Research On OSA Automatic Detection Algorithm Based On ECG Signal

Sleep Apnea is a common sleep-related breathing disorder,mainly with obstructive sleep apnea(OSA)which manifested as a partial or complete decrease in respiratory airflow from normal level.OSA always promote people to wake up from normal sleep and affect people's sleep quality seriously.Sleep apnea is also an important and independent factor of many chronic diseases such as high blood pressure and diabetes.Polysomnography is commonly used in diagnosis for sleep apnea in clinical,but it is not widely applied with drawback of high cost and uncomfortable process during detection of sleep apnea.As a result,more than 85% patients with sleep apnea did not get a diagnosis and treatment timely.Therefore,A lot of OSA detection methods based on single channel signal have been proposed in recent years,and Electrocardiograph(ECG)signal with advantage of high associated with OSA and non-intrusive acquisition way is the most popular signal among these single signals.However,the existing methods of OSA detection based on ECG are less effective in detecting OSA on high noise ECG data and large databases.This article studies the method of automatic detection of OSA which can be used in high-noise ECG data and large database.This paper extracts multiple shallow feature signals composed of many discrete points picked up from ECG signal as data basis which only require to detect R peak in ECG signal,with a high tolerance of noise in ECG data so that reducing influence of ECG signal noise on OSA detection.In addition,this article put forward three kind of signal preprocessing methods applied in shallow feature signal.The first method is enlarging amplitude of each signal to highlight the change of shallow feature signal.The second method is using one-dimensional convolution on each shallow feature signal to increase the amount of feature signal.The third method is combining of multiple convolution feature signal as input feature to make up the problem of single shallow feature only reflecting one side of ECG signal.Three preprocessing can effectively solve the problem of less information extracting from ECG signal.Because of the advantage of memory ability of long-term features and good generalized ability in long-short term memory recurrent neural network(LSTM-RNN),this paper uses LSTM-RNN to learn the fused convolution feature signals and classify per-segment ECG,which improving the accuracy of OSA detection.Experimental results show that the proposed methods is effective in detecting OSA on high noise ECG data and large databases.