Diabetes affects about 1.9% of the global population, mainly through Type 2 diabetes. Machine learning (ML) serves a pivotal role in enhancing diabetes prediction by analyzing complex datasets. Feature selection, a crucial ML pre-processing step, improved prediction accuracy by identifying relevant data and discarding irrelevant features. This study investigates the combination of metaheuristic algorithms and ML techniques to enhance diabetes prediction accuracy and computational efficiency. Utilizing the PIMA, Early Stage, and Vanderbilt datasets, experiments evaluated ten algorithm-model combinations based on metrics like accuracy, precision, the Wilcoxon test, and convergence curves. Key findings included that Firefly Algorithm-Logistic Regression, Bat Algorithm-Logistic Regression, and Cuckoo Search-Logistic Regression achieved 74.72% accuracy on PIMA; Firefly AlgorithmSupport Vector Machine and Cuckoo Search-Naïve Bayes achieved 83.39% accuracy and 96.15% precision on Early Stage; and Firefly Algorithm-Naïve Bayes achieved 92.88% accuracy and precision on Vanderbilt. These results highlighted the potential of integrating metaheuristics with ML methods to improve clinical diagnostics. Future research is recommended to validate algorithm robustness across diverse datasets to further optimize diabetes prediction strategies. Keywords complex_dataset; diabetes_prediction; disease_detection; feature_selection; prediction_accuracy