automatic sleep apnea event detection method with a multifeature fusion StCNN-BLSTM model.For the ECG signal data,the method fused two shallow features of R-wave interval and derived respiratory signal,which reflects the ECG signal information from two different perspectives of heart rate variability and respiratory signal.At the same time,the local spatiotemporal features of each ECG cycle are extracted from the fused signals according to the ECG physiological sense by the StCNN network,and then the global temporal features are learned by combining the BLSTN network,which fully exploits the ECG feature information.This method combined the ECG physiological sense to extract local spatiotemporal features and global timing features successively,which effectively improved the sleep apnea recognition accuracy.Finally,this thesis proposed a multifactorial integrated of OSAS classification approach,which integrated sleep apnea index,sleep structure disorder degree,Epworth sleepiness score and height and weight index,and it avoided the problem of low classification accuracy due to imprecise sleep apnea index obtained by non-clinical methods,and had high feasibility.Based on the OSAS intelligent screening method,this thesis further designed and implemented an OSAS intelligent screening system.The system can not only meet the demand of screening OSAS at home,but also effectively screen OSAS patients intelligently,which provides a strong support for the early screening of OSAS patients.