The growing prevalence of diabetes highlights the need for scalable, accurate, and privacy-conscious testing technologies. To train models, traditional machine learning (ML) techniques often rely on centralised infrastructure that aggregates patient data from multiple sources onto a single server. Even though these methods frequently produce highly accurate predictions, they raise serious concerns about data privacy, legal compliance, and computational scalability. This article presents a Federated Learning (FL) framework for diabetes risk prediction, utilis-ing the PIMA Indian Diabetes dataset to overcome these issues. Without sending raw patient data, the suggested approach utilizes a supervised Deep Neural Network (DNN) that has been cooperatively trained across several decentralised clients, including wearable health devices and institutional medical servers. Each client independently trains the model on its local dataset, sharing only model parameters with a central server. The global model is updated through Federated Averaging (FedAvg), thereby safeguarding data privacy while preserving diagnostic accuracy. Experimental results validate the effectiveness of the proposed approach in balancing performance and privacy, demonstrating its potential for real-world deployment in edge-based smart healthcare systems.