(630) Development and validation of a clinical risk prediction model for small-for-gestational age infants

Small for gestational age (SGA) is defined as a birth weight of < 10th percentile for gestational age and is associated with adverse neonatal risks and increased resource utilization. Prenatal prediction of SGA as late as birth hospitalization admission could assist with risk stratification and improved allocation of resources on labor units.

Study Design:

Data from the National Vital Statistics System (NVSS) Birth Certificate and clinical data from an academic medical center (AMC) were utilized for the development, internal, and external validations. Non-anomalous, singleton births at >= 35 weeks were included for this analysis with the outcome of SGA based on neonatal weight using Fenton growth charts. A predictive model was developed using a stochastic gradient boosting algorithm adjusting for demographic (age, race/ethnicity, payer), obstetrical (delivery gestational age (GA), parity), clinical (body mass index, gestational weight gain (GWG), pregestational weight, chronic hypertension, pregestational/gestational diabetes, hypertensive disorders of pregnancy), and neonatal sex. Data from the 2019 NVSS was used to train and internally validate two models predicting SGA at birth. External validation was conducted on AMC and 2021 NVSS birth data. Measures of model performance are expressed as areas under the receiver operator curve (AUC) and with 95% confidence intervals (CI).

Results:

Model development was performed using 3.3 million NVSS births from 2019 and externally validated on 8,557 AMC births and 3.3 million NVSS births from 2021 with an SGA prevalence of 7.9%, 4.2%, and 8.8%, respectively. Internal validation yielded an AUC of 0.80 (95% CI: 0.80, 0.81) with delivery GA, pre-pregnancy weight, and GWG as the most important predictors. External validation yielded AUCs of 0.71 (95% CI: 0.68, 0.73) and 0.81 (95% CI: 0.80-0.81) on AMC and 2021 NVSS data, respectively.

Conclusion:

SGA can be satisfactorily predicted using this clinically deployable machine learning model which can be implemented up until birth admission to improve resource utilization and team preparation in anticipation of birth.