In order to protect sensitive patient data across heterogeneous computer platforms, the federated learning (FL) technique has emerged in systems. In this work, we present the Federated Health Fog framework, which was meticulously created to address dispersed learning issues in IoT-enabled healthcare systems with limited resources, especially those that are energy-efficient and delay-sensitive. Conventional federated learning methods suffer from high compute requirements and communication costs. Their inability to gather global data on a single server is the main cause of this, leading to inefficient training models. By promoting a carefully positioned role in the network, we offer a novel solution to these issues. To optimize, greedy heuristic technique is employed to serve for every cloud. For three benchmark algorithms examined in this paper, the Federated Health Fog system dramatically reduces energy usage by 57.98, 34.36, and 35.37%, as well as communication delay by 87.01, 26.90, and 71.74%. The outcomes of our tests unequivocally demonstrate that Fed Health Fog is successful in lowering the quantity of global aggregation cycles when contrasted with cutting-edge substitutes. These results demonstrate how Federated Health Fog can revolutionize federated learning for applications that are resource-constrained and sensitive to delays in IoT contexts.