Stroke prediction using synthetic minority over-sampling technique and extreme gradient boosting

Stroke is the world's second leading cause of death; its early prediction benefits from interpretable, high-accuracy models that can guide prevention and care. Using the Kaggle stroke dataset, we applied SMOTE to balance class distribution and trained XGBoost, Random Forest, LightGBM, CatBoost, and SVM models. XGBoost achieved 97.26% accuracy with robust 10-fold cross-validation, outperforming prior baselines. Model outputs were interpreted using the Shapley Additive Explanations (SHAP) algorithm, which identified age and hypertension/blood pressure as dominant predictors, providing both case-level insights and global feature rankings. The proposed pipeline offers practical, interpretable stroke-risk prediction with state-of-the-art performance suitable for clinical decision support.