<![CDATA[Heart disease remains the leading cause of mortality globally, including Indonesia. However, developing accurate predictive models is often hindered by class imbalance in medical datasets, where positive cases significantly outnumber negative cases. This study optimizes heart disease prediction by applying SMOTE (Synthetic Minority Oversampling Technique) regionally to Decision Tree and Random Forest algorithms using the "Heart Attack Prediction in Indonesia" dataset from Kaggle, which contains rural and urban attributes. Following the CRISP-DM framework, SMOTE was applied separately for each region to capture local distributional diversity and reduce regional bias. Results demonstrate that regional SMOTE significantly improved recall and F1-scores for both algorithms, particularly in rural areas where Random Forest recall increased from 60.50% to 70.34%. Statistical significance was confirmed through paired t-tests and Wilcoxon signed-rank tests on 5-fold cross-validation results (p < 0.001). Fairness analysis using Demographic Parity Difference and Equalized Odds Difference confirmed equitable performance across populations (DPD < 0.005, EOD < 0.008). Random Forest consistently outperformed Decision Tree, achieving the highest F1-score of 66.19% in urban regions post-SMOTE. These findings support that regional SMOTE effectively enhances model sensitivity toward minority classes while maintaining spatial fairness in heart disease prediction]]>
Copyrights © 2026
