Optimizing Cerebral Stroke Prediction: Combining SMOTE-ENN and XGBoost on Imbalanced Medical Data

Cerebral stroke remains a major global health burden, ranking among the leading causes of death and longterm disability. Rapid and accurate identification of high-risk individuals is crucial for effective intervention and improved outcomes. However, clinical stroke prediction is often hindered by extreme class imbalance, where confirmed stroke cases are vastly outnumbered by non-stroke instances, resulting in high false negative rates with traditional machine learning models. To overcome this limitation, we propose a hybrid framework that integrates the Synthetic Minority Oversampling Technique with Edited Nearest Neighbors (SMOTE-ENN) for data balancing and noise reduction, followed by classification using eXtreme Gradient Boosting (XGBoost). Evaluated on a real-world dataset of 43,400 patient records, including 783 confirmed stroke cases, the model achieved 98.14% accuracy and 87.39% sensitivity, while reducing false negatives by 34% compared to the baseline approaches. This approach offers a promising direction for improving clinical decision support systems through earlier and more reliable prediction of stroke risk.