Cardiovascular disease remains a leading cause of death worldwide and causes nearly four out of every five premature deaths. However, the patient data needed for early risk prediction often sits locked away in separate hospitals and clinics. Centralized model training isn’t feasible because sharing raw health records risks privacy violations and data breaches. Federated Learning (FL) solves this problem by moving the model to the data instead of the other way around. In the FL setup, each client trains the same model on its own data and only sends weight updates to a central server, so sensitive records never move. In our work, we propose a privacy-preserving FL pipeline using Flower and TensorFlow with the Framingham Heart Study dataset. We split the data among six simulated clients using a Dirichlet-based non-IID scheme to mimic real-world differences in patient populations. To handle tabular medical features, we adopt the TabTransformer, whose self-attention layers automatically capture interactions, like how age, cholesterol, and blood pressure combine to influence risk, without manual feature engineering. We also use FedProx to add a proximal term that keeps local model updates from drifting too far. In just 20 communication rounds, our federated TabTransformer achieved 85.98% accuracy, about 4% better than the federated 1D-CNN and nearly on par with the centralized TabTransformer. Overall, our results show that pairing FL with attention-driven tabular models yields reliable cardiovascular risk predictions, safeguards patient privacy, and adapts seamlessly to diverse client datasets.