Design And Implementation Of A Federated Learning-Based End-Edge-Cloud Coordinated ECG Detection System

Arrhythmia is often associated with cardiovascular disease and in most cases increases the risk of heart failure,dementia,and stroke.Traditional methods of detecting abnormal heart rhythms are generally based on medical knowledge for diagnosis,which requires specialized knowledge,previous experience,and ECG monitoring equipment,and therefore there is a shortage of relevant medical resources.With the development of deep learning,the dependency of expertise has significantly decreased.However,most existing deep learning methods utilize only one view of the signal or spectrogram alone,ignoring information from a medical perspective,which can lead to relatively low detection accuracy.In addition,as multi-view models have large large amount of code and complex algorithms,and medical data has strong privacy,It is a major issue how to deploy the multi-view approach on heart monitors.To address the aforementioned detection accuracy,code volume,and privacy issues,this paper designs and implements a federated learning-based end-edge-cloud coordinated ECG detection system for accurate,efficient,and secure abnormal heart rhythm detection,reducing the intervention of professionals and alleviating the insufficient of medical resources.The system starts from multi-view learning and adopts the "cloud-edge-end" architecture to achieve long-term ECG acquisition,analysis,and management functions.The system achieves realtime detection by deploying efficient 1D convolutional neural network models at the end device.In addition,a multi-view model with raw ECG signal,time-frequency map,and Lorentz scatter plot as input is deployed at the edge to achieve end-to-edge collaborative detection and further improve detection accuracy.To ensure privacy,the cloud server is used as the federation learning server to achieve incremental training of the model on the edge end through the federation learning algorithm to achieve cloud-side collaboration and to realize the display and management of detection results.This paper contains three main works:(1)design a side-end collaborative ECG detection scheme integrating ECG signal acquisition,transmission,and analysis.An efficient arrhythmia detection algorithm combining medical knowledge and a one-dimensional neural network is designed and implemented,as well as a comprehensive analysis model based on multi-views,which finally achieves efficient end-edge coordinated detection.(2)To protect privacy as well as to ensure system robustness,design and implement an incremental training strategy based on personalized federal learning to achieve dynamic incremental updates of the model.(3)To design a cloud platform for easy viewing of ECG detection results and management of detection records,and to display and manage the detection results in work(1)through Web pages.Through testing,the system is able to achieve the expected goals well.The ECG detection function has high accuracy and low latency,which strikes a balance between performance and efficiency;through incremental training by federal learning,the system is able to provide accurate detection results for new data and patients;the cloud platform performance and functional tests meet the requirements.In summary,the ECG detection system has good application prospects and practicality.