Risk Stratification for Postpartum Hemorrhage in Vaginal Delivery: A Nationwide Cohort Study of Predelivery and Intrapartum Risk Assessment

AIM: To develop and validate machine learning models for clinically applicable risk stratification of postpartum hemorrhage (PPH) in vaginal delivery, and to assess whether incorporating intrapartum information improves detection while maintaining a similar screening burden. METHODS: This nationwide retrospective cohort study included 1 224 990 women undergoing vaginal delivery at secondary and tertiary care institutions in Japan between 2014 and 2023. Data from 2014 to 2022 were used for model development, and 2023 was reserved for temporal external validation. Machine learning models were developed using routinely available predelivery clinical variables, with additional intrapartum information incorporated into an updated model. A predefined alert-rate threshold was selected in the development cohort to maximize sensitivity while maintaining approximately 300 high-risk classifications per 1000 deliveries. Performance was assessed using clinically interpretable metrics per 1000 deliveries. RESULTS: PPH (blood loss ≥ 1000 mL) occurred in 7.5% overall and 9.7% in the validation cohort. In temporal external validation, the predelivery model identified 52.2 PPH cases per 1000 deliveries (sensitivity 0.537) with 302.5 high-risk classifications. The intrapartum model identified 57.7 cases (sensitivity 0.596) with 320.1 classifications. Thus, intrapartum updating detected approximately 5-6 additional PPH cases per 1000 deliveries with a modest increase in screening burden. Discrimination improved from 0.676 to 0.703. CONCLUSIONS: Machine learning models using routinely available clinical information may support screening and preparedness for PPH in vaginal delivery. Incorporating intrapartum information modestly improved detection while maintaining a similar alert burden, supporting its potential clinical utility.