Asynchronous Federated Learning-based ECG Analysis for Arrhythmia Detection

With the rapid elevation of technologies such as the Internet of Things (IoT) and Artificial Intelligence (AI), the traditional cloud analytics-based approach is not suitable for a long time and secure health monitoring and lacks online learning capability. The privacy issues of the acquired health data of the subjects have also arisen much concern in the cloud analytics approach. To establish a proof-of-concept, we have considered a critical use-case of cardiac activity monitoring by detecting arrhythmia from analyzing Electrocardiogram (ECG). We have investigated two Federated Learning (FL) architectures for arrhythmia classification utilizing the private ECG data acquired within each smart logic-in-sensor, deployed at the Ultra-Edge Nodes (UENs). The envisioned paradigm allows privacy-preservation as well as the ability to accomplish online knowledge sharing by performing localized and distributed learning in a lightweight manner. Our proposed federated learning architecture for ECG analysis is further customized by asynchronously updating the shallow and deep model parameters of a custom Convolutional Neural Network (CNN)-based lightweight AI model to minimize valuable communication bandwidth consumption. The performance and generalization abilities of the proposed system are assessed by considering multiple heartbeats classes, employing four different publicly available datasets. The experimental results demonstrate that the proposed asynchronous federated learning (Async-FL) approach can achieve encouraging classification efficiency while also ensuring privacy, adaptability to different subjects, and minimizing the network bandwidth consumption.