Research On Arrhythmia Classification Algorithm In Different Data Domain

Cardiovascular disease caused by arrhythmia is a major killer of human health.Nowadays,early detection and prevention of arrhythmia diseases are problems that need to be solved.Electrocardiogram(ECG)signal is a common auxiliary tool for diagnosing arrhythmia.Combining computer-aided technique analysis can reduce the doctor’s time and improve the classification accuracy,it has become a research hotspot.From the perspective of clinical medicine,this paper designs arrhythmia classification system for the same database and different databases,and compares the advantages.The Feature Cascaded Fusion Network(FCFN)is proposed for the first time to extract the spatial and temporal feature of ECG signals.Based on FCFN,this paper combine with unbalanced processing techniques,intra-class and inter-class joint loss functions,and class-balanced OHEM method to train the model.In this paper,a FCFN-based arrhythmia classification system is designed for the same domain data.The system includes four aspects:1.Feature acquisition.The deep abstract feature of the ECG signal is extracted by the cascaded fusion structure of one-dimensional convolution(1D-CNN)and long-term short-term memory network(LSTM),and the connection mode of multi-scale feature fusion is designed to combine the features of different sensing regions and increase its nonlinear fitting ability to enrich the amount of information.2.Unbalanced processing.The distribution of arrhythmia data in each class is different,which affects the accuracy of classification.This paper explores several unbalanced processing algorithms based on downsampling and upsampling,and proposes an integrated sampling method to improve sample distribution.It is verified by experiments that this method can improve the accuracy of classification of arrhythmia.3.Model training.The intra-class and inter-class joint loss function is designed to constraint the network parameters and make the model update towards the direction which has small differences within classes,large differences between classes.For the difficult-to-separate samples,a class-balanced OHEM training method is proposed,which focuses on improving the classification accuracy of confusing samples.4.Classifier application.This paper explore three typical classifiers: support vector machines,random forests,and Sofmax.By the theoretical analysis of different classifiers and considering the complexity of the system,the experimental results show that Softmax’s classification effect is better than the other two classification methods,the system finally uses Softmax classifier.For the different domain data,this paper introduce the feature transfer and model transfer idea in transfer learning,and design the Mirrored Deep Transfer Network(MDTN)based on FCFN.The ECG signal sharing network parameter to extract features between different domains.This network reduces the difference between feature domains through the adaptation layer and the maximum mean difference loss.The MDTN-based arrhythmia classification system can solve the problem that model difficult to be updated caused by the rare labeled data in the medical field.This system can improve the domain shift caused problem caused by patient specificity and environmental differences,and solve the problem of inaccurate classification caused by data differences.