COMPARISON OF RESAMPLING TECHNIQUES & MACHINE LEARNING ALGORITHM ON MATERNAL HEALTH RISK

Abstrak

The high maternal mortality rate drives the need for accurate early detection systems for pregnancy risk. This study compares nine classification model combinations — three resampling techniques (SMOTE, SMOTE-Tomek Links, ADASYN) with three tree-based algorithms (Decision Tree, Random Forest, XGBoost) — using the UCI Maternal Health Risk dataset (1,014 records, 6 features). Preprocessing includes outlier handling via the IQR method and label encoding. Data was stratified-split 70:30 and validated with 10-Fold Cross Validation within an ImbPipeline framework to prevent data leakage. The best result was achieved by XGBoost + SMOTE with an Accuracy of 99.67% and F1-Score of 0.9967, followed by Random Forest + SMOTE-Tomek Links with a CV F1-Score of 0.9957 (±0.0065). Feature importance analysis identified Blood Sugar (BS) as the most influential attribute (importance 0.5953), followed by Systolic Blood Pressure (0.1815) and Body Temperature (0.1014). This study confirms that XGBoost + SMOTE is the optimal combination for maternal health risk classification and has strong potential for implementation as a clinical decision support system in primary healthcare facilities.