Machine learning to predict adverse perinatal outcomes: a systematic review and meta-analysis.

Background: Machine learning (ML) provides a promising approach to predict adverse perinatal outcomes, supporting early intervention to reduce neonatal morbidity and mortality. There has been a surge of publications using ML algorithms applied to routine clinical data to predict adverse perinatal outcomes. We aimed to assess the current evidence regarding ML model performance, their comparison with logistic regression, and risk of bias. Methods: We searched PubMed, EMBASE, CINAHL, Global Health, Web of Science and IEEE Xplore for studies published from 1 Findings: We retrieved 38,322 studies, and 90 studies were included. Overall ML model performance was moderate, with a median average AUC based on internal validation of 0.73 (range 0.50-0.93) for PTB, 0.68 (0.62-0.85) for SGA, and 0.75 (0.57-0.98) for stillbirth. The best-performing ML models showed no significant difference in AUC compared with the best logistic regression models for predicting PTB (p = 0.07) or SGA (p = 0.31), but a higher AUC for stillbirth prediction (p = 0.03). No specific or universally strong predictors were identified across models. Almost all studies were at high risk of bias, and potential publication bias was observed. Interpretation: ML prediction models showed moderate performance, were at high risk of bias and did not always outperform traditional statistical methods. To improve perinatal outcome prediction, methodological improvements and better predictors are needed. Funding: Chinese Academy of Medical Sciences Innovation Fund for Medical Science, National Institute for Health and Care Research Senior Investigator Award.