A privacy-preserving federated learning framework with graph neural networks for enhanced heart attack risk prediction

Abstract

Cardiovascular disease, a predominant global cause of death, underlines the critical need for sophisticated and privacy-conserving predictive systems for early risk assessment. This paper introduces a novel federated learning (FL) architecture that incorporates graph neural networks (GNNs) to facilitate secure and efficient heart attack risk prediction utilizing decentralized healthcare data. Our methodology generates graph-structured representations from preprocessed client data, dispersed among three clients to maintain data locality and safeguard patient privacy. The GNN models are trained locally, with only the acquired weights transmitted to a central server for secure aggregation. The proposed framework attains a classification accuracy of 96.79 % through comprehensive ablation studies and hyperparameter optimization, exceeding baseline models such as Graph Attention Networks (GAT), 1D Convolutional Neural Networks (1D-CNN), conventional machine learning, and ensemble techniques. Additional validation using an external dataset supports the model's robustness, with an accuracy of 98.77 %. Furthermore, trials conducted on a consolidated dataset demonstrate consistent performance, hence strengthening the framework's generalizability. These findings illustrate the potential of integrating GNNs with federated learning for privacy-preserving, high-performance prediction of heart attack risk in practical healthcare contexts.