Federated Machine Learning for Cardiovascular Risk Assessment: A Decentralized XGBoost Approach
Md. Samiul Alom, Sharmin Sultana Akhi, Samsun Nahar Borsha, Naeem Mia, Fahim Shakil Tamim, Jubair Ahmed Nabin · 2025
Cardiovascular diseases (CVDs) remain the leading cause of mortality globally, emphasizing the need for accurate and privacy-preserving risk prediction models. This study proposes a decentralized Federated Learning (FL) framework employing the XGBoost algorithm to assess cardiovascular risk while safeguarding sensitive patient data across distributed clinical sites. The system allows multiple clients to train local models on their respective datasets without transferring raw data, followed by secure parameter aggregation into a global model. Experimental results demonstrate that the federated XGBoost model (Global Model) outperforms all individual clients, achieving a high accuracy of 94.7%, F1-score of 94.8%, and ROC AUC of 98.8%, significantly surpassing the local models that ranged from 80.4 % to 85.2 % accuracy. The framework effectively addresses challenges such as non-IID data distributions and class imbalance, showing robust class-specific performance and superior generalization across heterogeneous datasets. These findings highlight the viability of federated ensemble learning as a scalable and secure solution for real-time heart disease prediction in privacy-sensitive environments.