Abstract

Background: Observational data are often the only data available on drug use in pregnancy, and they can be slow and costly to produce. High throughput screening via active surveillance in large longitudinal databases might better identify potential adverse drug effects in pregnant persons that can be further evaluated using traditional causal inference frameworks.

Objectives: To design a rigorous and expeditious analytic approach to simultaneously screen for potential adverse infant outcomes among a wide variety of drug exposures.

Methods: We identified pregnant persons aged 12 to 55 with a live born infant in U.S. Medicaid data (2014-2019) and U.S. multi-site commercial claims data (2000-2023). We selected 50 drugs designated as potential teratogens by the Teratogen Information System (TERIS) and prioritized exposures of 4000+ pregnant persons. We compared first-trimester use of these drugs to an unexposed cohort with first-trimester exposure to comparator drugs for similar indications. We used conventional and high-dimensional propensity scores to adjust for many measured characteristics in the 90 days preceding pregnancy including concomitant exposures and utilization. After adjustment with stratification, we limited the analyses to 20 drugs with balanced characteristics and used Poisson tree-based scan statistics to identify imbalances in infant outcomes (e.g., malformation, seizure) in the 180 days post-delivery. With input from an expert panel of 8 clinicians and teratologists, we pre-defined ‘alerts’ as exposure-outcome pairs with a p-value ≤0.05 and a relative risk ≥2. Tree-based scan statistics use Monte Carlo simulation methods to control for inherent multiplicity in screening nearly 450 ICD-9/10 diagnostic codes.

Results: We observed 3.5M pregnancies, 35% in Medicaid. Exposed pregnancies ranged from 50,000+ for anti-infectives (e.g., first generation cephalosporins) to near the 4000 minimum for drugs used in assisted reproduction (e.g., leuprolide). We postponed evaluation of 30 lower prevalence drugs until we can add more years of Medicaid data. We identified 30 unique alerts among the 20 remaining drugs evaluated. Example alerts included congenital cardiac malformations following buspirone use, and small for gestational age outcomes following amlodipine use.

Conclusions: We have developed a rigorous, high throughput screening approach to monitor postmarket data on adverse infant outcomes following in utero exposure, seeking to detect any unusual increased frequency while minimizing false positives. Next, we will evaluate detected alerts for biologic plausibility with our clinician panel, and investigate potential outcome misclassification and other biases using patient profiles.