Comparative Analysis of Homogeneous and Heterogeneous Ensembles for Diabetes Classification Optimization

Abstract

Diabetes mellitus is a chronic disease with an increasing prevalence worldwide, including in Indonesia, reaching 11.7% by 2023. Early prediction of this disease is essential for more effective management. This study aims to develop a diabetes mellitus prediction model using an ensemble learning approach, including homogeneous (boosting and bagging) and heterogeneous (stacking and blending) techniques. In this study, the boosting algorithm using AdaBoost with Random Forest as the base estimator showed the highest accuracy of 98%, with balanced precision and recall. The bagging technique, which also uses Random Forest as the base estimator, achieved 97% accuracy, although slightly lower than boosting. The stacking technique, which combines XGBoost, Gradient Boosting, and Random Forest as base learners, with Random Forest as the meta-model, yields similar accuracy of 98%, but with lower prediction error, demonstrating its ability to cope with more complex data. Blending, which uses a similar approach but with training on the entire dataset, gave 98% accuracy with shorter processing time and more efficient memory usage than stacking.