OC09.06: Building Gestational Age Independent Features for First Trimester Viability Prediction

A key predictor of first trimester miscarriage is the discrepancy between observed and expected fetal development. The later requires a certain date of last menstruation period (LMP). However, a significant proportion of women are unsure of their LMP. This lack of gestational age (GA) absolute reference makes those patients ineligible for first trimester predictive model where GA adjustment is required. In this study, we assessed the use of relative references to build GA independent features. We studied 4 different predictors - MSD, CRL, progesterone and bhCG - based on data from 105 women with first trimester miscarriage and 1022 with viable pregnancies, prospectively recruited from a single centre from March 2014- May 2019. Using the subset of viable pregnancies with certain date of LMP, we first modelled the relationship to GA by LMP, and assessed the GA correlation. Estimated GA were then derived for all pregnancies. Doing so, the transformed predictor now demonstrates a constant growth rate in viable pregnancies, making its rate GA-independent. We computed the ROC AUC with 95% CI as predictive performance of those GA independent features. Progesterone was time-independent in the viable cohort and could be used without reference to GA with decent performance (AUC 0.799; 95% CI 0.730-0.869). MSD was linear in terms of GA, no specific transformation were needed to achieve constant rate (AUC 0.696; 95% CI 0.668-0.718.) For early pregnancies, using empirical CRL growth rate was acceptable without correction for the quadratic behaviour (AUC 0.852; 95% CI 0.813-0.891). Due to the large variability in bhCG concentration, obtaining accurate GA was more challenging, preventing the use of a growth rate method. However, using a combination of bhCG absolute and rate values, we were able to create a proxy variable with decent correlation to GA by LMP (R2 : 0.692), independent of the ultrasound-based GA. We have demonstrated that it is possible to derive meaningful predictor in the absence (correct) date of LMP.