Transformer Based Federated Learning Framework for Heart Disease Prediction

Heart related diseases are a life threatening which makes it crucial to develop efficient and automated predictive models for early diagnosis. Traditional methods rely on centralized data storage which imposes privacy risks and limiting access to diverse patient populations. This paper has proposed, Transformer-based Federated Learning (FL) model for heart disease prediction, enabling decentralized learning while preserving patient data privacy. The model utilized transformer model based self-attention mechanisms to find detailed feature patterns and improve predictive accuracy across multiple healthcare clients. FedProx aggregation strategy is used in proposed method to enhance model convergence and robustness in different client environments. Our novel approach is evaluated on a combined heart disease dataset, demonstrating 92.36% accuracy 91.06% F1-score. Also, comparative analysis with existing centralized and FL-based models further validates the superiority of our approach. Further, the optimized FL aggregation strategy for proposed method is FedProx. This study shows the promising results achieved through federated learning integrated with Transformer architectures for privacy-preserving, scalable, and high-performance heart disease prediction which plays the critical role for real-world clinical implementation.