Machine Learning Based Postpartum Depression Risk Prediction: A Case Study on a Bangladeshi Dataset with Transferability Discussions for Nepal

Abstract

Postpartum depression (PPD) is a common perinatal mood disorder that often goes undetected, affecting roughly 10-15 percent of mothers, with higher rates across the Middle East and Asia and about half of cases never reported. Awareness remains low in countries such as Nepal, making early detection difficult. This study used the dataset “Data for Postpartum Depression Prediction in Bangladesh” (Raisa & Kaiser, 2025) to test whether supervised machine learning can sort mothers into three EPDS-based risk categories (Low, Medium, High), and how much of that signal depends on the dataset's own depression-screening features rather than on sociodemographic and psychosocial variables. Four classifiers (Logistic Regression, Random Forest, XGBoost, and a Soft Voting Ensemble) were compared using a leakage-free pipeline, repeated stratified 5-fold cross-validation, paired McNemar and DeLong significance tests, bootstrap confidence intervals, and permutation/SHAP-based interpretability analysis. Under cross-validation, the Voting Ensemble and XGBoost were statistically indistinguishable (mean accuracy 76.5% versus 77.4%, macro-average ROC-AUC 0.91-0.92 for both), both modestly ahead of Logistic Regression and Random Forest, with no significant pairwise difference except a single class-level AUC comparison. Feature-importance and SHAP analysis showed that the antenatal PHQ-9 score dominates every model's predictions; removing all PHQ-2/PHQ-9 features cost every model 17 to 19 accuracy points (Random Forest fell from 76% to 59%), pointing to substantial redundancy between the EPDS-based target and the PHQ-based predictors, both collected in the same interview. Mutual-information feature selection showed that 10-30 of the ~87 encoded features recover most of the full model's performance, and SMOTE gave a small, consistent improvement on the minority Medium class. All models separated High- and Low-risk classes well (AUC > 0.93) but struggled with the Medium class (AUC 0.81-0.85), whose EPDS band overlaps the Low/High boundary. Gradient boosting and a soft-voting ensemble are competitive, well-calibrated screening aids under default hyperparameters, but a meaningful share of their accuracy reflects redundancy between two depression screens rather than novel signal from contextual risk factors, and transfer of these results to Nepal remains a hypothesis pending local validation.