Research And Application Of ECG Classification Based On Deep Learning

The ECG is easy to be collected,which doesn’t harm the human body.It has been widely used in detecting cardiovascular diseases in the clinic.However,since the ECG is a lowamplitude,low-frequency,and susceptible physiological signal,it is very difficult to filter out noise,and the variability and personalization of the ECG leads to difficulties in the clinical application of the ECG automatic classification algorithm.Therefore,the classification algorithm of ECG is still an area that needs to be studied.This article will study the ECG classification in following three parts:1)For the classification problem of single-lead ECG signals,we propose an end-to-end model combining multi-branch convolutional neural networks and residual networks to extract the characteristics of single-lead ECG.Thereby achieving normal and abnormal classification.In the experiment,we compared the four classical convolutional neural network models of AlexNet,VGGNet,GoogLeNet and ResNet in the classification of ECG signals.The results show that our proposed model has better performance in ECG signal classification.2)For the classification problem of multi-lead synchronous ECGs,we use an end-to-end model composed of multi-branch convolution and residual networks to capture the ECG signal characteristics of different leads.Then,the synchronism and independence of the multi-lead ECG signal are considered,and the extracted features are fused.In this paper,we propose three feature fusion methods,use different convolution and fully connected layers to achieve feature fusion,and map ECG signals to different kinds.In the experiment,the normal and abnormal ECG judgment can reach an average Accuracy of 87.04% and a sensitivity of 89.93%.3)Considering ECG signals are susceptible to interference,filtering algorithms are applied to improve the classification Accuracy of ECG.In the experiment,we used zero-phase IIR filter,band-stop filter and low-pass filter to filter the baseline drift interference,power frequency interference and myoelectric interference of ECG.The experimental results show that using the zero phase IIR filter to filter the baseline drift of the ECG can improve the classification Accuracy.Among the current classification and recognition algorithms for ECG,most of the databases used in the experiments are similar to the MIT-BIH ECG database.The data of these databases are from a small number of individuals,which leads to the obvious reduction of accuracy when the existing classification algorithm is applied in the clinic.This paper uses the Chinese Cardiovascular Disease Database,which has more than 190,000 clinical ECG data from different individuals.It makes the classification algorithm have good generalization performance.