Pruning Method Of ECG Signal Classification Model Based On Attention Mechanism

Objective: With the increasing precision and accuracy of ECG classification models,the application of deep learning models in ECG classification has become more and more complex.Nowadays,the demand for deploying models on terminals such as portable ECG detection devices has become very common.Due to the limited resources on such devices,the problems of large storage and slow running speed of classification models have become prominent,so it is necessary to simplify the deep neural network model.Pruning the model will affect the performance of the ECG classification model.In order to ensure that the classification performance of the ECG classification model after compression is not negatively affected and improved,we propose a pruning method based on the attention mechanism.Methods: In this study,pruning algorithm PEAM based the attention mechanism for ECG classification model is proposed.This method uses the attention module to explore the correlation between the filters,so as to prune the filters with less correlation to obtain a more concise compression model with less global loss.MIT-BIH database is used to train and test the model before and after pruning on ECG classification models based on VGG11,ResNet18 and ResNet34.And CMUH database is used to verify the generalization of PEAM on ResNet18.In order to further explore the influence of the pruning algorithm on the feature basis of the model classification decision,this study also compares the results of Score-CAM class activation map before and after pruning.Results: Compared with the other five pruning algorithms,the PEAM pruning algorithm proposed in this paper can obtain the best classification performance under great compression of the model.The accuracies of PEAM after pruning of VGG11,ResNet18 and ResNet34 on the MIT-BIH database are 98.91%,98.75% and 98.43%respectively,and the F1-scores are 99.05%,98.89% and 98.71%.Among them,PEAM has higher accuracy and F1-score after pruning of ResNet18 and ResNet34 than the original model.The accuracy of ResNet18 after pruning on the CMUH database is 97.63%,and the F1 value is 97.89%,which are also higher than the original model.The ECG signal feature annotation results of PEAM after Score-CAM activation mapping are closer to the annotation results of the original model than the other pruning methods.Conclusion: Compared with other pruning algorithms,PEAM,the pruning algorithm of ECG classification model based on attention mechanism proposed in this paper,can make the pruned model obtain better classification performance,and the visual results show that PEAM pruning algorithm has less impact on the decision-making basis of the model.This is of great significance in promoting portable ECG detection devices and helping doctors make clinical diagnosis of cardiovascular diseases.