Decentralized Reinforcement Learning Framework for Secure and Personalized Assistive Healthcare

The increasing demand for privacy-sensitive and personalised remote healthcare for people with disabilities necessitates adaptive, secure, and interpretable artificial intelligence frameworks. The proposed research is a secure generative AI-based decision-support system that combines personalised federated learning and adaptive optimisation models that are driven by reinforcement learning that allows the distributed generation of customised severity prediction and triage creation without aggregating the data in a central location. The framework uses FedPer strategy that constitutes a global feature extractor shared among subjects with local subject-specific layers to incorporate the patterns of a disease globally, coupled with the individual variation. The adaptive round-wise optimisation is gradually improving the local models using federated error feedback, whereas the explainability by SHAP guarantees clinical transparency. A triage system which learns the severity and incorporates the context to generate an actionable healthcare priority is called a context-aware triage mechanism. Experimental evaluation on 5875 longitudinal observations from 42 subjects shows that the personalised federated model delivers R2=0.7963 with stable convergence in 30 rounds, achieving performance close to centralised learning while strictly preserving data locality. The analysis of communication overhead and inference latency proves that it is feasible to deploy in real-time. The generative capability of the framework is in the dynamic generation of personalised severity assessments and adaptive decision-support outputs based on signal analytics distributed across a range of sensors as opposed to synthetic data generation.