Research On The GeneratGenion Of ECG Signals Based On Deep Erative Networks

The ECG signal diagnosis algorithms are of great significance to assistant medical diagnosis.Researchers often need to obtain a large number of effective,real and diverse biomedical data in the early stage to better train the algorithms and improve the performance of the algorithms,so as to make the algorithms more perfect in clinical application.However,the performance of current ECG signal diagnosis algorithms is greatly limited by the lack of data.As a result,many algorithms perform well when tested on a data set,but are not satisfactory when applied to clinical environment.In this paper,deep learning technology,which has developed rapidly in recent years,is used to conduct an in-depth study on the generation of ECG signals.The purpose is to expand the ECG data in terms of data quantity and data diversity,so as to improve the performance of ECG signal related diagnosis and processing algorithms.This paper proposes three algorithms for ECG signal generation,namely,WaveNet-based method,GAN-based method combined with short-time Fourier transform and GAN-based method combined with stationary wavelet transform.Firstly,this paper introduces the WaveNet-based approach.The data in the existing ECG database are compressed by u-law and encoded by one-hot.Then,the coded data are transmitted into WaveNet,a convolutional neural network with dilation factor,for training.The data are computed by softmax function after passing through the gate function,to determine the category of each sampling point.After training,the trained network is used to generate the data,and then the decoding operation is carried out to get the ECG signals.Secondly,this paper introduces GAN-based method combined with short-time Fourier transform.This method uses short-time Fourier transform to get the corresponding time-frequency spectrum of each data.Then,the time-frequency spectrum is input into a GAN network composed of convolution operation and transposed convolution operation for training.After training,ECG data are reconstructed by applying Griffin Lim phase reconstruction algorithm to the timefrequency spectrum learned by GAN.Thirdly,this paper introduces GAN-based method combined with stationary wavelet transform.This approach performs stationary wavelet transform on the data to obtain eight wavelet coefficients vectors including four approximate coefficients vectors and four detail coefficients vectors.Then the eight vectors are input into eight GAN networks composed of full-connection layers for training at the same time.After training,the corresponding wavelet coefficients vectors generated by the eight GANs are transformed into inverse stationary wavelet transform to reconstruct the ECG signals.Finally,this paper also proposes a performance evaluation algorithm for the above three ECG signal generation algorithms.Discrete wavelet transform is used to extract signal features,and then feature vectors are input into a support vector machine for classification.GAN-train scores and GAN-test scores are calculated to compare the performance of the ECG signal generation algorithms.Compared with traditional methods,the three algorithms proposed by us based on deep learning technology have greatly improved the quality,diversity,simplicity and extensibility of ECG signal generation.From the existing experimental results,the data generated by these three algorithms can be used in ECG signal processing algorithms and disease diagnosis algorithms based on ECG signals,so as to improve the performance of these algorithms in clinical environment.