Arrhythmia Automatic Diagnosis Based On Deep Learning

The improvement of living standards and lifestyle have seriously threaten human’s health.Cardiovascular diseases such as arrhythmia are plaguing more and more people.Finding and preventing the occurrence of heart disease can effectively reduce the mortality rate.ECG signals can effectively reflect the state of human heart health.The application of automatic detection and classification of ECG signals is significance in the prevention and treatment of heart disease.The paper summarizes the significance of the research topic and status at home and abroad.It also gives a basic explanation of ECG,ECG interference and arrhythmia.Secondly,ECG signal is weak in interference,utilizing good time-frequency domain and local analysis capabilities of wavelet transform to achieve noise processing;The traditional arrhythmia automatic diagnosis methods rely on the manual feature extraction;deep learning technology is applied to the diagnosis of arrhythmia,realized the feature automatic extraction and classification recognition as a whole.The innovative research of this article mainly includes four aspects:1.ECG signals are susceptible to noise interference.There are differences in frequency domain distribution between different types noise.The wavelet transform method is used to achieve noise filtering in ECG signals.Filter the low-frequency noise by zeroing the corresponding domain in the decomposition factor,and then reconstructing the scale coefficients of the processed wavelet;removing the high-frequency noise using the wavelet threshold denoising method and selecting the appropriate wavelet basis,decomposition scale,and use an improved soft and hard threshold compromise method to effectively remove noise.2.Extend the MIT-BIH ECG signal dataset in a data-enhanced manner.Perform R-wave localization and heartbeat segmentation for all modified single-lead Ⅱ in the dataset,and normalize the experimental dataset ECG_Ⅱ;Two kinds of effective solutions are proposed for unbalanced data of heart beat categories.3.The ECG is a timing signal,and the Recurrent neural networks have significant sequence data modeling capabilities.Through selecting performance optimization parameters,the paper proposes and constructs LSTM ECG network model suitable for ECG signal recognition.Comparative the different network structures,and integrate the average time-consuming and accuracy indicators to verify the validity of the LSTM_ECG model,finally analyze the model structure parameters and recognition effects.4.The timing of the ECG signal constitutes an electrocardiographic waveform that contains spatial information.Convolutional neural network focuses on spatial mapping and is suitable for the processing of ECG data.The paper analyzes signal characteristics and model performance parameters,builds CNN_ECG network model to achieve heartbeat recognition.Compare the different network structures to verify the feasibility of the CNN_ECG model,and give the model structure parameters and recognition effects.Finally,comparing the LSTM_ECG and CNN_ECG models proposed by this project and existing research results of others.