STACKED ENSEMBLE MODEL FOR EARLY DIABETES PREDICTION USING RFECV-BASED FEATURE SELECTION

Abstract

 Diabetes mellitus (DM) remains a widespread health issue globally, particularly in developing nations where early screening and detection are often inadequate. This research investigates the question: Can a stacked ensemble framework using optimized feature selection significantly enhance the early diabetes (Type 2) detection from structured tabular data? This work addresses the problem by implementing a stacked ensemble where LR (Linear Regression), RF (Random Forest), and XGBoost serve as base models, and a Gradient Boosting classifier functions as the meta-learner. Recursive Feature Elimination with Cross-Validation (RFECV) is employed to identify the most relevant features and improve the predictive accuracy of model. Experimental evaluation on a publicly available diabetes dataset demonstrates that the proposed ensemble approach significantly outperforms individual classifiers with accuracy of 99.3% and F1-score of 1. Moreover, the application of RFECV consistently boosts model performance across the board. This work focuses about the efficiency of ensemble learning with feature selection in constructing accurate, reliable, and scalable clinical decision support systems for early diabetes diagnosis. The findings highlight the potential of integrating advanced ensemble techniques with optimized feature selection to reduce misdiagnosis risks and support timely medical intervention. This study provides a scalable framework that can be adapted to other chronic disease prediction tasks in healthcare analytics.