Classification And Application Of Arrhythmias Based On Deep Learning

Electrocardiogram is an important basis for doctors to diagnose and treat heart diseases,and the interpretation of electrocardiogram depends on doctors’ subjective opinions and experience.With the increase of patients,the number of electrocardiogram to be diagnosed by doctors every day is very large,which is very easy to cause misdiagnosis of the disease.Moreover,the lack of real-time artificial recognition of electrocardiogram may delay the best time for treatment.Therefore,the application automatic analysis and recognition of ecg signals based on deep learning in the scientific research of heart disease can greatly reduce the workload of doctors and improve the accuracy of diagnosis at the same time,which will have great value for the diagnosis and treatment of heart disease in clinical application.In this article,MIT-BIH database of arrhythmia is used for signal preprocessing and automatic classification.Wavelet transform is used to decompose the ecg signal,and selecting appropriate threshold and threshold function for denoising the three main kinds of noise(baseline drift,emg interference and power frequency interference)in the ecg signal.After denoising,the ecg signals are divided into a heart beat,which is composed of 256 points that is selected forward and backward with R wave as the axis.Then Wigner time-frequency transformation is carried out on the heart beat sample to transform the one-dimensional signal into a two-dimensional image.Finally,two kinds of convolutional neural network models,LeNet-5 and VGGNet-16,and the improved models based on them respectively are used for image recognition and classification.On the basis of the classical LeNet-5 model,the improvement is that adding a convolutional layer and a pooling layer,the number of convolution kernels on the convolutional layer is increased to 32,64 as well as 128 respectively,and selecting the appropriate size of convolution kernel.In order to extract more abstract and distinguishing features from the data,on the basis of VGGNet-16,the improvement is that adding a convolution layer on the convolution layer module 5,and the number of convolution kernels on the convolution layer module 5 increases to 700 and then comparing the experiment results.This article selects a total of 16246 heart beat time-frequency image of 7 classes as the data set,the original LeNet-5 models of accuracy is 63.75%,the improved LeNet-5 model classification accuracy is 92.48%,VGGNet-16 network model for data classification accuracy is 97.43%,the recognition is higher than LeNet-5 model over five percentage,the improved VGGNet-16 model 1 of classification accuracy is 97.9%,improved VGGNet-16 model 2 of classification accuracy is 98.9%.Finally adding to the test set data with different intensity of gaussian white noise,the result shows that the lower SNR signal,the lower the accuracy,when the signal-to-noise ratio is between 25 and 17 dB,the improved LeNet-5 has little change about model accuracy,when the signal-to-noise ratio is less than 17 dB,the accuracy began falling fast,with accuracy of 62.79% in the end.On the other time,the accuracy of the VGGNet-16 model and the improved VGGNet-16 model decreased very slowly between 25 and 13 dB,when the signal to noise ratio is less than 13 dB,VGGNet-16 model and the improved VGGNet-16 model of accuracy fall fast,eventually at 70.95% and 75.26% respectively.