Metaheuristic Class Balancing and Hyperparameter Optimization for Diabetes Classification: A Two-Stage PSO–GA Framework

Background/Objectives: Accurate identification of individuals with diabetes or prediabetes from population health data can support screening and population-level disease management. However, substantial class imbalance in health datasets can bias machine learning classifiers toward the majority class and reduce the detection of affected individuals. This study proposes a two-stage metaheuristic machine learning framework combining Particle Swarm Optimization (PSO)-based class balancing with Genetic Algorithm (GA)-based classifier optimization for diabetes classification using health, behavioral, demographic, and socioeconomic information. Methods: The proposed framework integrates two optimization stages. First, Particle Swarm Optimization (PSO) is used to select majority-class training observations to address class imbalance. Second, a Genetic Algorithm (GA) is applied to optimize the hyperparameters of three classifiers: Multilayer Perceptron (MLP), implemented as an Artificial Neural Network (ANN); Extreme Gradient Boosting (XGBoost) version 3.4.1 ; and Random Forest (RF). The framework is evaluated using the CDC Diabetes Health Indicators dataset, comprising 253,680 observations and 21 predictors. Predictive performance is compared across the original imbalanced data, random undersampling, and PSO-based balancing using Accuracy, Specificity, ROC-AUC, and average precision. Results: The proposed PSO–GA framework demonstrates strong discrimination between individuals with and without diabetes/prediabetes. The best-performing model achieves a ROC-AUC of approximately 0.95, an average precision of 0.87, and an accuracy of 0.90 in the reported evaluation. These findings suggest that combining optimization-based class balancing with evolutionary hyperparameter tuning can support predictive modeling in imbalanced population health datasets. Conclusions: The proposed two-stage PSO–GA framework provides a systematic approach to diabetes classification by integrating majority-class sample selection with hyperparameter optimization of ANN, XGBoost, and RF classifiers. The findings demonstrate the potential of metaheuristic optimization for improving classification performance in imbalanced health datasets. Further validation using independent populations and prospective healthcare data is required before clinical deployment.