Research On Multi-lead And Multi-label ECG Classification Based On Deep Neural Network

Electrocardiogram(ECG)is an important tool for the diagnosis of cardiovascular diseases.The number of electrocardiograms received every year around the world is huge.If all of them require doctors to diagnose and screen them,it is time-consuming and laborious.Therefore,the computer-aided diagnosis comes into being.However,there are many types of ECG,and how to accurately use artificial intelligence methods for ECG-assisted diagnosis is still a research hotspot.In recent years,ECG-based research has mostly focused on the single-label problem,which means that a record corresponds to a label.However,in actual clinical applications,an ECG record may contain multiple diseases at the same time,and traditional computer-aided diagnosis only gives the prediction result of a single ECG,to a certain extent,ignoring the hidden dangers of the heart.Therefore,it is very important to study the multi-label ECG classification.The main research content of the paper is summarized as follows:(1)A multi-scale residual network(Multi-scale ResNet,MResNet)model is constructed.Using the residual network to extract deep features,and simultaneously perform multi-scale feature fusion,can realize the automatic classification of multilabel electrocardiograms.The proposed MResNet model is verified on the Chinese cardiovascular disease database.In order to determine the optimal choice of scale,experiments were performed with different convolution kernel sizes,which proved that MResNet reached the highest classification accuracy of the model when the two scales were 3 and 7,respectively.And compared with other classical neural networks.The results show that MResNet performs best.(2)A multi-scale ResNet model based on the channel spatial attention mechanism(Channel spatial attention with multi-scale ResNet,CSA-MResNet)is constructed.The channel spatial attention mechanism is introduced on the basis of MResNet,so that the model assigns more weight to more important ECG data fragments,so as to pay more attention to abnormal ECG data fragments.Finally,the proposed CSA-MResNet model is trained and tested on the Chinese cardiovascular disease database,and the model generalization verification is performed on the Hefei High-tech Cup ECG abnormal event detection data set.The experimental results show that when multi-label classification is performed on 9 ECGs,the CSA-MResNet model has the best classification effect compared with the benchmark model,and has a certain generalization performance.(3)Based on the CSA-MResNet model proposed in this paper,it is applied to the heart event real-time early warning cloud platform for system testing to prove its effectiveness in real-time ECG diagnosis.