Research On Key Technology Of Arrhythmia Automatic Classification Based On Convolutional Neural Networks

As an important application of computers in clinical practice,the benefit of automatic cardiac arrhythmia classification lies in the rational allocation of resources available for healthcare as well as the effective prevention of cardiovascular diseases.As people pay more and more attention to health management,both subjects and doctors put forward a higher requirement on the performance of automatic cardiac arrhythmia classification system for timely and accurately obtaining the information that reflects the functional state of the heart.The problems emerged in the current developments of telemedicine,digital health and e-home health,mainly involve:how to improve the capability in detecting arrhythmia;how to produce intelligent and reliable diagnosis results using computers? However,the conventional automatic cardiac arrhythmia classification methods cannot deal with the aforesaid problems,as the automatic cardiac arrhythmia classification is a comprehensive research objective involving multidisciplinary theory of modern sciences and plenty of key technologies.With the development of deep learning technologies in terms of image classification,it provides a new approach for a satisfactory performance of automatic cardiac arrhythmia classification to meet the medical market needs.To overcome the challenges mentioned above,this thesis proposes an effective and efficient automatic cardiac arrhythmia classification method and optimization strategies where a Convolutional Neural Network(CNN)as the most representative of the deep neural networks is employed to learn the MIT-BIH arrhythmia database.The process of the proposed automatic cardiac arrhythmia classification mainly encompasses the Electrocardiogram(ECG)signal processing,ECG beat classification and key techniques in the optimization strategies.The main contents and contributions of this study are presented as follows:1.An automatic arrhythmia classification method based on one dimensional convolutional neural network model(named ECG-SPP-net)is proposed.To solve the problem that the existing CNN models(e.g.VGGNet)cannot identify the images with variant different scales,this research firstly constructs a new ECG-SPP-net using Spatial Pyramid Pooling(SPP)method for different lengths of heartbeats.Meanwhile,the proposed ECG-SPP-net can avoid that the problem of limited weight sharing brought by the conventional SPP-net.In view of the ECG signal processing,median filter and wavelet analysis are adopted to remove noise within less computational burden.By carrying out experiments for the classification of heartbeat signals into six categories,the effectiveness of the proposed ECG-SPP-net is validated.2.An automatic classification method(named 2D-ECG-CNN)based on a one-hot encoding technique and a two-dimensional CNN model(named2D-CNN)is proposed.To handle the ECG signal processing,one-hot encoding is used to convert the ECG signals from one-dimension time series to binary images in the sense of constructing 2-D CNN for the ECG beat classification.Moreover,the morphology and rhythm of heartbeats are fused into a two-dimensional information vector.To avoid the limitation and complexity in terms of manual tuning learning rate,an adaptive learning rate called AdaDelta is utilized in the 2-D CNN.Comparing with the 1-D CNN,experimental results indicate that the proposed2D-ECG-CNN can improve the average classification accuracy to 98.8%.3.Two optimization strategies are designed to improve the classification performance of the proposed 2D-ECG-CNN.On the one hand,considering the convolution layer can filter the signal to some extent,an uncomplicated median filter replaces the original denoising method to preprocess the ECG signals for reduction of the computational complexity.On the other hand,Biased Dropout method is employed as a network compression technique for 2D-CNN to increase learning speed and recognition capability.After using these two optimizationstrategies,experimental results show that the classification accuracy can be improved to 99.1%.Besides,the results of the comparative experiments verify the universality and generalization of the proposed methods to different arrhythmia databases.Other than that,the proposed model was tested on base of the standard ANSI/AAMI EC57: 2012 by the Association for the Advancement of Medical Instrumentation to increase the reliability of the experimental results with uniform arrhythmia types and performance indicators for classification evaluation.