Labor
Poster Session 1
Reetam Ganguli, BS (he/him/his)
Elythea
San Jose, CA, United States
Stephen M. Wagner, MD
Assistant Professor
Beth Israel Deaconess Medical Center
Brookline, MA, United States
Postpartum hemorrhage (PPH) is a leading cause of maternal mortality in the United States. The widely adopted PPH risk assessment tool developed by the California Maternal Quality Care Collaborative only classifes 22% of severe PPH patients as high risk and includes variables only accessible during the labor period. We developd a machine learning (ML) model to identify patients at risk for severe PPH requiring transfusion using data available prior to delivery. We conducted a retrospective cohort study using the the Center for Disease Control’s (CDC) Vital Statistics System. All patients from the 2018-2020 years as well as PPH patients from the 2014, 2016, and 2017 years without missing data for the PPH outcome were included. Variables with >50% missing data were omitted and training variables consisted of demographics, medical history, and prenatal data. An extreme gradient boosting ML algorithm was validated on CDC data from the 2021 year. Youden’s J statistic value was calculated to set the classification threshold. The primary outcome was severe PPH requiring intrapartum maternal transfusion. 7,728,892 patients were used to train the model. The incidence of severe PPH necessitating a transfusion was 1.2% (n=89,819). The gradient boosting model was trained across 39 variables and achieved an AUC of 0.75 with a 78.2% accuracy, and 0.78 F1 score. The sensitivity was 0.65 after setting the optimal threshold at 0.46. The highest weighted factors for the model were body mass index and maternal age.
Early recognition could signal providers to schedule deliveries in higher-resourced hospital systems for patients in rural regions and enable earlier coordination with blood banks to prepare antigen matched blood prior to delivery.
Study Design:
Results:
Conclusion: ML methods can be utilized to identify patients at risk for severe PPH with nearly 3-fold greater high-risk sensitivity than the current PPH risk assessment system. Our model has the potential to improve care, avoid preventable PPH deaths, and reduce associated health services costs for PPH patients.