Predicting maternal health risk using PCA-enhanced XGBoost and SMOTE-ENN for improved healthcare outcomes

Maternal health remains a global priority, especially in low-resource settings where timely risk identification is critical. Traditional machine learning models often suffer from poor generalizability, data imbalance, and computational inefficiencies. This study proposes an enhanced predictive model combining SMOTE-ENN data balancing and Principal Component Analysis (PCA) with XGBoost to improve maternal risk classification accuracy using minimal, easily collectible clinical features. The dataset of 1014 maternal health records comprising seven physiological features was sourced from a public repository. Preprocessing involved standardization, label encoding, and class balancing using SMOTE-ENN. PCA was applied for dimensionality reduction to enhance computational performance and reduce overfitting. Several machine learning classifiers including Decision Tree, Random Forest, LightGBM, Gradient Boosting, and SVM were evaluated, with XGBoost selected as the final model. Performance metrics included accuracy, precision, recall, F1-score, ROC-AUC, and 10-fold cross-validation. The PCA-enhanced XGBoost model achieved the highest accuracy (97.73 %), precision (98 %), recall (98 %), and F1-score (98 %). It outperformed all other models, particularly in identifying high-risk cases with minimal false negatives. Cross-validation confirmed the model's robustness (mean accuracy: 98.39 %), and ROC-AUC scores exceeded 0.998 for all classes, indicating near-perfect classification performance. This study validates a maternal health risk prediction model that is scalable for use in resource-constrained environments and interpretable within the limitations of the selected dimensionality-reduction approach. Its simplicity, high accuracy, and generalizability make it a promising tool for early clinical decision-making and intervention. • Developed a PCA-enhanced XGBoost pipeline with SMOTE-ENN balancing for maternal health risk prediction. • Achieved 97.73 % accuracy and near-perfect ROC-AUC across all risk categories. • Utilized only six easily collectible physiological variables for scalable, low-resource deployment. • Demonstrated improved minority-class recall compared to existing public implementations. • Discussed trade-offs between performance and interpretability for clinical applicability.