Building a Predictive Model on Maternal Mortality Using Machine Learning: Comparison of Different Modelling Techniques

Abstract

Maternal mortality remains a critical public health issue in Ghana, influenced by complex socio-economic, health-related, and contextual factors. This study aimed to identify the most effective predictive modeling technique for maternal mortality and elucidate key risk factors. We developed and evaluated three predictive models: Logistic Regression, Random Forest, and Extreme Gradient Boosting (XGBoost) using data from the Ghana Maternal Health Survey (GMHS) 2017, which included 1,240 deceased women with detailed demographic, socio-economic, and health-related information. XGBoost emerged as the most robust and reliable model, achieving the highest average KFold score (0.8774), test accuracy (0.90), F1 score (0.47), and Jaccard score (0.82), indicating superior predictive performance. Significant predictors identified included place of death, marital status, blood pressure, traditional medication use, fever, season of death, and age at death. These findings highlight the need for targeted interventions to improve healthcare access, integrate traditional medicine into formal healthcare systems, and address socio-cultural barriers. Policy efforts should focus on enhancing healthcare access in rural areas, promoting gender equality in health decision-making, and targeting high-risk groups, particularly younger women and those with hypertension. Future research should include additional variables, conduct longitudinal studies, explore advanced machine learning techniques, and evaluate community-based interventions and policy impacts to further improve maternal health outcomes in Ghana.