Artificial Intelligence-Based Prediction of Postpartum Depression Using Structured Data: A Systematic Review

Abstract

Background: PPD is considered a substantial public health concern due to its negative consequences on both mother and child's well-being. Despite the potential for earlier recognition of PPD, early prediction of women who will develop this disease proves challenging due to the complexity of the PPD condition which includes demographics, clinical, obstetric and psychological variables. Machine learning (ML) is a promising tool for such risk prediction, but its clinical adoption is hindered by variability across data collection, pre-processing, model development, validation, and interpretability.
Objective: The aim of this systematic review was to evaluate systematically the ML-based prediction models for PPD based on structured data, with special focus on data set characteristics, predictor variables, pre-processing techniques, features selection and engineering, development, validation, performance and explainability.
Methods: Systematic literature review was performed following PRISMA 2020 guidelines in relevant electronic databases for studies published between 2015 and 2026. Studies included in this review were original articles written in English language which involve human participants applying ML-based prediction or classification models for PPD based on structured data. Study screening and selection were done independently by two authors resolving conflicts through discussion. Data extraction involved study characteristics, data source, predictors, pre-processing techniques, algorithms, validation strategy, performance measures and explainability.
Results: Nineteen studies met the inclusion criteria out of the initial set of 3,805 articles. Among the included studies, there was an increasing tendency towards the use of ensemble learning, deep learning, and hybrid models as well as conventional machine learning methods including Logistic Regression, Random Forest, Support Vector Machines, and XGBoost. Psychosocial variables – involving prenatal anxiety, depressive symptoms, stress, sleep disorders, social support, and previous history of mental health issues – appeared to be significant predictors in all studies. Nonetheless, considerable heterogeneity was present in the studies in relation to the adopted pre-processing steps, features selection techniques, outcome definition, and validation procedure. A formal four-domain PROBAST assessment found 18 of the 19 studies (94.7%) at overall high risk of bias, most often driven by inadequate sample size relative to the number of candidate predictors, validation on the training set or on internal cross-validation alone, or use of non-validated proxy outcome measures; only one study, which combined a large multi-site sample with external validation and reported calibration, was judged low risk across all domains.
Conclusion: There is great potential for applying machine learning for predicting PPD risk, but the evidence base remains at high risk of bias overall, and the current literature is limited by considerable methodological variability, scarce external validation, and problems concerning generalizability and clinical interpretability of the models. Substantial improvements in study design, sample size justification, and validation practice are needed before these models are ready for clinical translation.