Intelligent Chronic Heart Failure Detection Using Federated Continual Causal Graph Attention Networks

This study presents FedCau-GNN, a novel federated continual causal graph attention network for intelligent chronic heart failure (CHF) detection that addresses key limitations of existing federated learning approaches, including static model training, catastrophic forgetting, limited adaptability to new hospitals, and reliance on correlational explanations. The proposed framework models multimodal cardiac data as evolving causal-temporal graphs, integrates Elastic Weight Consolidation with privacy-preserving generative replay for continual learning, produces counterfactual causal explanations, and employs federated meta-initialization to enable rapid adaptation with limited local data. Experimental results demonstrate sustained diagnostic performance under long-term distributional drift (AUC > 0.94), significantly improved causal explanation consistency (0.88), enhanced resistance to membership inference attacks, and low-latency inference (<200 ms) while maintaining GDPR- and HIPAA-compliant privacy. These findings indicate that continual, causally informed federated learning offers a robust, scalable, and privacy-preserving solution for real-world CHF diagnosis across heterogeneous healthcare institutions.