Cardiovascular disease (CVD) remains a major global health burden and the leading cause of mortality worldwide. Although machine learning has shown promise for CVD risk prediction, centralized training on sensitive medical data from multiple hospitals raises serious privacy and security concerns. To address this issue, we propose FedCH, a novel federated learning method for collaborative CVD prediction. Specifically, FedCH quantifies client heterogeneity by measuring the similarity between local and global models together with client-specific prediction accuracy. These metrics are used to dynamically weight the global aggregation process, thereby improving the overall model performance while promoting fairness across clients. In addition, by incorporating historical global model parameters into the update process, FedCH mitigates local model fluctuations and enhances the generalization and robustness of the global model. Comprehensive evaluations on four public datasets and a real-world cardiovascular dataset from a tertiary hospital in China show that FedCH consistently outperforms five state-of-the-art federated learning baselines under heterogeneous client settings.