Machine Learning Prediction Models for Preeclampsia: Systematic Review and Meta-Analysis.

BackgroundPreeclampsia is a severe hypertensive disorder with rising global prevalence. While machine learning (ML) models for predicting preeclampsia are increasingly published, existing evidence shows high heterogeneity, and the distinction between internal performance and external transferability remains unclear.ObjectiveThis study aims to evaluate the performance of ML models in predicting preeclampsia through a systematic review and meta-analysis, while also exploring their potential clinical application value, in order to specifically enhance the quality of future research and the predictive capability of the models.MethodsFollowing PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines and PROSPERO registration, we searched PubMed, Web of Science, IEEE Xplore, and CNKI (China National Knowledge Infrastructure) for studies published through February 2025. We included studies using ML to predict preeclampsia in pregnant women. Bias was assessed using PROBAST (Prediction model Risk of Bias Assessment Tool). We calculated summary estimates using random-effects models and, crucially, computed 95% prediction intervals (PIs) to estimate performance in future clinical settings. Subgroup and meta-regression analyses were conducted to explore heterogeneity.ResultsIn total, 26 studies comprising 31 ML models were included. While the pooled area under the receiver operating characteristic curve was high at 0.91 (95% CI 0.87-0.92), extreme heterogeneity was observed (I2>99%). The 95% PI for sensitivity was wide (0.32-0.96), indicating that in some external settings, sensitivity could drop to 32%. Only 6 studies conducted external validation; in these, the pooled sensitivity decreased to 0.68, with a PI of 0.25-0.94.Subgroup analysis suggested that models incorporating laboratory biomarkers and neural networks outperformed others, though CIs overlapped.ConclusionsCurrent evidence suggests that a high area under the curve in ML models is more likely to reflect the "performance" of the model on the internal development dataset rather than its universal "effectiveness" and clinical utility in independent, diverse populations. The apparent performance exhibits significant contextual dependence. Future studies should conduct multicenter, prospective external validation and recalibration research to enhance transferability and reliability.Trial registrationPROSPERO CRD420251005830;https://www.crd.york.ac.uk/PROSPERO/view/CRD420251005830.