Research On Arrhythmia Classification Algorithm Based On Multi-branch Convolutional Neural Network

In contemporary society,there are smoking and drinking,irregular work and rest,lack of exercise,unreasonable diet structure,and work and life pressure.The double influence of many bad living habits and mental pressure leads to an increasing number of people suffering from cardiovascular diseases.Arrhythmia,as a typical disease of cardiovascular disease,has a high incidence and recurrence rate.It usually requires a cardiologist to observe the patient’s dynamic electrocardiogram for 48 hours.Not only is it time-consuming for doctors to diagnose arrhythmias,but many rural areas are faced with a shortage of medical resources,which makes diagnosis and treatment of arrhythmias difficult.Therefore,it is necessary to develop a computer-aided diagnostic system for arrhythmia.At present,many researchers have contributed a variety of algorithms for arrhythmia classification using the electrocardiogram,but some of these algorithms still need to be improved:First of all,most of the existing algorithms have low classification accuracy of noisy beats or lack experiments on noisy data sets,which reflects the poor noise robustness of their methods.The additional denoising steps will not only make the classification algorithm cumbersome,but also the unreasonable denoising algorithm will lose the detailed features of the heartbeat,and the improvement of the classification accuracy is limited.Secondly,the classification of unbalanced ECG data sets increases the difficulty of predictive modeling.Due to the lack of sufficient data for the category of abnormal heart rate,the classifier has a poor ability to depict sparse heartbeat samples,so the classification boundaries learned by the classifier tend to be in favor of the category of normal heart rate,leading to the deviation of classification boundaries and the reduction of classification performance.To solve the above problems,this thesis carries out the following research:(1)The arrhythmia classification algorithm based on multi-scale and multi-branch convolutional neural network and extreme learning machine(MBMS-CNN-ELM)is established.To solve the problem of insufficient extraction of correlation features before and after the heartbeat,firstly,the network applies multiple empty convolutions with different convolutional expansion rates to enlarge the receptive field and extract the correlation features before and after the heartbeat.Secondly,the network adopts a parallel structure of multiple convolution kernels of different sizes,which can simultaneously pay attention to the detailed features and global features of the beat and learn sufficient multi-scale beat information.Finally,aiming at the low classification accuracy caused by the poor generalization performance of the Softmax classifier,the extreme learning machine classifier is applied to classify the extracted features.The feature weight of the output layer of the extreme learning machine is obtained by solving the generalized inverse matrix theory.It has a strong generalization ability,fewer training parameters,and more ability to reach the global optimal than the feedforward neural network.The comparison experiments of different feature extraction and classifier algorithms show that the MBMS-CNN-ELM algorithm has better classification accuracy,and the overall accuracy rate is 99.3% in the MIT-BIH standard five classification experiment of the arrhythmia database AAMI.(2)A two-branch progressive learning strategy based residual network combined with the convolutional block attention mechanism(CBAM-ResNet)is proposed for arrhythmia classification.First,to solve the problem of low noise tolerance of the model,the network embeds the convolutional Block attention mechanism into the residual neural network to form ResNet-CBAM block as the backbone module for feature extraction of the beat.The convolutional block attention mechanism serializes the attention feature matrix in both channel and space,enabling the network to adaptively enhance the key features of each channel and position and suppress unnecessary noise information.Secondly,to solve the problem that the classification performance of the abnormal cardiac rhythm category is poor because the number of abnormal cardiac rhythm beats is far less than that of normal cardiac rhythm,the two-branch progressive learning strategy is adopted in the training of the network.In the initial training phase,the adapter assigns more weight to the general learning branch,and the model learns the common features of the overall ECG data set.As the iteration period increases,the weight assigned by the adapter to the rebalancing branch increases,and the model gradually focuses on the learning of abnormal rhythm characteristics.Finally,the two-branch feature aggregation is carried out and the loss function based on cross entropy band weight is applied to calculate the loss.The accuracy of this model has improved to 99.47% in the MIT-BIH arrhythmia database under Standard five categories of AAMI and demonstrated excellent performance in overall accuracy,noise tolerance,and anti-unbalance data robustness through ablation experiment,noise robustness experiment and anti-unbalance data robustness experiment.