Determinants of maternal postnatal care utilization in Bangladesh: A machine learning and SHAP-based analysis of BDHS 2022 data

Abstract

Postnatal care (PNC) plays a crucial role in minimizing maternal and neonatal morbidity and mortality, but the uptake of services in Bangladesh remains below the recommended level. Although logistic regression has been widely used, it may miss complex nonlinear interactions among social, economic, and healthcare factors. This study contributes to the body of knowledge by using machine learning (ML) to identify the most significant determinants of PNC and to enhance prediction accuracy. We compared logistic regression to several ML models, including Random Forest, XGBoost, CatBoost, Support Vector Machine, AdaBoost, and Gradient Boosting, using nationally representative data from the 2022 Bangladesh Demographic and Health Survey (BDHS) with ADASYN oversampling to correct class imbalance. Among all models, Random Forest achieved the highest AUC (0.9050), closely followed by XGBoost (0.9036) and CatBoost (0.9028), all of which substantially outperformed logistic regression (AUC = 0.8470). SHAP analysis of the Random Forest model indicated that delivery place, husband’s occupation, rural residence, wealth index, and media exposure were the most influential predictors of PNC utilization, alongside maternal education, women’s occupation, and age-related factors. The results indicate that ML is more effective than classical procedures for revealing latent patterns and making accurate predictions. Policy implications include encouraging facility-based deliveries, improving maternal education, reducing wealth disparities, and enhancing media coverage of health, particularly among rural and low-income groups. This paper not only identifies key drivers of PNC in Bangladesh but also demonstrates how ML can supplement traditional methods to reinforce maternal health policy and interventions.