Machine learning for early prediction of preterm birth

Abstract

Abstract Background Preterm birth (PTB), defined as delivery before 37 completed weeks of gestation, remains a major cause of neonatal mortality and long-term morbidity worldwide. Conventional risk assessment strategies, including cervical length measurement and biomarker-based screening, have shown limited predictive performance. Machine learning (ML) may improve PTB prediction by integrating heterogeneous clinical data and identifying complex risk patterns. Methods A structured narrative review was conducted by systematically searching PubMed, Web of Science, Scopus, IEEE Xplore, and Google Scholar for studies published between January 2010 and April 2024. Studies were screened according to predefined eligibility criteria, and 14 studies evaluating ML models for PTB prediction using human pregnancy datasets were included in the final narrative synthesis. Data on study characteristics, predictor variables, ML methods, validation strategies, and model performance were extracted and synthesized narratively. Results A total of 14 studies met the predefined eligibility criteria and were included in the final narrative synthesis. ML models generally performed better when they used longitudinal electronic health records (EHRs), repeated measurements, or richer maternal clinical histories. Ensemble approaches such as random forest, gradient boosting, and stacking models, as well as deep learning (DL) methods including artificial neural networks (ANNs), recurrent neural networks (RNNs), and long short-term memory networks (LSTMs), were frequently among the best-performing models. Across the included studies, previous PTB, cervical length, maternal age, hypertensive disorders, diabetes, and body mass index (BMI) emerged as the most consistently reported predictors of ML model performance. However, the evidence was highly heterogeneous in data sources, outcome definitions, validation methods, and performance reporting. External validation, calibration assessment, and real-world clinical evaluation were limited. Conclusion ML shows considerable promise for improving early PTB risk stratification, but clinical translation will depend on early prediction, interpretable outputs, and robust external validation. Although ML models generally demonstrated promising predictive performance, interpretation of these findings is limited by methodological heterogeneity and inconsistent external validation across studies. Future studies should prioritize standardized reporting, multicenter external validation, calibration assessment, explainability, and prospective implementation studies to facilitate routine clinical adoption.