Federated Reinforcement Learning Using MAML and Graph Neural Networks for Safe IoT-Based Chronic Disease Monitoring
Year: 2026
Type: conference-paper
Abstract: Accelerating prevalence of non-communicable chronic diseases imposes urgent demand for intelligent, patientspecific treatment systems capable of continuous operation over heterogeneous physiological data streams. Conventional centralized reinforcement learning (RL) architectures face three compounding limitations when deployed in Internet of Things (IoT) healthcare environments: inability to accommodate non-identically distributed patient data across edge nodes, exposure of sensitive health records during model training, and slow policy adaptation when patient clusters evolve dynamically. This paper addresses all three limitations through a unified federated framework that integrates Model-Agnostic Meta-Learning (MAML) for rapid cluster-level policy initialization, Dynamic GraphSAGE for inductive construction of patient similarity graphs from continuous IoT streams, Federated Proximal Policy Optimization (FedPPO) for distributed policy refinement without raw data transmission, Multi-Agent Deep Deterministic Policy Gradient (MADDPG) for cooperative inter-cluster treatment coordination, and SimCLR-based contrastive pre-training for robust physiological state encoding. Evaluated on synthetic datasets derived from MIMIC-III and UK Biobank distributions spanning 5,000 patient traces over 90-day windows, the proposed architecture converges 45-50% faster, improves treatment success by 20-25%, reduces chronic exacerbation incidence by 15-20%, and maintains communication overhead below 10% of centralized training bandwidth. These outcomes confirm that meta-learning-guided federated RL is a viable and privacy-compliant foundation for scalable chronic disease management over decentralized IoT infrastructure.
Institution: Dravidian University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204844845