Research On Arrhythmia Intelligent Diagnosis From Multi-lead And Multi-beat Perspectives

Cardiovascular disease is one of the most severe diseases that threaten human life and health.Effective detection of cardiovascular diseases is an urgent need of human beings.Electrocardiogram(ECG)is an effective means to detect cardiovascular diseases,and its acquisition method is simple and non-invasive.It is widely used in the diagnosis of various heart diseases.In recent years,many methods for detecting arrhythmias using 12-lead ECG have been proposed.In particular,deep learning methods,which have been widely used in recent years,still face numerous challenges.At present,the ECG diagnosis methods based on deep learning lack the usage of the physiological meaning of ECG signals.Based on the deep learning methods,combined with the physiological meaning of ECG signals,this paper proposes two models for ECG intelligent diagnosis from the perspective of multi-lead and multi-beat of ECG signals.The main contents are as follows:(1)The 12 leads of the ECG signal contain unique features and information.In order to make full use of this characteristic of the ECG,this thesis proposes a multilead combined ECG signal learning method.In multi-lead ECG analysis,most of the existing research models connect multi-lead ECG signals into a matrix as input.In this case,multiple leads are coupled during training,which causes lead-specific features to be ignored.In this paper,a network structure for intelligent diagnosis of multi-lead arrhythmias is designed,which combines the specificity of leads and the integrity between leads to extract more representative features to improve the ability of 12-lead ECG learning model.The first step of the model is to preprocess the original ECG data,and the second step is to use the multi-lead combined ECG learning method for training.We verified it on the data of the first China ECG Intelligence Competition.The F1 score reached 0.872,and won the second prize of the first China ECG Intelligence Competition(rank 8/545).The experimental results show that the model provides an effective method for ECG diagnosis.(2)In view of the multi-beat of ECG signals,this paper designs a multi-beat feature extraction and beat-level feature fusion network(BLF-Net).According to the contribution of the heartbeat to the diagnosis result,weights are assigned on the heartbeat level,so as to obtain a more comprehensive feature representation of the ECG.Our method is divided into three steps:(a)beats segmentation;(b)beat-level feature extraction;(c)inter-beat feature fusion.We test our algorithm on the PTB-XL database with an average macro-AUC of 0.931,outperforming the current state-of-the-art models with significantly reduced parameter size.Furthermore,the rationale of the model is clarified to some extent through the visualization of the attention weight mechanism.The network provides an efficient network structure for automatic classification of arrhythmias,which may help cardiologists in the diagnosis of arrhythmias.