Privacy-enhanced heart stroke detection using Federated Learning and Homomorphic Encryption

Heart stroke detection plays a vital role in early diagnosis and intervention leveraging machine learning (ML) models to predict stroke events from medical data. However, the use of such models often faces significant challenges related to data privacy and security, especially when sensitive health information is involved. To tackle these concerns, we present an innovative privacy-preserving approach for heart stroke detection using Federated Learning (FL) combined with Homomorphic Encryption (HE) utilizing the Cheon-Kim-Kim-Song (CKKS) scheme. FL is employed to enable edge nodes to locally train a Feed Forward Neural Network (FFNN) model on heart stroke data, without transferring any raw patient data to a central server, thus ensuring the data remains decentralized and private. During the communication between edge nodes and central server, CKKS encryption ensures that model updates remain encrypted throughout the aggregation process, allowing computations to be performed without decryption, thus enhancing privacy and security. Furthermore, the FedAvg is incorporated to aggregate model updates efficiently, ensuring robust model training while maintaining decentralized data integrity. Our model achieved an average accuracy of over 95 % in predicting heart stroke events, demonstrating the effectiveness of combining FL with an FFNN in terms of both performance and privacy. This approach allows the utilization of rich medical data for training while maintaining the security and confidentiality of sensitive health information, making it an effective option for real-world healthcare applications in contexts where privacy and security of data is of the utmost importance.