Work

A diffusion world model and multi-agent reinforcement learning framework for predictive routing in healthcare IoT

Year: 2026
Type: article
Abstract: Healthcare Internet of Things (H-IoT) networks support monitoring through wearable sensors, implantable devices, bedside units, mobile gateways, edge nodes, and cloud platforms. Reliable routing is challenging as medical packets are delay-sensitive, while networks face limited energy, mobility, lossy links, congestion, traffic bursts, and sudden node or link failures. This paper presents a diffusion world model and multi-agent reinforcement learning (MARL) framework for predictive H-IoT routing. The proposed design models the network as a dynamic graph, uses a Temporal Graph Transformer (TGT) to model spatial-temporal dependencies, forecasts link stability, congestion, and failure risk, and uses decentralized agents to select next hops. In ten paired, pre-configured software-emulation scenarios, the executable routing reference prototype achieved measured mean packet delivery ratio of $$56.24\pm 4.75\%$$ , delay of $$42.79\pm 3.34$$ ms, application goodput of $$6.11\pm 0.51$$ kbps, repair overhead of $$0.151\pm 0.024$$ , sensor-radio energy of $$8.94\pm 0.02$$ J, recovery time of $$0.0297\pm 0.0005$$ s, and restricted mean time to first sensor death of $$272.5\pm 44.9$$ s. Source-informed baseline adaptations achieved higher delivery and goodput and lower delay, while the prototype recovered successful routes faster. A separate 20-seed federated-predictor benchmark on emulated patient-group data reported patient-level DP-SGD accountant bound $$(\epsilon ,\delta )=(8,10^{-5})$$ . DP-FedAvg with rehearsal achieved task-1 and task-2 accuracies of $$72.26\pm 1.06\%$$ and $$70.16\pm 0.94\%$$ , respectively, with $$2.24\pm 0.84$$ percentage points of task-1 forgetting.
Language: English
Open Access status: gold
Estimated APC paid: $1,590
OpenAlex ID: W7231330563
FWCI:
Cites: 27
Cited by:
Related to: 10