Federated learning with homomorphic encryption for secure real time ECG anomaly detection: A multi institutional privacy preserving framework

In modern clinical settings, tools for rapid and secure diagnosis of cardiovascular diseases are becoming more important, particularly given the sensitive nature of medical data and the growing need for decentralized data processing and remote patient monitoring. To our knowledge, however, a comprehensive system architecture that enables AI assisted real time analysis of ECG signals for remote patient monitoring using Privacy Enhancing Technologies (PETs) has not been previously implemented. In this study, we propose a secure system that performs real time disease diagnosis based on three lead ECG signals and facilitates remote patient monitoring. The system leverages a model trained under various Federated Learning (FL) scenarios using the PTB-XL ECG dataset, partitioned across different hospitals via horizontal, vertical, and combined strategies to simulate clinical environments. We integrated Homomorphic Encryption during training to protect against honest but curious adversaries and mitigate gradient inversion attacks. Expert cardiologists evaluated the system and confirmed both proper dataset usage and diagnostic accuracy. In real time operation, the system displays ECG signals with pulse rate, allowing rhythm irregularities to be assessed based on clinical parameters and generating alerts when deviations from normal cardiac patterns occur. The proposed framework enables multiple healthcare institutions to collaboratively train an ECG diagnosis model without sharing raw patient data, creating an end to end secure structure that protects institutional confidentiality and model integrity. Simulation results demonstrated diagnostic accuracies reaching up to 99% across federated learning scenarios, with encryption induced overheads ranging from 30% in Horizontal FL to 456% in Vertical FL, offering clear insight into performance security trade offs. These findings highlight a unified, real time, and fully privacy preserving ECG analysis framework that integrates Federated Learning and Homomorphic Encryption across multiple data distribution settings.