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

Area of Science:

• Medical Informatics
• Biostatistics
• Obstetrics

Background:

• Preeclampsia is a severe hypertensive disorder with increasing global incidence.
• Existing machine learning (ML) models for preeclampsia prediction exhibit high heterogeneity and unclear external transferability.
• The distinction between internal model performance and real-world clinical effectiveness remains a critical gap.

Purpose of the Study:

• To systematically review and meta-analyze the performance of ML models in predicting preeclampsia.
• To evaluate the clinical application value of these ML models.
• To identify factors influencing model performance and guide future research quality.

Main Methods:

• Systematic review and meta-analysis following PRISMA guidelines.
• Searched major databases (PubMed, Web of Science, IEEE Xplore, CNKI) up to February 2025.
• Assessed bias using PROBAST, calculated summary estimates with random-effects models, and computed 95% prediction intervals (PIs).

Main Results:

• Included 26 studies with 31 ML models; pooled AUC was 0.91 (95% CI 0.87-0.92) but with extreme heterogeneity (I²>99%).
• Wide 95% PI for sensitivity (0.32-0.96) indicates potential performance drops in external settings.
• External validation (6 studies) showed lower pooled sensitivity (0.68, PI 0.25-0.94). Models with lab biomarkers and neural networks showed promise.

Conclusions:

• High internal AUC in ML models may not translate to universal clinical effectiveness.
• Model performance is highly context-dependent, emphasizing the need for external validation.
• Future research must prioritize multicenter, prospective external validation and recalibration for improved transferability and reliability.