| Cardiovascular diseases are the leading cause of death in China.Among them,Congestive Heart Failure(CHF)is a kind of cardiovascular disease with high mortality,high treatment cost and difficulty in early detection,which has attracted wide attention of researchers.Traditional diagnostic methods in hospital settings are not convenient for real-time monitoring of patients’ physical conditions at home.In view of this,taking peak-to-peak interval of short-term ECG signals as input,deep learning was used to construct a CHF diagnostic network based on multi-scale residual UNet++.Considering the risk of privacy disclosure of data shared by multiple institutions,federated learning was used to simulate collaborative training among multiple medical institutions,so as to effectively assist the early diagnosis of heart failure.The main contents of this thesis are as follows:Diagnostic model of heart failure based on short-term ECG R peak-R peak(RR)interval and multi-scale residual UNet++.The early diagnosis method of heart failure needs to use the long-term dynamic ECG signal,but the acquisition of long-ter m signal will bring a lot of inconvenience to the life of patients,and t he practicality of the model needs to be improved.In this thesis,we propose an end-to-end classification model based on the RR interval of ECG signals,which integrates the outputs of codec,decoder and intermediate unit through unified scale operation t o preserve the low-level details of the input RR interval fragments and extract high-level pathological-related features.In addition,residual variants with block convolution and compression excitation modules are used to enhance the network representation ability.Experimental results based on public and benchmark data sets show that the classification accuracy of this model for normal sinus rhythm and CHF is 89.83%,which is better than existing similar models.Considering that non-heart failure may incl ude multiple types of arrhythmias in practice,data on paroxysmal or persistent atrial fibrillation and apnea are further introduced.The accuracy of proposed method achieves 90.61% for a variety of non-CHF and CHF classifications,which proves the robustness of the method,and indicates that the fusion of RR interphase fragment and multi-scale residual UNet++ model can effectively assist the diagnosis of CHF.Diagnostic model of heart failure based on federated learning.Considering that it is difficult for a single institution to collect a large number of diversified data samples and the potential privacy and security problems caused by data transmission,this thesis dispersed the data to different clients and federalized the CHF diagnostic model,which promoted multi-institution data collaboration training while maintaining data anonymity.Experimental results based on public and benchmark data sets show that the federated model achieved 87.54% accuracy in distinguishing normal sinus rhythm from CHF,comparable to existing centralized diagnostic models for CHF.When more types of non-CHF signals are added,the accuracy of the proposed method is 87.71%.The performance of the federalized diagnostic model is similar to that of the centralized learning diagnostic model,indicating the feasibility of using multi-site data to cooperatively train the CHF diagnostic model without sharing patient data.Diagnostic toolbox for CHF based on ECG RR interval.To facilitate the popularization and use of the above methods,a CHF diagnostic toolbox is developed.It mainly includes: diagnostic model training module,which is convenient for medical institutions to update the model after adding data;diagnostic model prediction module,which can infer whether the current subject has CHF online,and promotes the application of related technologies in mobile health monitoring.This thesis has 27 figures,12 tables and 82 references. |
PDF Full Text Request