Explainable Federated Learning for Secure and Transparent Medical Diagnosis in IoT-based Smart Hospitals

The fast uptake of IoT-enabled medical devices to create smart hospitals has revolutionized healthcare delivery, affording real-time monitoring and personalized diagnostics. But, the central training of AI models based on patient data raises serious issues of privacy, transparency, and trust. This work proposes an Explainable Federated Learning (XFL) framework that jointly integrates Explainable AI (XAI) mechanisms with Federated Learning (FL) to enable a secure, privacy-preserving, interpretable medical diagnostic across decentralized healthcare systems. The proposed XFL framework uses edge-based federated optimization to train deep neural models on decentralized patient data, maintaining the locality of that data. The framework enables mutual co-existence of model-agnostic explainability mechanisms such as SHAP and LIME which generate human-comprehensible justifications for a diagnostic decision, improving clinical trust and accountability. To enhance model robustness and confidentiality of the data, we integrate differential privacy, secure aggregation, and lightweight blockchain logging, allowing for auditability and protection against adversarial attacks. We provide experiments using real-world healthcare datasets such as COVIDx and MIMIC-III to demonstrate that our framework can build diagnostic models with competitive accuracy, while providing strong privacy assurances and information for clinicians. This work responds to an urgent requirement in AI-enabled healthcare; not only should it ensure what a model predicts, but also why, so that smart hospital ecosystems may be ethically aligned, safe, and transparent.