Predicting Postpartum Depression Using Imbalance-Aware Machine Learning

Postpartum depression (PPD) affects nearly one in five mothers, yet many cases remain undetected. This study demonstrates the value of a leakage-resistant, rigorously designed machine learning pipeline for early prediction using psychosocial data from 1,430 women during pregnancy. Concept-level feature harmonization addressed data artifacts, while repeated group-aware cross-validation, calibration, and recall-oriented thresholding ensured robust evaluation. Class imbalance was handled using weighting and SMOTE. Models showed strong performance (ROC-AUC: 0.762-0.801; PR-AUC: 0.400-0.490), with logistic regression and ensemble methods performing best. High recall (>0.90) supports screening use, though moderate precision highlights the need for clinical follow-up. SHAP analysis identified anxious attachment and coping factors as key predictors. Overall, well-designed ML pipelines can enable early, reliable PPD risk stratification while minimizing bias and leakage.