Explainable Machine Learning Approaches to Predict Knowledge of Pregnancy Danger Signs Among Women in the Haor Region of Bangladesh

Background Knowledge of pregnancy danger signs is critical for timely maternal healthcare, yet awareness remains limited in vulnerable regions such as the Haor area of Bangladesh. This study aimed to predict women’s knowledge of five key pregnancy danger signs and identify the associated predictors using machine learning approaches. Methods Cross-sectional data were collected from 800 women across 40 Union Parishads via face-to-face interviews. Ten machine learning models such as Logistic Regression (LR), Random Forest (RF), XGBoost, LightGBM, CatBoost, AdaBoost, SVM-RBF, KNN, ANN, and Decision Tree (DT) were trained and compared for predictive performance. Feature importance was evaluated using global (importance scores) and local (LIME and SHAP) explainability methods. Results Slightly over half of women (53%) recognized at least five pregnancy danger signs. Among all models, AdaBoost achieved the highest predictive performance (AUROC with 95% CI = 0.91 (0.87-0.94)). Feature importance analyses consistently identified maternal knowledge of neonatal danger signs, maternal education, husband’s occupation, household health expenditure, facility accessibility, and exposure to postnatal care and breastfeeding messages were associated with knowledge of pregnancy danger sign. Conclusions AdaBoost provides the most accurate prediction of maternal knowledge of pregnancy danger signs in this setting. Interventions should focus on improving maternal knowledge while addressing structural and socioeconomic barriers to optimize maternal health outcomes. These findings inform targeted, data-driven public health strategies in marginalized regions, contributing to the achievement of SDG 3 and SDG 10.