ECG Signal Feature Extraction,Sparse Representation And Abnormal Heartbeat Classification

Cardiovascular disease?CVD?has become the leading cause of death among non-infectious diseases over the world,it has greater number of deaths than any other diseases.From the 21^st century,the global population,especially the aging population,continues to increase,as a result,the cost of cardiovascular treatment rises rapidly and there is a shortage of medical resources along with unbalanced distribution of medical level.All the disadvantages exert a direct effect increasing the cardiovascular morbidity.To realize early detection and effective prevention of CVD,daily electrocardiogram?ECG?monitoring has provided a feasible technology as the development of body sensor networks and signal processing techniques.Based on the body sensor networks,new type of mobile cardiovascular monitoring system is able to capturing rare instantaneous heartbeat.After analyzing and processing these beats,the results will be sent to patients,doctors or families,realizing remote cardiovascular monitoring,medical and health care.However,there is still a far way to implement daily ECG monitoring in clinical use,several critical issues are still challenging such as dry electrode,signal sensing,typical wave location,feature extraction and heart beat classification.In this study,all the mentioned challenges were investigated and discussed in the view of sensing,characterization and recognition,including:synchronous signal collection,signal quality evaluation and comparison,R-peak detection,low-dimensional feature extraction,heart beat classification.This study mainly focused on the following problems:?1?Synchronous signals were collected,the signal quality differences were compared and quantified to demonstrate the clinical diagnostic performance of the AgNW electrode;?2?An adaptive and time-efficient ECG R-peak detection algorithm was developed,providing the accurate R position for the following feature extraction procedure;?3?Low-dimensional beat features were extracted based on wavelet multi-resolution analysis and sparse representation,providing input features for the classifiers;?4?A beat classification method combining low-dimensional features,10-fold voting scheme and support vector machine was proposed to test the property of R-peak detection and feature extraction method.The highlights of this study included:?1?A synchronous signal collection scheme was proposed by using AgNW and Ag/AgCl electrodes.Signal quality disparity was systematically evaluated and quantified by signal quality indices and heart rate variability.The results proved that the AgNW electrode was a potential substitute for the Ag/AgCl electrode,meanwhile,the method has provided diverse selection of parameters for signal quality evaluation;?2?A frequency band selection method was proposed based on wavelet multi-resolution analysis,the method could be used not only for signal denoising and enhancement,but also for feature extraction.A sparse representation-based feature extraction method was also proposed for heartbeat classification;?3?An adaptive and time-efficient ECG R-peak detection algorithm was proposed,the R-peak detection accuracy could reach up to 99%,and saving 30%of computation time;?4?The amplitude and time interval thresholds could change adaptively according to the signal amplitude and heart rate,and the feasibility of the threshold operation was validated.The amplitude threshold coefficient could be any value from[0.2,0.3],and the time interval threshold coefficient could be any value from[0.42,0.48];?5?A dimensionality adjustable scheme was proposed for feature extraction,the feature dimension was tunable according to the signal characteristics and classification requirements.The results has indicated that for the classification using wavelet features or sparse features,the classification rates remain unchanged and show strong consistency;?6?A 10-fold voting based heart beat classification method was developed to recognize 4 types of heart beat:premature atrial contraction,premature ventricular contraction,normal,others.Using the individual-based scheme,the proposed method could realize an overall recognition rate of more than 70%.This research has further enriched the study of ECG monitoring technology and related methods and techniques.It is hoped that this study could provide theoretical support and method reference for ECG sensing,continuous monitoring and early prevention of cardiovascular diseases,and intelligent diagnostic evaluation.