Research On Multi-label Classification Method Of ECG Signal Based On Deep Learning

ECG,a physiological signal that indicates the status of the heart,is usually collected by surface electrodes and recorded by amplification.The signal is analysed and the abnormal bands are labelled(i.e.tagged),on the basis of which the doctor can detect and diagnose the corresponding heart disease.Various computer-aided diagnostic methods based on ECG signals are important in order to relieve physicians of the heavy burden of evaluating ECG signals.Single-label classification of ECG signals has been well studied and has produced good results.However,in practice,not only a single type of abnormality exists in a segment of the ECG signal,but also multiple abnormalities at the same time.In clinical diagnosis,it is necessary to label all the abnormal ECG signals present at the same time,which means multi-label classification.Most of the existing multi-label classification methods for ECG signals are based on independent modelling of individual labels without considering association information between labels,resulting in unsatisfactory classification performance for a part of the labels;although some studies have used feature-based selection to introduce relationships between labels,they are mostly based on machine learning,and the existing methods are mostly studied for small or relatively homogenous datasets.To this end,this paper explores issues related to multi-label classification of ECG signals on three datasets from a widely sourced dataset,PhysioNet/Computing In Cardiology Challenge 2020(Cinc2020),by introducing association relationships between labels based on a deep learning approach.The main research components are as follows:(1)Firstly,the experimental study of multi-label classification was carried out using three neural networks(ResNet,GoogLeNet,Transformer),assuming that the labels were independent of each other and without considering the association relationship between the labels.For labels with sufficient samples in the dataset,the macro F1 scores all managed to exceed 80%,while for labels with smaller sample sizes,the macro F1 scores were lower,not exceeding 50%.(2)Referring to the above experiments,a graph convolutional network for label embedding is proposed with consideration of relationship between labels.Based on both word embeddings and a priori probabilities of label co-occurrence relationships modelling relationship between labels,feature representations containing association relationships between labels are learned by Graph Convolutional Network(GCN)and fused to ECG signal feature learning to further focus on the information between labels.Validation of the method was carried out on the Cinc2020 dataset,where some improvement in macro F1 scores was obtained,especially for labels with small sample sizes.(3)Furthermore,a dynamic graph convolutional network based on ECG signal features is proposed in order to mine relationships between labels based on the ECG signal itself.The feature mapping of the ECG signal is obtained from ResNet,the feature representation of the labels is obtained from the class generation module,a shared trainable correlation matrix is used to model the relationship between the labels from the full training sample,a correlation matrix is further constructed for each ECG record,and the feature representation containing the correlation relationship between the labels is learned by GCN.Validation of the method on the Cinc2020 dataset through comparative experiments and module performance experiments further improves the performance of multi-label classification.Alternatively,by replacing the ResNet part of the ECG signal feature extraction with two other network structures,the method also improves the results of multi-label classification.To sum up,by introducing multi-label relationships using a deep learning approach,this paper improves the performance of multi-label classification in ECG signals,especially for labels with a small number of samples.The research method has only been validated on the Cinc2020 dataset and can be subsequently extended to more datasets for application and further optimisation of the model parameters and structure.