Using predictive modeling to identify pregnant members to increase enrollment in a maternity disease management program

Many adverse pregnancy outcomes can be avoided through early and effective prenatal care. The goal of the maternity Disease Management (DM) program of a large southeastern managed care organization (MCO) is to enroll as many pregnant women as they can identify. However, only 23% of known pregnant women were actually being enrolled, presumably due to low identification. We extracted 21 months of medical and pharmacy claims data for two groups of members: 1) those known to be pregnant and 2) those believed not to be pregnant. Specifically, we examined claims for evidence of prenatal vitamins, ultrasounds, pregnancy tests, and other diagnoses and procedure codes suggesting a member may be pregnant. Using data mining techniques, we created a predictive ensemble model which consisted of a classification & regression tree (CART) combined with a logistic regression model. Every week, members are scored using the model, and those with at least a 70% probability of being pregnant are contacted by DM staff and encouraged to participate in the program. On a validation dataset of 54,220 members, we correctly predicted 96% of the members to be pregnant. Since the inception of the model, enrollment in the maternity disease management program has increased 59%. We are able to identify 87% of known pregnant members. Approximately 85% of enrollment is now coming from the weekly report. Identifying pregnant members was a major problem plaguing our maternity DM program. As a result, 59% more members—an average of approximately 200 members a month—are participating in a program that promotes a healthier pregnancy and birth. Initial analyses of pregnancy outcomes suggest that although adverse outcomes have been increasing over time, since the inception of the weekly report, the trend not only slowed, but is now actually on the decline.