Current methods for assessing risk of postpartum hemorrhage (PPH) rely on traditional statistical models or expert judgment. The study team believes machine learning could enhance PPH prediction by enabling more complex, data-driven models.
Objective
The research aimed to advance prediction models for PPH and evaluate how machine learning compares to traditional statistical approaches.
Methods
The researchers developed predictive models using the Consortium for Safe Labor data set, which included data from 12 U.S. hospitals covering the years 2002-2008. The study focused on transfusions of blood products or cases of PPH defined by an estimated blood loss of 1000 mL or more. For a secondary outcome, transfusions of any blood products were tracked. The models included 50 characteristics related to pregnancy, childbirth, and hospital factors. Several machine learning methods, including logistic regression, support vector machines, random forest, multilayer perceptron, and gradient boosting (GB), were evaluated. Performance was compared using the area under the receiver operating characteristic curve (ROC-AUC) and the area under the precision/recall curve (PR-AUC).
Results
The study analyzed 228,438 births, finding that 3.1% (5,760) involved PPH, 2.8% (5,170) required a transfusion, and 5.6% (10,344) met the criteria for both PPH and transfusion. Models based on both antepartum and intrapartum features showed the highest positive predictive values, with the GB machine learning model outperforming others (ROC-AUC=0.833, 95% CI 0.828-0.838; PR-AUC=0.210, 95% CI 0.201-0.220). Key predictive factors included delivery mode, oxytocin dose during labor, use of intrapartum tocolytics, presence of anesthesia support, and hospital type.
Conclusions
Machine learning models demonstrated superior discriminability over traditional logistic regression in predicting PPH risk. However, the study notes limitations with the Consortium for Safe Labor data set due to subgroup variability, which affects model accuracy and generalizability.
References:
Homa Khorrami Ahmadzia, Alexa C Dzienny, Mike Bopf, Jaclyn M Phillips, Jerome Jeffrey Federspiel, Richard Amdur, Madeline Murguia Rice, Laritza Rodriguez. *JMIR Bioinform Biotech* 2024 (Feb 05); 5:e52059
