Research On Feature Extraction And Fusion Method Of Ecg Signal Based On Deep Learning

The incidence of cardiovascular disease is obviously rising now due to various reasons such as working pressure,lifestyle,and aging population,which seriously threatens human health.Therefore,it is necessary to effectively strengthen the prevention and diagnosis of cardiovascular disease.Since the electrocardiographic(ECG)signals can characterize the condition of the heart,it is an important indicator for diagnosing clinical cardiovascular diseases.However,the traditional methods to diagnose cardiovascular disease are heavily relay on subjective observation,thus resulting in the inaccurate results.Therefore,the computer-aided diagnosis technology is necessary for objectively diagnosing diseases.Computer-aided disease diagnosis can be seen as a classification task,which includes three processes: data preprocess,feature extraction and feature processing.Feature extraction is the most important step.Currently,many traditional methods that used to extract features of ECG signals have some shortcomings such as highly dependence on subjective experience,large calculations in the transform domain,and insufficient capabilities of generalization.As a result,feature extraction methods based on deep learning have been increasingly used for ECG signal feature extraction.In this thesis,two methods based on deep learning are studied to extract ECG image features and numerical features,respectively.Then the auto-encoder is used to achieve the fusion of image features and numerical features,which can make the classification of ECG signals accurate and efficient.First,this thesis proposes a multi-level convolutional neural network(ML-CNN)model to extract image features of ECG signals.This model introduces multi-level ideas to divide the original heartbeats data into three levels,and design three convolutional neural network for these data.Therefore,ML-CNN can extract the whole-level features,local sequence-level features and local band-level features of ECG signals.These features are input to the classification network,and then the result of classification are given.Finally,the gradient descent method is used to train the overall network.The ML-CNN model comprehensively considers the local characteristics,local sequence characteristics and overall trend characteristics,which reduces the loss of effective features and increases the receptive filed to improve the accuracy of classification.Second,this thesis converts the timing information of ECG signals into band sequence information and transfers it to the recurrent neural network(RNN),and proposes a feature extraction method of the band-RNN.This method is more suitable for extracting the numerical features of heartbeats.The experimental results illustrate that the classification results of the band-RNN are better than that of the 1-D CNN,and the classification results of the band-bidirectional RNN are no different from that of the convolution-RNN.Finally,based on the previous work,this thesis introduces auto-encoder and to propose a feature fusion method that can mix the image features and numerical features of ECG signals.In addition,this thesis also introduces sparse auto-encoder(SAE)and denoising auto-encoder(DAE)to analyze the effects of different auto-encoders on feature fusion.The experimental results prove that the classification results of the feature fusion method based on the auto-encoder are stronger than that of the individual features.The SAE is suitable for the feature fusion extracted by methods with great distinction and DAE gives the mixed features with high capabilities of generalization.