Research On 12-Lead ECG Signal Classification Method Based On Attention Mechanism

Objective: Cardiovascular disease has always been one of the most threatening diseases to human life.Thus,early screening of cardiovascular disease is extremely important.Traditional diagnosis requires manual screening by clinicians.At present,a large number of classification algorithms based on deep learning have achieved good results in the detection of 12-lead ECG signals,but it still needs further improvement from the perspective of clinical application.At the same time,whether these algorithms can gain the trust of clinicians and patients is crucial to the application of classification network,so the interpretability of ECG classification network has attracted more and more attention.This paper proposes a convolutional neural network(SCBAM)based on attention mechanism to detect 12-lead ECG signals,and uses the Score-CAM interpretability method to visually analyze the results of SCBAM network.Methods: The classification method proposed in this paper is composed of 16-layer CNN architecture and attention module(SCBAM).Based on the CBAM module,the Soft-Pool pooling method is introduced to improve the SCBAM module.Firstly,the ECG signal is sent into the CNN architecture to obtain the feature maps of multiple channels through feature extraction,and then these feature maps are fed to the SCBAM attention module.The SCBAM module uses multiple pooling structures and convolution structures to extract the importance of channel and spatial from the feature map information and it’s multiplied with the corresponding feature map as a mask.Finally,the feature map is used for classification.In order to determine the best application location of attention module and the most appropriate network depth,this paper uses the 50-fold cross-validation technology to carry out the best application location experiment of attention module and the best depth experiment of SCBAM network on the PTB-XL12-lead ECG database.Finally,we also use the five-fold cross-validation technology to carry out a five-category and nine-category comparative experiment of cardiovascular disease on the PTB-XL 12-lead ECG database and the CPSC2018 12-lead ECG database,respectively,using the SCBAM network and seven neural networks(CBAM,ECA,Senet,Resnet,34 Layer,Mobilent V2,NCBAM)to verify the effect of the network proposed in this paper.Results: In the application location experiment of attention module,the effect of applying attention module to CNN architecture is better;In the optimal depth experiment,the SCBAM network with 16 layers of convolution layer achieved the best effect;In the comparative test,the SCBAM network obtained the AUC value of 93.117 ± 1.069 and the maximum F1-Score of 87.860 ± 0.438 when performing five classifications in the PTB-XL 12-lead ECG database.The SCBAM network obtained the AUC value of93.014 ± 0.962 and the maximum F1-Score of 90.574 ± 0.791 when performing nine classifications in the CPS2018 12-lead ECG database.Conclusion: Compared with other excellent ECG classification networks,the SCBAM network has achieved better results in two open 12-lead ECG databases,and the Score-CAM visualization results show that the characteristics of SCBAM network used for classification are consistent with clinical ECG knowledge.The above results show that the SCBAM network can effectively complete the classification of 12-lead ECG signals and its classification results are reliable.