This study presents a hybrid, metaheuristic-driven optimization framework for power hyperparameter tuning in predictive modeling based on large-scale annual health examination data. Different from conventional grid and random search strategies, the proposed method directly incorporates particle swarm optimization, artificial bee colony, and gravitational search algorithm into the training pipeline of multiple machine learning models, enabling adaptive exploration of high-dimensional parameter spaces under clinical data constraints. The approach was evaluated on a comprehensive dataset comprising 93 clinical attributes and 1,000 patient records, with a specific focus on ischemic stroke risk prediction. Random Forest, decision tree, support vector machine, and logistic regression models were optimized using the proposed hybrid structure and benchmarked against baseline configurations. Experimental results demonstrate consistent and statistically significant reductions in mean squared error, mean absolute error, and root mean squared error, alongside improvements in R² and classification accuracy exceeding 99% for optimized logistic regression models, while maintaining computational efficiency suitable for routine clinical deployment. Beyond performance gains, the study introduces a stacked ensemble architecture guided by metaheuristic-tuned base learners, enhancing model robustness and generalization across training and independent test sets. These findings demonstrate the practical novelty of integrating swarm and numerical optimization into clinical predictive pipelines, providing a scalable and domain-agnostic solution for high-accuracy risk decision support in preventive healthcare and other data-intensive applications.