Federated Learning is an emerging and promising approach for AI healthcare applications, which enables collaborative model training without direct data sharing. However, existing approaches are suffering from various challenges such as poor generalization across non-IID medical data distributions, rigid privacy budgets that degrade model performance, and impractical communication overhead for clinical deployment. The proposed study offers an adaptive privacy-preserving Federated Learning framework that provides three key innovations such as context-aware differential privacy with dynamic noise injection based on data sensitivity and clinical urgency, patient similarity-driven client selection to mitigate non-IID bias, and gradient sparsification with medical-adaptive compression. The study is evaluated according to the diabetic benchmark for mortality, that achieves robust performance across 10 clients, while ensuring a final privacy budget of \(\epsilon \) = 5.0. In this study, a theoretical analysis is also presented that proves the approach provides ( \(\epsilon \) , \(\delta \) ) differential privacy guarantees while maintaining the utility required for clinical decision support.