Research On Attention Based Deep Learning Network For Automated Arrhythmia Classification

In recent years,with ageing of population,the problem of cardiovascular disease has become increasingly serious.As the most common persistent arrhythmia,atrial fibrillation can cause a series of complications such as stroke and systemic thromboembolism,resulting a great threat to people’s health and life.In order to provide accurate and reliable diagnostic information and reduce the morbidity and mortality of patients,the automated detection of atrial fibrillation has important clinical and social significance.With the development of artificial intelligence,deep learning models have been widely used in the automated analysis of ECG signals.However,the existing automatic detection algorithms of atrial fibrillation cannot provide the interpretability for classification results,and the classification accuracy of the model is still need to be further improved.Therefore,on the basis of deep learning model for automated arrhythmia classification,this paper attempts to optimize it by introducing different attention mechanism.In addition,Physio Net 2017 challenge data set are used to train and verify the proposed methods respectively.This paper studies and analyzes the influence of different forms of attention mechanism on the classification performance of atrial fibrillation.The main research contents of the paper are as follows:(1)The hybrid attention based deep learning network(HADLN)method is proposed to implement arrhythmia classification.HADLN not only has the interpretability of the model itself,but also solves the problem of ignoring context correlation and gradient dispersion in traditional deep convolutional neural network models,and further improves the classification accuracy of the model.In this model,the residual network(Res Net)and the bidirectional long short-term memory network(Bi-LSTM)are used to extract multiple features of the original ECG signals.The attention mechanism is used to fuse different extracted features and extract high-value information from them.For single-lead long-sequence ECG signals,the experimental results show that,compared to the traditional deep convolutional neural network,the proposed HADLN method achieves better AF classification performance.(2)A deep learning model with multi-head attention mechanism is proposed implement arrhythmia classification.In this paper,the deep residual network is used to extract the local features of the ECG signal firstly,and then the bidirectional long and short-term memory network is proposed to extract the global features on this basis,and finally the multi-head attention mechanism layer is used to extract the key features,and multiple modules are connected through the method of assembly and play the role of each module.The focus of multi attention mechanism is to spread the weight learning to different subspaces,so as to obtain different position features at different levels,which can effectively extract the key features,but its performance is affected by the number of subspaces.For single-lead long-sequence ECG signals,the experimental results show that,the proposed multi-head attention mechanism based deep learning method reaches the optimal value in 8 subspaces,which has a great improvement in arrhythmia classification.The two attention mechanism based deep learning models are proposed for arrhythmia classification in this paper.And the unified Physio Net 2017 challenge data was used to train and validate the proposed deep learning model.The proposed deep learning models can achieve good performances in terms of accuracy and has a better self-explanatory.