An Ensemble Machine Learning Approach for Maternal Health Risk Prediction with Emphasis on Sensitivity and Interpretability

Abstract

Maternal mortality is a topical health issue worldwide, especially in low-resource states where few developed diagnostics exist. To minimize morbidity, it is imperative to detect risks early; however, the current models often sacrifice the clinical imperative of sensitivity at the altar of predictive accuracy. Subsequently, a Sensitivity-Aware Ensemble Framework was proposed to address this issue and predict maternal health risks based on routine physiological parameters. The framework uses a soft-voting ensemble of three gradient-boosting classifiers, namely CatBoost, XGBoost, and LightGBM, all of which are strictly calibrated to alleviate class imbalance. The methodology also involves Multivariate Imputation by Chained Equations (MICE) to establish sound data handling and sets sensitivity-based decision boundaries to enable screening-based implementation. The results of the application to the UCI Maternal Health Risk data (n=1,014) have an Area Under the Curve (AUC) of 0.96 with a recall of 0.81, which exceeds the performance of the individual constituent models. Moreover, SHAP (Shapley Additive explanation) analyses also showed that the most common predictors were blood glucose and systolic blood pressure, which enabled clinicians to have a clear explanation of why they made a decision. Taken together, these findings indicate that a sensitivity-oriented ensemble learning system supplemented with explainability instruments can provide concrete screening assistance for maternal care in resource-limited settings.