Machine-Learning Models May Accurately Predict Postpartum Hemorrhage Risk

A new study published in the Journal of Medical Internet Research shows that machine-learning (ML) models can effectively predict the risk of postpartum hemorrhage using data pulled from de-identified patient EMRs.

Research indicates that postpartum hemorrhage is the leading cause of maternal mortality worldwide. It places affected patients at an increased risk of comorbidities related to the heart, kidneys, and liver. Interventions to address postpartum hemorrhage also carry associated risks. Blood transfusions, one of the most common interventions, can cause anaphylactic reactions, lung injuries, antibody formation for future pregnancies, and risk of infections.

The adverse outcomes of postpartum hemorrhage have spurred an interest in using AI to predict which patients may be at risk for the condition. As with many medical conditions, early detection of postpartum hemorrhage is key to improving patient outcomes and mortality rates, the study authors noted.

To develop their prediction model, the researchers began by gathering retrospective data from 30,867 women aged 18 to 55 who underwent obstetric delivery at New York University Langone Health Tisch Hospital from July 1, 2013, to Oct. 31, 2018. They identified 2,179 cases of postpartum hemorrhage, defined in this study as blood loss of greater than or equal to 1000 mL at the time of delivery, regardless of delivery method.

For the models to generate predictions, 497 variables were extracted from the study cohort’s EMRs, including demographic information; obstetric, medical, surgical, and family history; vital signs; laboratory results; labor medication exposures; and delivery outcomes. The study cohort was split into smaller groups for model development and testing, with 70 percent assigned to the training cohort and the remaining split evenly into the validation and independent test cohorts.

Regression-, tree-, and kernel-based machine-learning methods were used to create the classification models. Some models were created using all collected data, while others were limited to data available prior to the second stage of labor or at the time of the decision to proceed with cesarean delivery. Additional models were built to make predictions based on the mode of delivery.

The tree-based methods achieved the best discrimination out of all the models. The model that included all collected data slightly outperformed the model relying on second-stage data. The highest-performing models achieved an accuracy of approximately 98 percent. Models stratified by delivery mode achieved good to excellent discrimination, but these lacked the sensitivity necessary for clinical applicability.

These results indicate that ML methods can be used to identify patients at risk for postpartum hemorrhage who may benefit from individualized preventative measures and that this work is worth pursuing because ML predictions may be superior to human risk assessment, the authors stated.

However, they also noted that more study in this area is necessary to validate the research findings and create successful models based on mode of delivery.