A Clinically Practical Postpartum Depression Predictor: Machine Learning Model Based on Simplified Indicators

Purpose: Postpartum Depression (PPD) is the most prevalent psychiatric disorder following childbirth, posing significant risks to both maternal and infant health. Consequently, prevention is paramount. This study aims to develop a clinically practical prediction tool using simple, readily accessible indicators to facilitate early intervention. Methods: In this multi-center study, data from 5,011 postpartum women were collected through structured questionnaires and electronic health records (EHRs). PPD was defined as an Edinburgh Postnatal Depression Scale (EPDS) score ≥10 at 6 weeks postpartum. Predictors within the dataset encompassed sociodemographic characteristics, pregnancy factors, delivery experiences, and infant care practices. Following feature selection, 11 machine learning(ML) models were constructed and evaluated. Model interpretability was enhanced using SHapley Additive exPlanations (SHAP). Results: The prevalence of PPD was 13.3%. Nine core predictors were ultimately identified: support from spouse, parents, parents-in-law, siblings, and friends; fetal distress; birth experience rating; marital status; and marriage duration. Among the algorithms evaluated, the Random Forest model demonstrated relatively superior performance (Training set AUC: 0.725, 95% CI: 0.697-0.752; Validation set AUC: 0.612, 95% CI: 0.568-0.656). SHAP analysis identified optimizing the birth experience and strengthening the social support system as key, clinically actionable intervention targets. Conclusion: This study confirms that a ML model based on simplified indicators provides moderate-performance risk stratification for PPD. This clinically practical tool equips frontline clinicians in resource-constrained settings in safeguarding vulnerable mothers through proactive early-warning systems during their critical transition period, thus reducing severe PPD and preventing devastating consequences.