Construction and comparison of preeclampsia prediction models using logistic regression and machine learning.

OBJECTIVE

Preeclampsia is a major cause of maternal and perinatal morbidity and mortality. Early identification of women at high risk is essential for timely prevention and improved outcomes.

METHODS

In this study, clinical data as well as ultrasound examination results from 404 pregnant women, among whom 45 suffered and 359 did not suffer from preeclampsia, were retrospectively analyzed. Independent predictors were found via multivariable logistic regression. With these predictors, four classification models involving logistic regression, support vector machine, random forests, and extreme gradient boosting are developed in this paper. The performance of all these four classification models is assessed via metrics such as area under receiver operating characteristic curves, calibration plots, and decision curves.

RESULTS

Maternal age of 35 years or more, assisted reproductive technology, history of eclampsia, history of fetal growth restriction, and two placental blood flow indices - flow index and vascular flow index, were independent predictors. In the internal validations, logistic regression was characterized by the most favorable calibrated method, based on 10-fold out-of-fold prediction, with an intercept of − 0.325, slope of 0.811, and Brier of 0.090, suggesting its suitability as a probability-based predictive tool, while random forest produced the highest AUC of 0.781.

CONCLUSIONS

Logistic regression appears most suitable for early screening as a probability-based risk tool due to its favorable calibration and interpretability, pending external validation and recalibration. These models quantify statistical associations between routinely collected predictors and preeclampsia in this retrospective cohort; they do not imply causality and require external validation before prospective clinical use.