Federated learning enables collaborative model training across distributed healthcare institutions without centralizing sensitive patient data. However, existing approaches face challenges in balancing privacy protection and model utility. This paper proposes an adaptive differential privacy mechanism for federated learning in healthcare data aggregation that dynamically adjusts privacy budgets based on data sensitivity and model convergence. We introduce a privacy-utility trade-off optimizer that allocates privacy budgets adaptively across training rounds while maintaining rigorous privacy guarantees. Our experimental evaluation on the MIMIC-III and eICU datasets demonstrates that the proposed approach achieves 7.3% higher model accuracy compared to fixed differential privacy mechanisms while providing equivalent privacy protection with ϵ=2.5. The system reduces communication overhead by 31% through selective gradient aggregation and achieves convergence 23% faster than baseline federated learning approaches.