Integrating Bleeding Velocity and Clinical Indicators into Machine Learning Models for Predicting Postpartum Hemorrhage: Development and External Validation Study

Objective To develop and externally validate predictive models for early detection of postpartum hemorrhage by combining clinical, hemodynamic, and temporal blood loss indicators. Design Multicenter retrospective model development and external validation study. Setting A Brazilian tertiary referral hospital and pooled clinical trial and observational datasets from low- and middle-income countries. Population or Sample Model training used data from a Brazilian tertiary hospital cohort (n = 270) and external validation employed pooled datasets from randomized and observational studies (n = 2,379). Methods Predictive models were developed using maternal demographics, delivery characteristics, pre-delivery hemoglobin, blood loss at 30 minutes postpartum, and bleeding velocity. Statistical and machine learning approaches including logistic regression, random forest, and XGBoost were evaluated. Model discrimination and performance were assessed using area under the receiver operating characteristic curve (AUC), F1-score, sensitivity, and additional validation metrics. Main Outcome Measures Binary; occurrence of postpartum hemorrhage defined as blood loss ≥500 mL (yes/no), along with model discrimination, calibration, and clinical predictive performance. Results Early bleeding velocity and blood loss at 30 minutes were the strongest predictors of PPH. Generalized linear models achieved high discrimination (internal AUC ~ 0.98) and strong agreement. Machine learning models performed well but did not consistently outperform logistic regression. In external validation, simpler regression-based models demonstrated equal or superior discrimination (AUCs up to 0.996) and more stable calibration. Reduced predictor models retained good performance (external AUC ~0.89–0.94), while expanded models maintained robust performance (AUC ~0.92–0.98). Conclusions Integration of early hemodynamic and temporal bleeding measurements could substantially improve early prediction of postpartum hemorrhage. These findings support scalable, data-driven tools for timely intervention and reduction of maternal morbidity.