Research On ECG Abnormality Diagnosis Based On Deep Transfer Learning

The incidence rate of cardiovascular diseases is increasing year by year.The range of cardiovascular diseases is very wide,including arrhythmia,myocardial infarction and so on.If we want to distinguish these diseases accurately,we need to use ECG equipment to obtain the patient’s ECG.In general,the discrimination of ECG needs medical experts to make artificial discrimination,which is laborious and easy to make mistakes.Therefore,it is very important to diagnose ECG by computer aided system.Deep learning has the unique ability to learn features from original data,but there is data dependence.Transfer learning solves this problem.It allows the migration of existing knowledge to solve problems in the target domain of a small number of datasets.In this paper,deep learning and transfer learning are combined to study the diagnosis of ECG.The main work of this paper is as follows:(1)Aiming at the problem that ECG data sets are limited and difficult to obtain,this paper proposes to apply transfer learning to convolutional neural network.In order to solve the problem that convolutional neural network needs fixed image size,spatial pyramid pooling is used to replace the pooling layer between the last convolutional layer and the full connection layer.Experiments on the data set of Physio Net challenge in 2017 show that the accuracy of the proposed method is better than other methods.(2)In view of the timing of ECG data,a gated recurrent neural network is added to convolution neural network,which can combine the spatial and temporal information of ECG signal,and improve the accuracy of ECG classification.In order to make the model pay more attention to the important aspects of ECG signal,attention mechanism module is introduced.Experiments on MIT-BIH arrhythmia database show that the proposed method is superior to other methods.(3)Aiming at the research of myocardial infarction diagnosis,this paper proposes to transfer the arrhythmia model of ECG signal to the research of myocardial infarction diagnosis,and then fine tune the network model.The proposed method can transfer the knowledge learned from arrhythmia to the target task of myocardial infarction diagnosis research.Experiments on PTB database show that the knowledge learned from the task can be successfully transferred to the task of myocardial infarction prediction using ECG signals.The experimental results show that the method based on deep transfer learning is a feasible scheme for ECG abnormal diagnosis.This paper mainly uses the network-based deep transfer learning method.When building the network model of the target domain,it reuses some of the pre trained networks in the source domain,including their network structure and parameters,and then uses the target data set to fine tune.On the one hand,it solves the problem of data dependence of deep learning model;on the other hand,it improves the work efficiency by using the pre trained network.