JMIR Publications

ABSTRACT

Background:

Postpartum depression (PPD) affects up to 20% of new mothers globally. Early detection is vital for better outcomes, yet traditional screening lacks scalability, personalization, and predictive power. Artificial intelligence—through machine learning, deep learning, and natural language processing—can enhance the early identification of mothers at risk for PPD with greater accuracy and efficiency.

Objective:

To systematically map the existing literature on AI-based methods for detecting and predicting PPD.

Methods:

Following the PRISMA-ScR framework, we systematically searched eight databases (MEDLINE, EMBASE, PsycINFO, CINAHL, Scopus, IEEE Xplore, ACM Digital Library, and Google Scholar) up to February 28, 2025. Two independent reviewers screened records and extracted data from eligible studies that applied AI techniques to detect or predict PPD. Narrative synthesis was used to summarize findings.

Results:

Out of 65 included studies, most were conducted in the U.S., China, and Bangladesh, using open-source healthcare surveys or EHRs—primarily small-scale clinical datasets. Common features extracted included demographics, clinical variables (e.g., maternal anxiety, delivery mode, breastfeeding, hypertension), LIWC-based linguistic markers, and limited biomarkers. Data preprocessing mostly relied on basic scaling (78.5%) and some missing-data imputation (44.6%). Short-term postpartum outcomes (≤12 weeks) were most frequently assessed (30.8%). Machine learning dominated (80%), especially Random Forest, SVMs, logistic regression, and gradient-boosted trees. Neural networks and stacking ensembles were rare. Internal validation (k-fold, hold-out) was standard, while external validation was scarce. Accuracy (75.4%) and sensitivity (73.9%) were commonly reported, but metrics like specificity, AUROC, and F1-score were often omitted.

Conclusions:

AI offers strong potential for scalable, personalized PPD risk prediction, but current studies are limited by small, non-diverse datasets, inconsistent validation, and a lack of explainable or ethical AI use. To advance the field, future work should focus on developing and sharing large, multimodal datasets, applying advanced preprocessing (e.g., robust feature selection, class imbalance handling), and using rigorous validation methods—including external testing and regularization—to enhance generalizability and clinical relevance.