Metaheuristic Hyperparameter Optimization of Machine Learning Models for Diabetes Prediction

Abstract

This study investigates the effect of metaheuristic hyperparameter optimization on machine learning-based diabetes prediction using a unified experimental framework. Five classifiers—Logistic Regression, K-Nearest Neighbors, Decision Tree, Random Forest, and XGBoost—were optimized using the Dhole Optimization Algorithm (DOA), Grey Wolf Optimization (GWO), Particle Swarm Optimization (PSO), and RIME. After removing duplicate records, 96,146 observations were retained and divided into stratified training and independent test subsets. Hyperparameter tuning was conducted only on the training data using five-fold cross-validation, while model performance was assessed using accuracy, precision, recall, F1 score, Matthews correlation coefficient, ROC-AUC, and PR-AUC. Among the 20 model-optimizer combinations, XGBoost-PSO achieved the highest test F1 score of 0.8119, whereas XGBoost-GWO produced the highest ROC-AUC of 0.9781 and PR-AUC of 0.8849. Decision Tree also showed competitive performance, while Random Forest-RIME achieved the highest recall but with lower precision and F1 score. At the optimizer level, GWO provided the strongest average predictive performance, whereas DOA offered the greatest computational efficiency. Cross-validation and independent test results were generally consistent, indicating satisfactory generalization. Correlation and permutation feature importance analyses identified HbA1c level and blood glucose level as the most influential predictors. Overall, the findings show that model and optimizer selection should jointly consider predictive performance, generalization, interpretability, and computational efficiency.