Research On ECG Data Augmentation Method Based On Generative Adversarial Network

Computer-aided automatic classification of ECG signals is an important method for effective diagnosis and prevention of cardiovascular diseases.At present,scholars at home and abroad have carried out a large number of studies on ECG signal classification using public ECG databases,and have achieved a series of effective research results.However,these public ECG databases generally have the problem of small amount of data and serious imbalance of data categories,which seriously affects the improvement of the accuracy of ECG classification and hinders the development of ECG classification research.In response to the above problems,this article takes ECG signals as the research object,and processes the ECG signals based on the theory of generative adversarial networks and machine learning methods.This article focuses on ECG signal preprocessing,ECG data Augmentation,and augmened data classification performance verification.The main research contents are as follows:(1)ECG signal preprocessing.First,the discrete wavelet transform method is used to denoise the ECG signal in the MIT-BIH heart rate abnormality database.Then use the PT algorithm to locate the R wave of the ECG signal.The heart beat segmentation is performed according to the positioning result.Finally,according to the AAMI standard and various heartbeat data characteristics,9 types of heartbeat signals are classified and extracted to provide data support for subsequent ECG data Augmentation and augmened data classification performance verification.(2)ECG data Augmentation.Three heartbeat data Augmentation methods based on generative adversarial network are proposed,namely ECG-GAN,ECG-CGAN,and ECG-LSTM-GAN.The ECG-GAN model sets the hidden layer of the generator network model to the Batch Norm1 d one-dimensional fully connected layer architecture,so that the input of each layer maintains the original distribution characteristics of the data and avoids the problem of gradient disappearance.The ECG-CGAN model provides additional information of class tags to the generator and discriminator,and guides the heartbeat data generation process in a supervised manner;The ECG-LSTM-GAN model combines the GAN model and the LSTM model to retain the time correlation of the sample data,thereby further improving the quality of the generated data.At the same time,after the heartbeat data is generated by the three types of GAN networks,it is denoised by the band stop filter and the DB8 wavelet double filter.Finally,the data Augmentation performance of the generated confrontation network model is judged by generating the heartbeat data pattern,the PRD index and the RMSE index.The experimental results show that the PRD value and RMSE value of the three models are relatively small.Among them,the PRD value and RMSE value of the ECG-LSTM-GAN model are the smallest.(3)Augmen the verification of data classification performance.In order to verify the effectiveness and robustness of the three ECG data Augmentation methods,this paper uses three commonly used traditional machine learning and two deep learning classifiers to evaluate the classification performance.In the evaluation,30% of the original heartbeat data was used as the test data,and the augmened heartbeat data and the original 70%heartbeat data were trained to compare their Top1 and Top5 performance indicators.The final test results show that it is compared with the original data.Using the heartbeat data generated by the ECG-GAN,ECG-CGAN,and ECG-LSTM-GAN ECG generation models proposed in this article,the ECG classification has improved the classification performance indicators.Among them,the advantages of the ECG-LSTM-GAN generative model are more obvious.This research provides new ideas and methods for the current automatic classification of ECG based on the open ECG database.