Temporal validation of machine learning models for pre-eclampsia prediction using routinely collected maternal...

Area of Science:

• Obstetrics and Gynecology
• Medical Informatics
• Public Health

Background:

• Pre-eclampsia (PE) is a major global cause of maternal and newborn mortality.
• Early detection and intervention are crucial for reducing PE complications.
• Existing risk prediction models require rigorous validation for clinical application.

Purpose of the Study:

• To temporally validate three existing pre-eclampsia (PE) prediction models: two machine learning (ML) models (XGBoost, random forest) and one logistic regression model.
• To compare the predictive performance of these validated models.
• To assess the clinical utility of PE prediction models.

Main Methods:

• Temporal validation of XGBoost, random forest, and logistic regression models using antenatal data from South-East Melbourne, Australia (July 2021-December 2022).
• Evaluation of discrimination using Area Under the Receiver-Operating Characteristic Curve (AUC) and calibration using slope.
• Comparison of AUCs via bootstrapping and assessment of clinical net benefits.

Main Results:

• The temporal dataset included 12,549 pregnancies, with a PE incidence of 3.43%.
• XGBoost (AUC 0.75) and logistic regression (AUC 0.76) showed similar discrimination; random forest had AUC 0.71.
• Logistic regression demonstrated excellent calibration (slope 1.02); XGBoost showed poorer calibration (slope 1.15).
• Temporally validated logistic regression and XGBoost models maintained stable discrimination, outperforming random forest.
• Logistic regression and XGBoost models offered improved clinical net benefits over default strategies at specific risk thresholds.

Conclusions:

• Logistic regression and XGBoost models demonstrate stable predictive performance upon temporal validation for pre-eclampsia.
• Neither machine learning model significantly outperformed the logistic regression model in PE prediction.
• Logistic regression is a viable option for routine, first-stage screening in a two-stage PE detection approach, identifying women for further assessment.