Developing a predictive model for perinatal morbidity among small for gestational age infants
Nathan R Blue,Amanda A Allshouse,William A Grobman,Robert C Day,David M Haas,Hyagriv N Simhan,Samuel Parry,George R Saade,Robert M Silver
DOI: https://doi.org/10.1080/14767058.2021.1980533
Abstract:Background: While neonates with birth weight <10th percentile are at increased risk of morbidity and mortality, most of these are constitutionally small and not at increased risk. There are no current strategies that reliably distinguish constitutionally small neonates from small neonates at the highest risk of morbidity, so additional tools for risk stratification are needed. Objective: Our objectives were to identify factors that are independently associated with perinatal morbidity among neonates with birth weight <10th percentile (small for gestational age, SGA) and to create predictive models of perinatal morbidity among SGA neonates based on the timing of information availability. Study design: This secondary analysis of the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be, was a nested case-control study. Participants were prospectively enrolled at eight U.S. centers, with data collection occurring at three standard time points during pregnancy and again after delivery. Our analysis included neonates with birth weights <10th percentile and excluded those with major congenital malformations or suspected or confirmed aneuploidy. The primary outcome was a composite of perinatal morbidity, defined as NICU admission >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.