This study presents a robust machine learning framework to enhance stroke risk prediction and support global health objectives to reduce stroke-related morbidity and mortality. Using the Kaggle Stroke Prediction Dataset, the research integrates advanced data preprocessing, exploratory data analysis, and five machine learning algorithms: Logistic Regression, Random Forest, Gradient Boosting, XGBoost, and LightGBM. A stacked ensemble model serves as the core methodology, achieving superior predictive performance (accuracy 97%, F1-score 0.97, AUC-ROC 0.99) compared with individual models and prior works. Bayesian Optimization ensures optimal hyperparameter selection, while explainability methods such as SHAP and LIME bolster model interpretability to meet clinical transparency demands. The methodology aligns with the NZ Ngā Tikanga Paihere Framework, embedding principles of data privacy, cultural sensitivity, and fairness. Addressing challenges such as data imbalance and computational scalability via SMOTE and distributed computing, the model demonstrates robust performance, validated through bootstrapping and cross-validation. This research advances ethical, accurate, and actionable AI-driven stroke prediction for healthcare applications.