HFedLMR: Personalized and Hierarchical Federated Learning for Consumer Healthcare IoT With Non-IID Data

Federated learning (FL) in Consumer Healthcare IoT faces challenges in non-IID data heterogeneity, scalability, and personalization. This paper introduces Hierarchical Federated Layer-wise Model Recombination (HFedLMR), a novel framework that enhances model generalization and adaptation by dynamically selecting and recombining important model layers. Unlike traditional FL approaches, HFedLMR leverages layer importance scores to optimize knowledge transfer, ensuring better personalization while reducing communication overhead. Extensive evaluations on MedMNIST (PathMNIST, BloodMNIST) and CIFAR-10 demonstrate that HFedLMR achieves up to 7% higher accuracy in highly skewed Non-IID settings, 6% in moderate cases, and outperforms HFedMR by 4%. These results confirm that HFedLMR effectively balances local model personalization and global generalization, making it a highly efficient federated learning solution for real-world healthcare AI applications.