A HYBRID ANN-XGBOOST FRAMEWORK FOR EARLY DIABETES RISK PREDICTION UNDER CLASS IMBALANCE: EXTERNAL CLINICAL VALIDATION IN AN AFRICAN HEALTHCARE SETTING

Abstract

Diabetes is one of the leading chronic diseases worldwide and remains a major public health challenge, particularly in low- and middle-income countries where limited diagnostic resources often delay early detection and treatment. Although machine learning has shown considerable promise for diabetes prediction, most existing studies rely on publicly available datasets, compare only a limited number of algorithms, and rarely validate their models using real-world African clinical data. This study proposes a Hybrid ANN-XGBoost framework for early diabetes risk prediction under class-imbalanced conditions. The proposed architecture combines the nonlinear feature representation capability of Artificial Neural Networks (ANNs) with the gradient boosting classification mechanism of XGBoost. The framework was evaluated on a large-scale dataset containing 100,000 patient records and compared with eight baseline models, including Logistic Regression, Support Vector Machines, Random Forest, Gradient Boosting, Light GBM, CatBoost, standalone ANN, and standalone XGBoost. External clinical validation was subsequently performed using 2,182 patient records collected at the Endocrinology and Metabolic Diseases Center of CNHU-HKM in Cotonou, Benin.