This study develops a data-driven framework for early three-level maternal health risk prediction using routinely collected vital signs. The Maternal Health Risk Dataset, containing 1,014 anonymised records, was preprocessed through removal of clinically implausible values, SMOTE-based class balancing, Z-score normalisation, and second-order polynomial feature engineering. A soft-voting hybrid ensemble combining multi-layer perceptron, random forest, and XGBoost classifiers was then developed. Across repeated train-test partitions, the proposed model achieved approximately 0.98 accuracy and macro-F1, outperforming baseline models by about 10-11 percentage points. The framework shows strong potential for low-resource maternal health decision support.
Gao et al. (Tue,) studied this question.