Development and validation of a novel prediction model for hypertensive disorders of pregnancy based on maternal cardiovascular function and placental blood flow metrics at 22 to 24 gestational weeks using machine learning

Abstract

Abstract Hypertensive disorders of pregnancy (HDP) remain a leading cause of maternal and perinatal morbidity worldwide, and current screening strategies have limited predictive value in low-risk populations especially for late-onset HDP. This study aimed to develop and internally validate an interpretable machine-learning model for predicting HDP using noninvasive parameters of maternal systemic hemodynamics, endothelial function, and utero-placental and feto-placental blood flow measured in mid-pregnancy. In this cross-sectional cohort, 577 normotensive women underwent impedance cardiography and Doppler ultrasonography at 22 + 0 to 23 + 6 weeks’ gestation. The incidence of HDP was 16.6% (96/577) most (87.5%) occurring at term (≥ 37 weeks), including 73 cases of gestational hypertension (12.6%) and 23 cases of pre-eclampsia (4.0%). An optimized predictive model with seven physiological features was developed using Extreme Gradient Boosting (XGBoost). Hyperparameters were optimized using stratified 10-fold cross-validation maximizing average Precision Recall Area Under the Curve (PR–AUC), and the decision threshold was selected by maximizing the geometric mean of sensitivity and specificity on a separate validation set. On an independent test set ( n  = 58; 10 HDP), the model achieved an Area Under the Receiver Operating Curve (ROC–AUC) of 0.82 (95% CI 0.65–0.95) and a PR–AUC of 0.57 (95% CI 0.26–0.84). At the optimized operating point, sensitivity was 70% (95% CI 0.40–1.00), specificity 79% (95% CI 0.67–0.90), precision 41% (95% CI 0.18–0.65), and negative predictive value 93% (95% CI 0.84–1.00). This interpretable, non-invasive mid-gestation model demonstrates strong discrimination and excellent negative predictive value, supporting its integration into routine second-trimester screening for risk stratification without reliance on biochemical markers.