A sustainable AI approach to diabetes prediction using metaheuristic feature selection

Abstract

Accurate diabetes prediction is a critical challenge in preventive healthcare due to the disease's rising global prevalence and its long-term societal, economic, and clinical impact. While machine learning models have demonstrated strong predictive capabilities, their effectiveness is often limited by redundant diagnostic features, reduced interpretability, and increased computational cost. This study describes diabetes prediction as an optimization problem that seeks to maximize predictive performance while minimizing feature dimensionality and computational effort. A unified optimization framework is developed that integrates wrapper-based and metaheuristic feature selection methods with a fixed Random Forest classifier. Recursive Feature Elimination, Genetic Algorithms, and Particle Swarm Optimization are evaluated on two heterogeneous realworld datasets from India and Iraq. The optimization objective explicitly models the trade-off between predictive accuracy and feature reduction. Experimental results based on repeated stratified cross-validation show that optimized feature subsets consistently improve F1-score and accuracy compared to baseline models while substantially reducing the number of required diagnostic features. Metaheuristic methods achieve competitive performance with highly compact feature sets, whereas Recursive Feature Elimination provides a favorable balance between performance gains, interpretability, and computational cost. The findings demonstrate that feature selection is a key enabler for developing efficient, interpretable diabetes prediction systems, particularly in resource-constrained clinical environments.