Machine Learning Prediction of Persistent Postpartum Hypertension Following Hypertensive Disorders of Pregnancy: A Secondary Analysis of a Prospective Cohort

Background: Persistent postpartum hypertension (PPPH) is a clinically significant complication of hypertensive disorders of pregnancy (HDP) that is frequently underrecognized due to high rates of loss to follow-up after hospital discharge. Early pre-discharge risk stratification represents an important unmet clinical need. This study aimed to develop and internally validate machine learning (ML) models using exclusively pre-discharge data to predict the risk of PPPH in women with HDP before hospital discharge. Methods: A secondary analysis of a prospective cohort from the National Hospital of Obstetrics and Gynecology in Hanoi, Vietnam, was conducted. The primary outcome was PPPH, defined as a systolic blood pressure (SBP) ≥140 mmHg and/or a diastolic blood pressure (DBP) ≥90 mmHg, or ongoing antihypertensive therapy at 6 weeks postpartum. Feature extraction was strictly limited to the period from hospital admission through 72 hours postpartum. Three ML algorithms were developed: Ridge logistic regression (LR), Random Forest, and eXtreme Gradient Boosting (XGBoost). Internal validation was performed using a 500-repeat bootstrap framework with Synthetic Minority Over-sampling Technique (SMOTE) applied within each resampling loop. Model calibration was assessed using the Brier score and calibration plots, while clinical utility was evaluated via decision curve analysis (DCA). Model interpretability was achieved using Shapley Additive Explanations (SHAP). Results: PPPH was observed in 39 of 301 women (12.96%). The Random Forest model, incorporating 25 pre-discharge clinical and biochemical features, was selected as the final model based on its superior calibration (Brier score: 0.130; calibration intercept: −0.794; calibration slope: 0.759; observed-to-expected [O/E] ratio: 0.544) and area under the receiver operating characteristic curve (AUROC) of 0.638 (95% confidence interval [CI]: 0.632–0.643). To prioritize the identification of women at risk of PPPH, the decision threshold was set at 0.157, achieving a sensitivity of 80.2% and a negative predictive value (NPV) of 92.5%. DCA demonstrated that the Ridge LR model provided the greatest net clinical benefit across the 15%–25% threshold probability range. SHAP analysis identified alanine aminotransferase (ALT), blood urea nitrogen (BUN), and maternal age as the top influential predictors. Conclusions: An ML-based pre-discharge risk-stratification tool for PPPH was developed and internally validated. Although the current model demonstrated modest discrimination and requires external validation in larger multicenter cohorts, the Random Forest model produced well-calibrated risk estimates and a high NPV (92.5%), offering a proof-of-concept for individualized pre-discharge risk stratification and the targeted allocation of postpartum monitoring resources.