Diabetes imposes a substantial public health burden; according to the International Diabetes Federation, there were about 3.4 million diabetes related deaths worldwide in 2024, and in Iraq, the Federation reports that one in nine adults lives with diabetes in 2024, with 14,683 adult deaths attributable to diabetes and a total diabetes related health expenditure of 2,078 million United States dollars. The dataset analyzed in this study contains 1,000 records collected in 2020 from two Iraqi teaching hospitals and includes multiple clinical and laboratory measurements with three outcome classes, namely Non diabetic, Pre diabetic, and Diabetic, with a low prevalence of the Pre diabetic class and an imbalanced overall class distribution; the data are challenging because they contain many outliers, non homogeneous covariance matrices across classes, exact duplicate rows that were removed before modelling, and linear correlations among certain variables. The study objective was to train and evaluate models that discriminate among the three classes and yield accurate, well calibrated predictions for future cases in similar clinical settings, but the diagnostic properties of the data limited the applicability of classical discriminant functions; therefore two supervised learners were employed: Classification and Regression Trees (CART) and Extreme Gradient Boosting (XGBoost), together with preprocessing that removed exact duplicate rows and excluded VLDL because it is algebraically derived from triglycerides in mmol per liter as VLDL equals triglycerides divided by 2.2, which would introduce redundancy and multicollinearity. On the heldout test set, XGBoost achieved higher Accuracy at 98.18 percent compared with 97.58 percent for CART and higher Balanced Accuracy at 93.84 percent compared with 88.16 percent for CART, indicating that XGBoost provided the strongest overall operating point for this three-class task while CART remains useful when simple and transparent rules are required.