Research On Few-shot ECG Signal Classification Based On MAML And Attention Mechanism

Purpose: The traditional ECG signal classification model has achieved very good results by using deep learning with the support of big data.In the experiments of various public data sets,it can very accurately classify common abnormal ECG signals.However,in reality,many cardiovascular diseases are very rare,and the amount of labeled ECG signal data that can be obtained for such diseases is very limited.In the field of automatic recognition of ECG signals,the existing models require a large amount of data.The ECG signal of this part of rare cardiovascular diseases is very poor in the traditional deep learning classification network and even cannot be automatically classified using the deep learning classification network.In order to solve this problem,this study combines the small sample learning method with the deep learning network and proposes the MAML/MSConv+ECA model,which is used in the classification of ECG signals to realize the classification of ECG signals in the case of small samples.Methods and Datasets: This study uses the model-independent meta-learning method MAML algorithm as the main framework,and uses the one-dimensional full convolutional network as the benchmark,and combines multi-scale convolution and efficient channel attention module(ECA-Net)to construct MAML/MSConv +ECA model realizes the classification of ECG signals in the case of small samples,and gets rid of the limitation that ECG signal classification must rely on big data.The data set used in this study comes from the world-recognized ECG database PTB＿XL database.The ECG data of 71 categories in the PTB＿XL data set are preprocessed and screened to obtain the final 26 categories,and 21 categories are selected.As the training set,the remaining 5 categories are used as the test set.Results: The accuracy rates of the MAML/MSConv+ECA model proposed in this study reached 0.80,0.87,0.55 and 0.68 respectively in the experiment of 2-way1-shot,2-way 5-shot,5-way 1-shot and 5-way 5-shot.In the comparison of MAML/MSConv+ECA model with MAML/Conv model,MAML/MSConv model and MAML/Conv+ECA model under the same MAML framework,the accuracy rate of MAML/MSConv+ECA model is all above 3.5%;in Compared with the Prototypical Network and the comparison Matching Net in the Few-shot learning algorithm,the accuracy rate is increased by more than 2.5%;in the machine learning algorithm with the KNN algorithm,random forest and SVM algorithm,the accuracy rate is increased by more than 20%;in the comparison with the CNN network in deep learning,the 2-way 5-shot experiment is used.In the case of the same accuracy rate,MAML/MSConv+ECA The data size required by the model is only 10.8% of the CNN network.Conclusion: This study innovatively proposes the MAML/MSConv+ECA model based on MAML and attention mechanism.Experiments and comparisons show that this model is better than other Few-shot learning algorithms,machine learning algorithms and deep learning algorithms in ECG signal data volume.The MAML/MSConv+ECA model solves the classification problem of ECG signals in the case of a small amount of data,so that some rare diseases can also be classified by ECG through the network model of this study.