(B-260) Development of a Pregnancy Loss Risk Score to Address Potential Confounders for Pregnancy Loss

Background: Accurately adjusting for confounding factors associated with pregnancy loss is vital in pregnancy safety studies. Existing obstetric risk scores focus on identifying high-risk pregnancies rather than pregnancy loss.

Objectives: Therefore, we aimed to develop a pregnancy loss score to adjust for confounders in studies assessing prenatal exposure to medications and the risk of pregnancy loss (PL).

Methods: We selected covariates using diagnosis codes reflecting conditions included in 4 previously published papers for general obstetric risk. To avoid adjustment for mediators or similar biases, we focused on the pre-conception period and removed any pregnancy-related complications (e.g., pre-eclampsia). We added a composite covariate indicating previous pregnancy complications and prior PL.
In Merative MarketScan Commercial Claims Data, we created a cohort of livebirth (LB), spontaneous abortion, and stillbirth pregnancies from 2005-2021 with continuous enrollment from 365 days before conception until 30 days after delivery. We excluded any LB with extreme prematurity (GA < 28 weeks). We modeled all covariates using logistic regression to derive the predicted probability (score) of PL. We tested the score’s ability to adjust for confounding by comparing the hazard of PL among pregnancies exposed to pregnancy-safe medications, including insulin, sertraline, and methyldopa, versus those unexposed without or with risk score adjustment. Adequate adjustment should yield no risk difference.

Results: In 2,234,133 eligible pregnancies with a mean age of 30.7 for LB and 32.2 for PL, the most common conditions were mental health disorders (14.8%) and pain (14.3%). Among 39 risk factors, the strongest predictors for pregnancy loss were hypertension (OR 1.27, 95% CI 1.25-1.29) and prior pregnancy loss (OR 1.21, 95% CI 1.19-1.23). Overall, the model had a c-statistic of 0.6. We identified 9,380 pregnancies exposed to insulin in the first trimester, 31,550 exposed to sertraline, and 8,849 exposed to methyldopa, all with significant crude HR (HR 1.33, 95% CI 1.28-1.38; 1.12, 1.10-1.14;1.38, 1.33-1.43). With pregnancy loss score adjustment, insulin and methyldopa had non-significant HRs while sertraline remained slightly significant (HR 1.06, 95% CI 1.04-1.08).

Conclusions: The PL score’s discriminant validity was moderate, which was expected due to sole focus on pre-pregnancy risk factors, However, our application examples demonstrate good ability to balance comparisons groups, which in some scenarios might need further augmentation with more granular disease-specific measures (e.g., disease severity).