Data heterogeneity remains one of the most significant challenges in federated learning (FL), impacting model performance, convergence, and scalability. This issue is especially critical in healthcare, where data is distributed across multiple institutions, devices, and geographical regions, and privacy preservation is paramount. In this study, we propose a novel hybrid algorithm, Dynamic-Cluster Personalized Federated Learning (DCP-FL), which integrates Federated Averaging (FedAvg), Personalized Federated Averaging (p-FedAvg), and dynamic clustering techniques. DCP-FL enables clients to maintain personalized models while contributing to a shared global model, achieving a balance between generalization and personalization. The algorithm clusters clients based on the similarity of their model updates, allowing for targeted aggregation that mitigates the effects of non-IID data distributions. We evaluate DCP-FL using the cardiovascular disease (CVD) dataset and the Breast Cancer Wisconsin (Diagnostic) dataset under realistic non-IID settings in a simulated FL environment using the Flower framework. Experimental results show that DCP-FL achieves 86.8% global model accuracy, 84.5% local model accuracy, and convergence in 35 communication rounds, outperforming FedAvg, p-FedAvg, FedNova and FedClust in both performance and convergence speed. While the approach incurs slightly higher communication costs, the accuracy gains justify the trade-off. These results demonstrate the potential of DCP-FL for privacy-preserving, heterogeneity-aware model training in healthcare and other domains.