Mathematical Models to Predict Maternal Vitamin D (VitD) Status During Pregnancy and Lactation

Abstract

Abstract Objectives VitD status is important during pregnancy and lactation as maternal status is linked with fetal and neonatal outcomes: Maternal vitD status during pregnancy predicts neonatal vitD status; similarly, maternal vitD status during lactation predicts infant vitD status as mother is the sole source of vitD to the unsupplemented lactating infant. The best indicator of vitD status is measurement of circulating 25(OH)D concentration; however, results may be delayed and add to health care costs. A prediction model taking into account certain factors would assist health care providers in identifying those individuals most at risk for vitD deficiency. OBJ: Develop and test a prediction model for vitD status that utilizes information readily available to any health care provider in the primary care setting. Methods Three existing datasets (2 pregnancy, n = 405; and 1 lactation, n = 451; total n = 856 women) were combined for development of this prediction model. Variables included in the linear regression models to predict maternal total circulating 25(OH)D at baseline were maternal race/ethnicity, location/study site, maternal age, degree of skin pigmentation (forehead, forearm, inner arm, stomach and knee), body mass index (BMI), season (April-October and November-March) and vitD intake. To test the data, each woman was randomized to either the training dataset (80%) or to the testing dataset (20%). The fitted model was used to predict circulating 25(OH)D concentration in the testing dataset. Results In the full linear regression model, race, season, center, and age were significant predictors of the log transformed 25(OH)D (P-value < 0.05) and the prediction error was 12.97%. When only the significant predictors of the full model were applied in a submodel, the prediction error rose to 15.92%. The best model for prediction of 25(OHD) in this study was the full model. Conclusions The ability to predict vitD status from readily available sociodemographic and clinical parameters is possible with an accuracy of 87%. When only significant factors were utilized to predict vitD status, the accuracy fell slightly. Mathematical prediction models as shown in this example allow accurate prediction of particular health risk factors and will likely play a role in advancing nutritional and medical therapies. Funding Sources NCATS/NIH #UL1 TR001450