<![CDATA[Diabetes is a growing global health challenge that requires advanced approaches for early detection and prevention. Previous research has often been limited to evaluation using a single data split, which can potentially yield unreliable model performance estimates. This study addresses that limitation by conducting a comprehensive and rigorous evaluation of eight machine learning algorithms—including classical models, ensemble methods, and Multi-Layer Perceptron (MLP)—using the Pima Indians Diabetes dataset. The applied methodology includes data preprocessing, systematic hyperparameter optimization, and, most importantly, robust performance validation through Multi K-Fold Cross-Validation (K=5,10,15,20). Initial results showed perfect accuracy (100%) for the K-Nearest Neighbors (KNN) model; however, this finding was proven to be an artifact of a fortunate data split (lucky split) after undergoing cross-validation procedures. The more reliable validation results instead revealed the exceptional superiority of the Multi-Layer Perceptron (MLP) model, which achieved 94.96% accuracy with high stability (standard deviation 0.0356) in 20-fold cross-validation. Meanwhile, classical models such as Logistic Regression and Support Vector Machine (SVM) demonstrated high robustness and consistency. These findings significantly contribute to the field of health informatics by emphasizing the importance of rigorous validation methodology and identifying MLP as a highly strong predictive model candidate for diabetes detection. For practical application, this study recommends MLP along with stable classical models as the foundation for developing reliable clinical decision support systems, with the note that further external validation testing is necessary.]]>
Copyrights © 2026
