<![CDATA[Stroke remains a leading cause of mortality and long-term disability worldwide, including in Indonesia, highlighting the urgent need for early risk identification. Machine learning models for stroke prediction often suffer from severe class imbalance, where stroke cases constitute only 4.9% of clinical datasets, leading to biased predictions that favor the majority class. This study evaluates three ensemble and kernel-based algorithms Random Forest, XGBoost, and Support Vector Machinecombined with two resampling strategies (SMOTE and SMOTE-ENN) using the Healthcare Stroke Dataset (5,110 records, 11 clinical attributes). To prevent data leakage, resampling was strictly applied within each training fold of 5-fold stratified cross-validation, while all evaluations were conducted on the original imbalanced test set. The results demonstrate that XGBoost integrated with SMOTE-ENN achieved the highest minority-class sensitivity, improving PR-AUC by 23.5% (0.1537 vs. 0.1244 with SMOTE alone), while detecting 24% of stroke cases (12 out of 50) in the test set. Although cross-validation results indicate strong class discrimination with AUC-ROC values above 0.98, the low PR-AUC reflects the operational challenge of extreme class imbalance and the inevitable trade-off between recall and precision, resulting in an increased number of false positives. Consequently, the proposed model is best positioned as a first-tier population screening tool that flags high-risk individuals for confirmatory clinical diagnostics, rather than as a standalone diagnostic system. The approach maintains computational efficiency (training time < 0.12 seconds) and substantially improves model stability, evidenced by a 73% reduction in cross-validation variance. These findings support the integration of hybrid resampling techniques with ensemble learning as a practical and scalable framework for early stroke risk screening in resource-constrained primary healthcare settings.]]>
