Diabetes Mellitus (DM) is a chronic and rapidly increasing health condition, affecting millions worldwide due to factors such as modern lifestyles and inadequate early detection methods. Current clinical diagnostics, while effective, often fail to identify early-stage DM, resulting in delayed treatment and higher risks of severe complications. This study proposes a hybrid ensemble machine learning model that combines both parallel and sequential ensemble methods with forward and backward feature selection techniques to enhance the early prediction of DM. The ensemble methods include J48, Classification and Regression Trees (CART), Decision Stump, Random Forest for parallel ensemble methods, and Gradient Boosting, XGBoost, and AdaBoostM1 for sequential methods. The study utilized a diabetes dataset containing features such as glucose levels, blood pressure, insulin levels, and BMI, applying the ensemble models to improve prediction accuracy. The experimental results showed that Random Forest, from the parallel ensemble methods, achieved a classification accuracy of 100%, significantly outperforming individual classifiers. Similarly, Gradient Boosting, from the sequential ensemble models, also yielded 100% accuracy. The combination of these models through a voting ensemble further enhanced the system's performance, producing superior prediction results with minimal errors. The findings emphasize that combining multiple ensemble techniques with feature selection can dramatically improve predictive performance. This study contributes a robust and scalable model for real-time diabetes prediction that can assist in the timely diagnosis and management of diabetes, potentially reducing global health risks associated with this disease.

Diabetes Mellitus ensemble learning feature selection machine learning Early Detection