Contribution-Aware Hierarchical Federated Learning: A Novel Framework for Heterogeneous Medical IoT Devices in 6G Networks

With the rapid development of 6G technology and the evolution of healthcare Internet of Things (IoT), federated learning has gained significant attention for privacy-preserving distributed model training in medical scenarios. However, existing approaches face challenges in addressing device heterogeneity, communication overhead, and data privacy in healthcare IoT environments. This paper proposes a resource-efficient hierarchical collaborative federated learning framework designed for next-generation 6G-enabled healthcare IoT. We introduce an adaptive edge aggregation frequency adjustment mechanism and a contribution-based dynamic model aggregation weight adjustment mechanism to enhance system training efficiency and model performance. Additionally, a resource-balanced client selection algorithm and a self-organizing federated collaborative training approach are adopted to mitigate the impact of device heterogeneity and improve resource utilization. Experimental results on two medical datasets demonstrate that our approach achieves accuracy improvements of 7.96% and 8.69% over standard federated averaging, reaching 91.58% and 87.05% accuracy respectively. The framework reduces communication overhead by up to 48.4% while maintaining effective communication ratios of 78.6% in stable network environments.