Developing a predictive model for perinatal morbidity among small for gestational age infants.

BACKGROUND While neonates with birth weight 48 h, NEC, sepsis, RDS, mechanical ventilation, retinopathy of prematurity, seizures, grade 3 or 4 IVH, stillbirth, or death before discharge. Cases were SGA neonates that experienced the primary outcome, and controls were SGA neonates that did not. Maternal factors for potential inclusion in predictive modeling were drawn from a broad list of variables collected as part of the NuMoM2B study, including demographic, anthropometric, clinical, ultrasound, social/behavioral, dietary, and psychological variables. Characteristics that were different in bivariate analysis between cases and controls then underwent further evaluation and refinement. Continuous and multi-category variables were assessed using multiple approaches, including as continuous variables, using standard categories (such as for BMI) as well as empirically-derived cut-points identified by receiver-operating characteristics methodology. The approach for each variable that resulted in the best performance was selected for use in modeling. After variable optimization, multivariable analysis was used to derive prediction models using factors known at mid-pregnancy (Model 1) and delivery (Model 2). RESULTS Of the original cohort, 865 were eligible and analyzed, with 134 (15.5%) experiencing the primary outcome. After bivariable and multivariable analysis, these variables were included in Model 1: BMI, stress level, diastolic blood pressure, narcotic use (all in 1st trimester), and uterine artery pulsatility index at 16-21 weeks. Model 2 added the following variables to Model 1: preterm delivery, preeclampsia, and suspected fetal growth restriction. When models 1 and 2 were empirically tested and compared to predicted performance to demonstrate calibration, observed morbidity rates approximately followed expected rates within deciles. Models 1 and 2 had respective areas under the receiver-operating characteristic curve of 0.72 (95% CI 0.67-0.76) and 0.84 (0.80-0.88), to predict the composite morbidity. CONCLUSION Using a deeply phenotyped cohort of nulliparous women, we created two models with the moderate-good prediction of perinatal morbidity among SGA neonates. TRIAL REGISTRATION clinicaltrials.gov ID: NCT01322529.