Research On Heartbeat Classification Based On Deep Learning

With the rapid development of my country’s economy and society and the gradual aging,the incidence of cardiovascular diseases continues to increase,people pay more and more attention to self-health management,and the requirements for heart monitoring systems are also increasing.Therefore,accurate classification of heartbeat types is one of the important ways to prevent these diseases.The traditional way of diagnosis based on ECG by relevant experts is time-consuming and prone to unobjective diagnosis results.Therefore,the use of computer-aided technology to accurately classify heartbeat types is one of the research hotspots in the diagnosis of cardiovascular diseases.Due to the rapid development of deep learning in recent years,it has received high attention in the detection and research of cardiac diseases.Therefore,this paper uses the characteristics of deep learning with self-learning features to build a deep learning classification model to predict the heartbeat type,which not only reduces the classification steps such as manually extracting the features of heartbeat data,but also avoids the limitations of manual feature extraction,thereby improving diagnosis.Efficiency and Accuracy.In this paper,two classification models based on convolutional neural network are used to predict the category of heartbeat.The main research contents are as follows:First of all,the heartbeat signal data used in this paper has a serious category imbalance problem.In order to prevent the classification model prediction results from biasing the heartbeat category with a large number of samples,SMOTE upsampling technology and comprehensive sampling technology are used for processing.Medium SMOTE upsampling works better.Secondly,the classification model research of feature self-learning is carried out.In this paper,a one-dimensional convolutional neural network with 6 convolutional layers and a 13-layer residual network are constructed.The experimental comparison shows that the residual network has a better classification effect.Specifically,the precision,recall,and F1 value of the heartbeat category with the smallest sample size increased by 1.11%,0.56%,and 0.83%,respectively.However,there is still a certain gap between the heartbeat prediction results of this category and the other three types of heartbeat prediction results,so the loss function of the residual network classification model is improved,and the Focal Loss in the background of target detection research is used as the loss function.Other experimental environments With the conditions unchanged,it can be found that compared with the residual network model prediction based on Cross Entropy,the classification model based on the residual network based on Focal Loss has a certain degree of improvement in the overall accuracy and the index values of each category..Among them,the recall rate of the category with the smallest number of samples increased by 0.4%,and the precision rate and F1 value also increased slightly.