Diabetes mellitus is a major global health challenge, motivating reliable early-detection tools for timely intervention. This paper provides a comparative evaluation of traditional machine learning models, Logistic Regression, Support Vector Machines (SVM), Random Forest, and K-Nearest Neighbors (KNN), for diabetes prediction on the DiaBD dataset (5,288 patient records from Bangladesh). To address the pronounced class imbalance (6.5% diabetic vs. 93.5% non-diabetic), we apply the Synthetic Minority Over-sampling Technique (SMOTE) to the training data and assess performance using accuracy, precision, recall, F1-score, and ROC-AUC. Experimental results show that Random Forest yields the best overall performance after SMOTE, achieving 93.67% accuracy and 0.8798 ROC-AUC, outperforming the remaining traditional baselines. Feature-importance analysis further indicates that hypertension, glucose-related measures, and blood pressure are the most informative predictors. These findings support the practical use of traditional, interpretable machine learning models for diabetes risk assessment in resource-constrained healthcare settings.