Building a Machine Learning Prediction Model for Postpartum Depression Based on Ensemble Learning Strategies

Postpartum depression (PPD) is a common perinatal mental health disorder that leads to adverse maternal and infant outcomes. Existing screening strategies rely on self-reported symptom presence, potentially delaying early identification. This study aims to integrate biological, psychological, and social predictive factors using an ensemble learning (EL) strategy to develop a superior early prediction machine learning (ML) model for PPD.

By integrating sociodemographic characteristics, clinical baseline data, and psychological test results of 1323 postpartum women, PPD status was classified using the Edinburgh Postnatal Depression Scale (EPDS). Eight selection methods were used, and eleven base ML models were optimized through EL strategies (Voting and Stacking) to identify the model with the optimal predictive performance.

Twenty predictor features were selected for the model construction. The Voting top 5+XGBoost (eXtreme Gradient Boosting, Weight) model demonstrated the best overall performance (area under the curve [AUC] = 0.835) and the highest specificity (0.906), indicating that it is more suitable for differential diagnosis following an initial positive screening result. The Voting (Synthetic Minority Over-sampling Technique [SMOTE]) model performed exceptionally well in sensitivity (0.758) and clinical net benefit at low-risk thresholds, indicating its suitability for early screening for PPD.

The results demonstrate that ensemble learning models outperform individual ML prediction models. Voting top 5+XGBoost (Weight) and Voting (SMOTE) show respective advantages in predictive specificity and sensitivity. This suite of models can provide distinct data-driven prediction strategies tailored to different clinical application contexts.