Learning from Imbalanced Stroke Data: A Comparative Study of SMOTE and SMOTE-ENN on Structured Health Records

Abstract

Stroke prediction with machine learning is challenged by severe class imbalance, which can produce models that appear accurate but fail to detect true stroke cases. This study evaluates Logistic Regression, SVM, and XGBoost on the Kaggle Stroke Prediction Dataset under three training regimes: no resampling, SMOTE, and SMOTE-ENN. Performance is assessed using imbalance-aware metrics (precision, recall, F1, and PR-AUC). Results show that models trained on the original data achieve high accuracy, yet near-zero recall for the stroke class, while SMOTE substantially improves minority-class detection and overall clinical usefulness. SMOTE-ENN offers additional sensitivity gains in some settings but introduces trade-offs in precision. These findings highlight the importance of resampling and appropriate evaluation metrics when developing stroke risk prediction models on imbalanced health records. Repeated cross-validation, SHAP-based explainability, and calibration analysis further indicate the models are best used as exploratory screening aids rather than deployment-ready clinical tools. These findings are expected to contribute to predictive maintenance in health monitoring for better future health and life.