With the growing awareness of data privacy in healthcare applications, Federated Learning is emerging as a promising alternative to traditional centralized machine learning methods. Unlike conventional ML which relies on centralized data storage, FL allows decentralized training by holding data locally on devices, thus significantly reducing privacy risks. This paper compares the performance of Federated Learning (FL) models against traditional machine learning (ML) models for diabetes prediction. By the employment of a real-world dataset on diabetes, we experimented with four traditional ML algorithms that consist of Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbors (KNNs). For FL, we employed FedAvgM and FedProx because these allow decentralized training preserving privacy over data. FL models outperform traditional ML models in terms of accuracy, F1-score, and recall, with the FedAvgM achieving high performance. Further, FL ensures important privacy benefits because sensitive data remains local on devices. The findings present the promise of Federated Learning as a superior privacy-preserving approach to machine learning applications in healthcare.