The Research On Compression And Reconstruction Algorithm Of Electrocardiogram Signals

The prevalence and mortality of CVD in China is still increasing,and remote ECG monitoring plays a very important role in the prevention and treatment of CVD.Patients utilize wearable devices to collect personal ECG signals uninterrupted,and transmit the signals to the cloud to give a detection report.In the field of signal processing,as a breakthrough technique to break Nyquist’s sampling theorem,the compressed sensing technique,which not only alleviates the complexity of the signal acquisition end but also reduces the energy required during the retransmission process,is highly attractive for the application of wearable devices with limited resources in remote ECG monitoring.In compressed sensing technology,the key issue is how to guarantee a high accuracy reconstruction signal,so the reconstruction algorithm has been a hot issue in the research field of compressed sensing.However,traditional ECG compressed sensed reconstruction algorithms often require complex signal processing and prior knowledge,and the reconstruction process is time-consuming,which limits the application of compressed sensing in remote ECG monitoring systems.In order to solve the problems of traditional ECG compressed sensing reconstruction algorithms,this paper explores a data-driven deep learning reconstruction algorithm CSNet.It models directly the mapping relationship between the transposed projection signal and the original signal of an measurement vector(that is,a compression signal)and does not need to consider any prior information about the ECG signal.Specifically,inspired by transposition convolution,first,make matrix multiplication between the measurement vector and transposition vector of the sensing matrix,which guarantees that the signal of the input network and the original signal have the same shape.The CNN was then utilized for initial reconstruction.Next the quality of the initial reconstructed signal was further improved using LSTM to realize the function of nonlinear signal reconstruction.Finally,the data set of MIT BIH arrhythmia is verified.Compared with traditional ECG compression sensing reconstruction algorithm,CSNet can improve the reconstruction speed by 26 times at least when the reconstruction quality meets the clinical requirements.Also,the robustness was validated in larger scale ECG datasets(MIT-BIH NSRDB,AFDB and EDB).The ultimate aim of the remote ECG monitoring system is to analyze the ECG signal,and at present,the field of ECG compressed sensing is to study only the remodeling algorithms without further analysis of the remodeling signal.In this paper,we propose a CS-Res Net model for AF identification by addressing the problem of AF detection to verify the availability of remodeling signals.Specifically,the original signal was first classified using CS-Res Net,while comparing with the state-of-the-art method to demonstrate the validity of the model,and then the reconstructed signal with different compression ratios was classified using CS-Res Net,compared with the classification results of the original signal,even if the original signal was compressed to 1 / 10,AF detection performance was less than 1% from the original signal;most The problem of interpretability of the reconstructed signal similar to the original signal classification results was investigated using the Grad-CAM method,informed by visualization of the reconstructed signal at different compression ratios with the original signal as CS-Res Net input,and the model’s discrimination between AF and non AF signals was identified by the presence or absence of P-waves,even when the compressive magnitude was large,as long as the reconstructed signal could be included in the classification model The presence or absence of the P wave was identified.The experimental results proved that the reconstruction algorithm is sufficiently effective to guarantee that the reconstruction signals are identified critical regions by the deep learning model.