Predicting Maternal Outcomes Using Tree-based Methods in Machine Learning

Abstract

Maternal health, with its global significance for maternal mortality rates, is a paramount concern. This study focuses on leveraging tree-based algorithms to aid healthcare providers in informed decision-making for expectant mothers. Analyzing 4,000 antenatal care records in Nigeria's Niger Delta area (2018–2022) identified 15 critical features using Principal Component Analysis (PCA) to predict outcomes like stillbirth, full-term birth, preterm birth, miscarriage, placenta previa, and maternal mortality. Decision Tree (DT) prioritizes Hemoglobin Level (HL), Random Forest (RF) includes HL, Pulse Rate (PR), and Packed Cell Volume Level (PCVL). AdaBoost (ADA) emphasizes HL, Maternal Weight (MW), and Preeclampsia (PREE). Gradient Boosted Trees (GBT) consistently prioritizes HL, PREE, and MW, with Extreme Gradient Boosting (XGB) aligning with these features. A bar chart illustrates precision scores, with XGBoost leading at 0.95, GBT at 0.93, Random Forest at 0.92, AdaBoost at 0.91, and DT at 0.90. These findings offer valuable insights for healthcare professionals and researchers aiming to enhance maternal health outcomes. Future research avenues include exploring the synergy of tailored logistic regression models with gradient-boosted algorithms to enhance discrimination and calibration. Additionally, combining gradient-boosted trees algorithms with SHAP (Shapley Additive Explanations) could provide deeper insights into feature importance and predictive performance improvements.