Research And Hardware Implementation Of Sleep Apnea Detection Algorithm Based On ECG Signal

Sleep apnea(SA)not only affects people’s sleep quality and increase the potential risk of driving accidents,but also may cause cardiovascular disease and renal damage,threating people’s health.In order to detect SA more conveniently and replace expensive polysomnography,it is necessary to develop a lightweight SA detection algorithm and low-power hardware so that they can be embedded in wearable health monitoring devices.To address this issue,a SA detection algorithm based on electrocardiogram(ECG)signals and corresponding low-power hardware design is proposed in this thesis.Firstly,as heart rate variability(HRV)and ECG-derived respiratory(EDR)features,the RR interval and R-S amplitude of each heartbeat in the ECG signal are extracted by the proposed QRS complex detection algorithm which is improved through the refractory refreshing mechanism.Then,a convolutional neural network(CNN)is adopted to classify whether a SA event occurs based on these two features and the size of the CNN model is optimized to reduce the overhead of wearable devices.Finally,a SA detection processor is designed for wearable health monitoring devices.The contributions of this thesis are as follows:First of all,the refractory period method in the conventional threshold-based QRS complex detection method is improved with refractory period refreshing,which makes the method more effectively.With refractory period refreshing,the accuracy of the improved QRS detection is 99.24% on MIT-BIH arrhythmia Database.Secondly,a features fusion method based on point-wise convolution is adopted to improve the classification performance of the CNN for classification task based on multifeature.In addition,through the multi-period classification method based on network block,the amount of calculation and parameters of CNN is reduced without decrease of classification performance.Experimental results show that the proposed SA detection algorithm achieves an accuracy of 86.41% on the Apnea-ECG dataset.Thirdly,the low-power SA detection processor is designed based on the proposed algorithm,including a FIR filter with configurable coefficients and orders,a QRS complex detector,and a reconfigurable neural network computing engine that supports multiple neural network structures.The design is implemented on FPGA and the power of the processor is evaluated through EDA tools.The result of experiment shows the dynamic power of the processor is only 9mW.