Machine Learning-Based Prediction Models for Adverse Pregnancy Outcomes: A Systematic Review

BackgroundAdverse pregnancy outcomes, including gestational diabetes mellitus, preeclampsia, preterm birth, postpartum hemorrhage, fetal complications, and postpartum mental health disorders, remain important challenges in maternal healthcare. Early identification of high-risk pregnancies enables timely intervention and personalized care. Machine learning (ML)-based prediction models have emerged as promising tools for analyzing complex clinical data; however, evidence regarding their performance and clinical applicability remains heterogeneous.Objective:This systematic review aimed to evaluate ML-based prediction models for early identification of adverse pregnancy outcomes, focusing on algorithms, predictive factors, validation approaches, and model performance.Methods:A systematic review was conducted following PRISMA guidelines. PubMed/MEDLINE, Scopus, Web of Science, Embase, IEEE Xplore, ScienceDirect and SpringerLink were searched for studies published between January ۱, ۲۰۲۱ up to August ۲۰۲۶. Original studies developing or validating ML prediction models for pregnancy-related adverse outcomes were included. Data regarding study characteristics, algorithms, predictors, validation methods, and performance measures were extracted.Results:Of ۹۰ screened records, ۳۸ studies met the eligibility criteria and were included in the qualitative synthesis. Gestational diabetes mellitus and preeclampsia were the most frequently investigated outcomes. Common algorithms included logistic regression, random forest, support vector machine, XGBoost, gradient boosting, deep learning, and ensemble methods. Major predictors included maternal characteristics, obstetric history, laboratory biomarkers, blood pressure, ultrasound findings, and electronic health record variables. Reported AUC values ranged approximately from ۰.۷۰ to ۰.۹۵. Limited external validation and insufficient assessment of clinical applicability were frequently observed.Conclusion:ML models show potential for early prediction of adverse pregnancy outcomes; however, further multicenter validation, standardized reporting, and clinical integration are required before routine implementation.

کلیدواژه ها:

Machine Learning ، Artificial Intelligence ، Pregnancy Complications ، Predictive Models ، Maternal Health