Perinatal & Pediatric
Can we use classification and regression trees to compare cesarean delivery rates across populations? Ruby Barnard-Mayers* Ruby Barnard-Mayers Martha Werler
Background: When researchers estimate the effect of delivery method on birth outcomes, cesarean deliveries (CDs) are often compared to a reference group of vaginal delivery. However, this reference group does not always satisfy the assumptions necessary to use vaginal delivery as a counterfactual to CD (e.g. cesarean sections due to breech births or twin gestation). The Robson Classification (or 10-group classification) is one of the most common systems used for the purpose of comparing cesarean section rates across populations (based on pregnancy characteristics) and is recommended by the WHO. However, there has not been much testing of the Robson Classification System for the purposes of epidemiologic analyses.
Objective: The purpose of this analysis is to test machine learning techniques by conducting classification and regression tree analysis (CART) and random forests to compare to the Robson classification system.
Method: Data for this analysis comes from the Pregnancy and Early Life Longitudinal (PELL) data system, comprising all birth certificate records in Massachusetts from 2011 to 2018. We used classification and regression trees (CART) and random forests to create groupings of individuals, using the same set of variables as used in the Robson classification system. We used the mean Gini decrease to calculate variable importance.
Results: Eight leaves (final groups) resulted from the CART analysis, two fewer than the Robson classification system. Trial of labor was the most important variable and preterm delivery was the least important. In the Robson classification system, the first branch for classification is plurality. Rates of CD varied across groups for each system. The highest contributor to CD rates were births with no trial of labor for the CART analysis and multiparous births to people with a prior CD for the Robson System.
Conclusion: The CART created different groups than the Robson system. These groupings may be better for causal analysis of CD.
