Maternal health risk prediction in Bangladesh using machine learning

Maternal mortality risk in Bangladesh remains a critical public health challenge, compounded by rural access gaps and the absence of scalable, data-driven early-warning systems. This study presents a reproducible, interpretable machine learning framework for maternal health risk classification using an IoT-collected dataset of 1,014 patient records and six physiological indicators; a deduplication audit identified 562 repeated sensor readings, a finding which is documented in the exploratory analysis. A rigorous pipeline was implemented encompassing five clinically grounded engineered features - Mean Arterial Pressure, Shock Index, Pulse Pressure, BP Ratio, and Composite Risk Score - alongside SMOTE-based class imbalance correction applied strictly post-split to prevent data leakage. Seven classifiers were systematically evaluated across two experimental tracks: the raw six-feature dataset and the eleven-feature engineered dataset. On the raw six-feature dataset with SMOTE (training: 811 → 1{,}218 samples; test: 203 samples), Random Forest achieved the best overall performance (Accuracy: 88.2%; Macro Recall: 0.889; F1: 0.888; AUC: 0.966), confirming its suitability as the champion model. XGBoost achieved the highest AUC (0.967) with marginally lower Macro Recall (0.868). Feature importance analysis revealed Blood Sugar (28.4% MDI) and the engineered Composite Risk Score (12.2% MDI) as the two dominant predictors, validating the clinical feature engineering approach. Feature engineering benefited weaker models most (Logistic Regression +3.3 percentage points in Macro Recall) while the strongest tree ensembles marginally preferred the SMOTE-balanced raw feature space. An interactive Tableau dashboard translates predictive outputs into accessible visual analytics for clinical and policy decision support.