Decentralized Intelligence for Cardiac Health: A Federated Learning Approach to Privacy-Conscious Disease Prediction

Abstract

Cardiovascular disease (CVD) remains the leading cause of mortality worldwide, with early detection being vital for reducing life-threatening complications and improving patient prognosis. Traditional machine learning (ML) models for CVD prediction require centralized access to large-scale clinical data, but such access is often restricted due to privacy regulations and institutional barriers. To overcome these challenges, this study proposes a privacy-preserving federated learning (FL) framework that enables multiple healthcare institutions to collaboratively train predictive models without sharing sensitive patient data. The framework employs a multilayer neural network trained using the Federated Averaging (FedAvg) strategy, combined with data preprocessing techniques including standardization and class balancing via Synthetic Minority Oversampling Technique (SMOTE). Experimental evaluations using a real-world CVD risk dataset demonstrate that the proposed federated model achieves a classification accuracy of 94.0%, an F1-score of 93.5%, and outperforms centralized and local baselines. Communication overhead is kept minimal, averaging ~75 KB per round, and the model shows strong convergence across heterogeneous client data. An ablation study further confirms the critical role of SMOTE and local training configurations in model performance.