256: Machine learning-based prediction models for postpartum hemorrhage

Abstract

Given the burden of maternal morbidity due to postpartum hemorrhage (PPH), a prediction model is needed. Our objective was to construct and validate models using machine learning that predict a woman’s risk of PPH using clinical data available at the time of admission for labor. Estimated blood loss was reported on births from 10 of 12 hospitals in the U.S. Consortium for Safe Labor Study from 2002-2008. PPH was defined as an EBL ≥1000 mL. Fifty-five risk factors were considered. Logistic and lasso regression, random forest, and Extreme Gradient Boosting models were derived to predict PPH. Temporal external validation was performed from the first phase (2002-2006) for model derivation and the second phase (2007-2008) for model validation. Hospital-specific (geographic) and temporal validation were combined by using each hospital once as a validation sample, with the remaining hospitals used for model derivation during the first phase. Model performance was measured by c-statistics, calibration, and decision curves. Of the 152,279 births, 7,279 (4.8%, 95% CI 4.7, 4.9) had a PPH. All models had good to excellent discrimination during temporal validation (c-statistic [95% CI]: logistic regression 0.87 [0.86,0.87], lasso 0.87 [0.86,0.88], random forest 0.92 [0.91,0.92], Extreme Gradient Boosting 0.93 [0.92,0.93]) and hospital with temporal validation (c-statistic: logistic regression 0.87 [0.86,0.87], lasso 0.87 [0.86,0.88], random forest 0.92 [0.91,0.92] and Extreme Gradient Boosting 0.93 [0.92,0.94]). The Extreme Gradient Boosting model (Figure 1) had the best performance across time and hospitals (Figure 2a). All models provided superior net benefit when clinical decision thresholds were between 0 to 80% predicted risk with the boosting model providing the largest net benefit (Figure 2b). These prediction models using machine leaning had excellent discrimination ability to predict PPH. Clinical application of these models could allow providers to be clinically prepared and facilitate triage of women at high-risk of PPH to the appropriate level of maternity care.View Large Image Figure ViewerDownload Hi-res image Download (PPT)