Risk patterns in Medicaid-enrolled pregnant women: Using classification trees to predict poor birth outcomes

Among many health policy makers, increasing limitations on public funding for enhanced prenatal services have been coupled with an awareness that neither budgets nor best clinical and community practice knowledge can support a “one-size-fits-all” solution for a diverse population of pregnant women. Crafting more strategic interventions requires a refined sense of the ways in which multiple factors interact to produce poor birth outcomes and an enhanced ability to predict these outcomes. With advances in computational power and the proliferation of larger data warehouses in fields ranging from health care to business sectors, statisticians and computer scientists have moved forward new methods of analysis and modeling techniques focused on the specific problems of and opportunities in extracting new learning from very large datasets with many records and fields. One family of these models – classification – is specifically geared toward identifying what factors distinguish members of one class (e.g., women who experience a poor birth outcome) from another (e.g., women who have a normal birth outcome). These classification methods can have several potential advantages above and beyond traditional statistical methods such as regression modeling. First, they focus on producing the most accurate algorithms possible to predict new cases – often producing substantially better predictive performance than traditional regression techniques. Second, these models are typically able to discover and take in to account complex interactions and subtleties in how risk factors impact the outcomes of interest. We present findings and implications from classification tree models developed to identify risk classifiers and pathways for poor birth outcomes in Medicaid-insured pregnant women, using data from Michigan's linked maternal-child Medicaid Data Warehouse for birth cohorts from 2004-2006. We developed classification tree models to predict low birth weight, preterm delivery, and NICU admissions given various behavioral, demographic, psychosocial, and chronic disease factors. These models correctly classify women who experienced poor birth outcomes versus those who had normal birth outcomes with notable and encouraging levels of accuracy. Of particular interest, we identify distinct risk pathways for Black and non-Black women, suggesting that more specific attention to different risk factors may help target more powerful interventions for these subpopulations. Finally, we present our work exploring the potential of using these models to produce risk scores that can be subsequently incorporated into multilevel models of birth outcomes. Such models illuminate county-level differences and the potential impact of enhanced prenatal care programs on birth outcomes. Implications for policy and practice are discussed.