To develop and externally validate a machine learning model for predicting hypertensive disorders of pregnancy (HDP) using fetal aortic hemodynamics, and to interpret key predictors. This multicenter retrospective study included 800 pregnant women, divided into training (n = 350), internal testing (n = 150), and independent external validation (n = 300) cohorts. Twenty-eight candidate variables were screened via logistic regression. Eight machine learning algorithms were trained using fivefold cross-validation. Model discrimination, calibration, and clinical utility were assessed via AUC, calibration plots, and decision curve analysis. SHapley Additive exPlanations (SHAP) evaluated feature interpretability. Six fetal aortic parameters emerged as independent predictors: systolic blood pressure, mean arterial pressure, peak systolic velocity, resistance index, pulsatility index, and blood flow volume. XGBoost significantly outperformed other algorithms, achieving an AUC of 0.825 internally and 0.805 in external validation, with good calibration and favorable clinical net benefit. SHAP identified fetal aortic resistance and pulsatility indices as the most influential predictors; higher resistance and pressure increased predicted HDP risk, whereas greater blood flow volume was protective. This interpretable XGBoost model provides a non-invasive and generalizable tool for predicting HDP risk based on fetal aortic hemodynamics, reflecting fetal circulatory adaptation to placental dysfunction.
ZHENG et al. (Fri,) studied this question.