Joining IoT, federated learning and explainable LLMs can help improve the privacy of data, make machine learning more understandable and better support healthcare teamwork involving multiple agents. In healthcare using the Internet of Things, wearable devices, diagnostics systems and patient monitors create useful health data all the time which calls for effective privacy measures. Because federated learning offers a distributed way to train models, there is no need for private data to be centralized which helps prevent data breaches. Still, issues such as non-IID data distribution, not enough local computational power and communication issues exist in multi-agent scenarios. The results generated by models are not always easily understood which further reduces the medical staff’s trust and confidence in their use. This hybrid framework uses explainable LLMs in combination with FL to support the learning of agents, help the model converge and let users explain decision-making. This approach allows every agent to collaborate, does not give up their privacy and makes sure the rationale for forecasts is open to both doctors and stakeholders. A new optimization approach is created to deal with model drift and an uneven amount of contributions from each user. The system is checked with scenarios and an artificial setting for the Internet of Things to be sure it works as expected and stays within privacy rules. The analysis of quantitative performance looks at accuracy, latency, the cost of communicating and how explainable the model is. Multi-modal charts and tables show that the proposed method outperforms traditional ways of FL. The research creates a way to train AI models that is privacy-conscious, leaves results understandable and minimizes resource needs in healthcare using IoT, supporting responsible and easy-to-understand AI usage in complex environments.