Predictors of preterm birth in birth certificate data.

Demographic factors have been shown to be moderate predictors of preterm birth in prior studies which used hospital databases and epidemiologic sample surveys. This retrospective study used de-identified 2003 North Carolina birth certificate data (n=73,040) and replicated the statistical and computational methods used in a prior study of an academic medical center's data warehouse. Receiver Operating Characteristics (ROC) curves were used to compare results across methods. Due to differences between the data collected for birth certificates and the original clinical database, five of the seven demographic variables in the clinical database model were available for model testing (maternal age, marital status, race/ethnicity, education and county). Even with a reduced model, multiple methods of statistical and computational modeling supported the earlier findings of demographic predictors for preterm birth. The reduced model AUC results were acceptable (logistic regression = 0.605, neural networks = 0.57, SVM = 0.57, Bayesian classifiers = 0.59, and CART = 0.56), but lower than in the prior study as might be expected for a reduced model. On a population level, these results support a prior demographic predictor preterm birth model generated from a clinical database and the use of computational methods for model formation. Additional testing for stronger predictor models within birth certificate data is suggested as birth certificate data is a parsimonious population dataset already routinely collected.