Application Research Of Convolutional Neural Network In The Classification Of Multi-lead Electrocardiogram

Cardiovascular disease is a significant threat to human health and is the leading cause of death for non-communicable diseases.Because of the inexpensive,effective and noninvasive properties of electrocardiogram,it is widely used in clinical medical testing and is critical for diagnoses of various cardiovascular diseases such as arrhythmia,coronary heart disease and myocardial infarction.Automatic electrocardiogram interpretation which known as electrocardiogram classification is playing an increasingly important role in clinical electrocardiogram diagnosis.It can enhance the accuracy and efficiency of electrocardiogram diagnosis.Traditional electrocardiogram classification methods rely on domain knowledge of experts and it require complex preprocessing of electrocardiogram signals and then manually build features.Those methods are difficult to implement and have poor generalization.Therefore,they are hard to achieve the standard of clinical diagnosis.Electrocardiogram classification based on deep learning can utilize the property of deep models learning representative features from raw data.Both classification accuracy and generalization of deep learning methods surpass traditional methods,which make them become hot spots of this field in recent years.In this thesis,an end-to-end multi-lead electrocardiogram classification method based on CNN is proposed.Based on the similarities between different leads in multi-lead electrocardiogram,an adaptive network structure that can learn shared features among multiple leads is proposed to strengthen the representative ability of multi-lead electrocardiogram for this model.In addition,this research explore the influence of the global pooling on electrocardiogram classification.This study applied a large-scale clinical electrocardiogram dataset to verify the availability of the proposed method.The experimental results indicate that the method can precisely recognize various arrhythmias from multi-lead electrocardiograms.Since CNN focus on local morphological features of electrocardiogram signal,it is difficult to process the time domain features of electrocardiogram signal or clinical features of patients.However,those features play an important role in electrocardiogram diagnosis.This thesis further proposes an electrocardiogram classification method that integrates deep features of CNN,time-domain features of electrocardiogram signals and clinical features of patients.The experimental results indicate that the classification accuracy of this method surpass the previous method which relying on a single end-to-end CNN model.