Methods/Statistics
Examining the standard practice of race adjustment in prenatal screening biomarkers: a first step in a larger investigation Kharlya Carpio* Kharlya Carpio Demia Tyler Hannu Koistinen Janet Catov Jennifer J. Adibi
Background: Standardizing prenatal screening biomarkers involves adjusting for gestational age and other sources of variability, including race, to minimize nuisance parameters and enhance accurate risk prediction for adverse outcomes. This can be achieved without the inclusion of race, revising a practice that perpetuates the harmful idea that race differences are biologic. The first step is to establish a correctly specified model that is not conditional on race. Serum levels of the hCGa biomarker are lower in Black vs. white women and are associated with SDOH and psychosocial stress. Methods: hCGa was measured in maternal serum at the 1st, 2nd, and 3rd trimesters, labor and delivery, and two years postpartum, in pregnant people who delivered between 2020-2021 and enrolled in the Magee-Womens Hospital Biobank (N = 58). Overall distributions were examined as crude, logarithmic, and squared transformations. Log-linear and exponential equation models were used to examine the relationship between hCGa collected pre- and postpartum, in which covariate combinations of maternal age, pre-pregnancy weight, and gestational age at time of blood draw were included. We examined standard Akaike Information Criterion (AIC) and a bias detection tool to evaluate if model properties were considered “fair” across racial groups (fairmodels R package). Results: Multiple imputation was used to account for participants with missing serum levels in different trimesters. Log transformation reduced skewness but biased the gestational age association. Including all variables in an exponential model yielded the highest AIC, indicating the most accurate shape. None of the models satisfied the fairness criteria. Conclusion: Applying transformations for analyzing biomarkers in pregnancy is essential. More empirical and methodologic work is needed to reconsider the impact of race-based algorithms on accurate risk prediction to avoid potential harm. This work is a first step in that process.