FedShield: A Federated Learning Approach for Secure Heart Disease Monitoring

FedShield aims to develop a privacy-preserving system for heart disease prediction using federated learning (FL), enabling collaborative learning without centralized data storage while safeguarding sensitive patient data. The FedShield system employs the Flower framework with TensorFlow, utilizing FL across six simulated clients with horizontally partitioned, Non-IID datasets. Local training with Adam and FedAvg aggregation were combined with preprocessing techniques like Isolation Forest, SMOTE, and MinMaxScaler. Fernet encryption ensured secure model update transmission. FedShield achieved a maximum training accuracy of 99.16% over 45 FL rounds, with precision of 99.8%, recall of 99%, and F1 score of 0.999, demonstrating superior performance compared to centralized methods. FedShield effectively applies FL in healthcare, achieving high diagnostic accuracy while preserving data privacy. Future directions include scaling for larger datasets and integrating realtime monitoring for broader applications.