Research On Classification And Recognition System Of ECG Signal Based On Deep Learning

Cardiovascular diseases listed by WHO as leading cause of death.As a non-invasive diagnostic tool,electrocardiogram has been widely used in the diagnosis of cardiovascular disease in clinical practice.The classification and recognition of ECG signals is an important research topic in the fusion process of medicine and computer technology.The key is to extract the effective characteristics of ECG signals more accurately.The traditional ECG classification methods first perform feature extraction and selection,and then perform classification.However,these methods rely heavily on manual labor and cannot fully dig deep features hidden inside a large number of ECG signals.Nowadays,deep learning is booming.Deep learning has the characteristic advantages of learning features automatically without the need for an explicit feature extraction process,it opens up a new path for the automatic classification and recognition of ECG signals.Therefore,this paper applies deep learning technology to the classification and recognition of ECG signals in order to improve the accuracy of classification and recognition of ECG signals.The main research contents are as follows:Firstly,In view of the denoising for ECG signals,this paper proposes a wavelet denoising method based on adaptive threshold,which can dynamically adjust the threshold of different decomposition scales.The basic steps are: First,the sym8 wavelet function is used to decompose ECG signals on 8 scales.Then,set the approximate coefficients on the 8th scale and the detail coefficients on the 1st scale to zero directly,and use the adaptive threshold which is proposed in this paper and soft threshold function to process the other wavelet coefficients.Finally,by reconstructing the signals,the denoised ECG signals are obtained.Secondly,on the base of deep learning technology,this paper builds two ECG signal classification models: DBN model and CNN model.In the DBN model,in view of the problems of local optimization and long training time that are prone to occur during the training phase of the traditional DBN model,the idea of adding momentum and using batch training to update the parameters in the parameter update process can effectively improve the traning performance.In the CNN model,in order to achieve rapid convergence,reduce overfitting,and improve the generalization ability of the model,some optimization techniques such as Xavier weight initialization,Relu activation function,BN,and Dropout are used.Two models are used to realize the automatic classification of the four arrhythmia beats based on the AAMI standard.The four types are normal heartbeat(N),supraventricular ectopic heartbeat(S),ventricular ectopic heartbeat(V),fusion heartbeat(F).Thirdly,training and testing the DBN model and CNN model respectively based on the MIT-BIH arrhythmia database.For the DBN model,the classification performance is evaluated and analyzed from three aspects: the setting of the number of hidden layers,thecomparison of classification results before and after parameter tuning,and comparison of classification results before and after data synthesis.For the CNN model,the comparison experiments in three aspects which including the parameters of the convolution layer,the setting of the learning rate,and the number of iterations are performed to obtain the optimal parameters of the CNN model,and the optimal CNN training model is obtained and used for testing.Finally,The DBN model and CNN model obtain overall accuracy rates of 98.46%and 98.10% individually.By comparing with other research methods under the intra-patient and inter-patient schemes,the results show that the proposed methods of ECG signals classification which are based of DBN and CNN have high classification performance.