Leveraging machine learning for early detection of diabetes: A dataset-driven comparative study

Diabetes mellitus is one of the most prevalent chronic diseases worldwide. This study investigates the application of machine learning algorithms for early diabetes prediction, a crucial step in preventing severe health complications. Two benchmark datasets, the Pima Indians Diabetes dataset and the Scikit-learn Diabetes dataset (converted to binary classification), were analyzed using Logistic Regression, Support Vector Machine (SVM with RBF kernel), and XGBoost. The methodology included preprocessing, train-test splitting, feature scaling, and model evaluation with metrics such as Accuracy, ROC-AUC, Precision, Recall, F1-score, and Confusion Matrices. Logistic Regression, serving as the baseline, performed best on the Pima dataset with 75.3% accuracy and a ROC-AUC of 0.815, demonstrating its strength in structured medical data. On the Scikit-learn dataset, XGBoost achieved the highest accuracy (76.4%), while SVM produced the best ROC-AUC (0.841), showcasing its ability to capture complex non-linear patterns. Findings highlight the importance of dataset-specific model selection and the integration of linear and non-linear approaches for reliable healthcare decision-support systems.