<![CDATA[Diabetes is a chronic metabolic disease and an increasingly widespread disease around the world, and early diagnosis is crucial. Methodology In this study, the performance of three machine learning models (Logistic Regression, K-Nearest Neighbors (KNN), and Naive Bayes) is reviewed under the task of diabetes classification using the Pima Indians Diabetes Dataset. To tackle the class imbalance, we applied imputation, SMOTE for the data preprocessing, and min-max scaling to enhance the prediction performance. Further, we have applied ensemble learning and stacking, where all three models have been used as meta-classifiers. The results indicate that KNN had the best individual model performance (accuracy 77.27%, AUC 0.8444), but the stacking ensemble with a meta-model being Logistic Regression is superior to any model (accuracy 80.52%, AUC 0.8604). This suggests that ensemble learning can also improve the accuracy of diabetes diagnosis. These findings demonstrate that combining multiple classification approaches may provide more stable predictions across different patient conditions and clinical attributes. In addition, the preprocessing stages contributed to reducing noise and improving data consistency before model training. The study also highlights the potential use of ensemble-based systems in supporting healthcare professionals during preliminary diabetes screening, particularly in environments with limited medical resources and increasing numbers of diabetes cases requiring rapid assessment.]]>
Copyrights © 2026
