Diabetes is a chronic metabolic disorder characterized by increased blood sugar levels due to impaired insulin secretion. Accurate diagnosis is essential to prevent complications such as cardiovascular diseases, kidney failure, etc. In this study, we proposed a refined XGBoost based model for diabetes diagnosis. Our experiments on a heterogeneous Kaggle dataset of over 100,000 records show our model achieves 97% accuracy, 96% precision, 98% recall and a 0.98 Area Under Receiver Operating Curve (AUROC), outperforming five other machine learning models whose accuracies range from 65.1% to 94%. We further analysed global and regional diabetes prevalence to underscore the disease burden and discuss the implications of machine learning–driven diagnostics for improving public health outcomes.