Decentralized Horizontal Federated Learning Using Graph Neural Networks for ECG Signal Analysis

Abstract

Data privacy is a major concern in data analysis due to ethical and legal considerations. To overcome this issue, federated learning has transformed data analysis by putting data privacy first throughout the model training process. We proposed a horizontal federated learning system in this study after analyzing ECG data. Our methodology makes use of the decentralized method of federated learning, in which local models are trained on ECG datasets separately without the need for a central server. We collected data for our studies using two publicly available ECG datasets: the PTB-XL and Chapman datasets. Furthermore, we used a variety of preprocessing approaches to denoise the signals, as well as different feature extraction techniques for both datasets to discover important characteristics in the signals. In this method, the graph neural network (GNN) served as the foundation for the local models. As a result, local models are trained separately on the client side to create the global model for our proposed federated learning system. This approach attempts to create an efficient federated learning model with high accuracy, as well as a promising approach to data analysis in a variety of domains. Our experimental findings underscore the usefulness of federated learning in sensitive data analysis by showing that the suggested method achieves excellent accuracy while protecting data privacy. Future research on safe and decentralized machine learning applications in several domains is made possible by this work.