(265) Finding the Forest in the TreeScan: A Big Data Driven Approach to Screening Drug Use in Pregnancy

Background: Data on drug exposure during pregnancy and infant outcomes is sparse. There is a need to screen pregnancy exposures to triage potential health impacts quickly and robustly, and semi-automated approaches can ensure more drugs are screened. Active comparator selection in high throughput studies is challenging to ensure rigor without sacrificing speed.

Objectives: To examine semi-automated approaches to identify appropriate controls in high throughput scenarios that simultaneously screen multiple exposures and outcomes. Appropriate controls using our high-throughput methods target a 1:5 ratio of exposed:referent pregnancies.

Methods: We identified pregnant persons aged 12 to 55 with a livebirth in Merative™ MarketScan® Research Databases data from 2010 to 2022. We tested three semi-automated active comparator selection strategies among unexposed patients: a) requiring a dispensing for a drug in a similar class or related classes, b) requiring a diagnosis for the indication in the U.S. clinical classification system, or c) either. We tested the strategies with four example drugs: two for chronic conditions (buspirone and montelukast) and two for acute (terconazole and clindamycin). We compared these restricted, unexposed pregnant populations to all unexposed pregnancies when evaluating potential comparators. We evaluated the three approaches by comparing propensity score (PS) distribution, overlap, and sample size retention with balanced characteristics in a traditional PS model adjusted for 24 characteristics and a high-dimensional PS model adjusted for 224 characteristics.

Results: We found no single approach for active comparator selection that resulted in the desired overlap and sample size ratio in our small pilot. The unrestricted scenario resulted in reference populations too large and dissimilar to the exposed, leading to ill-fitting PS models with poor overlap. In clindamycin, restricting to the dispensing requirement, but not the diagnosis requirement, retained the sample size ratio while providing good overlap. In buspirone, the opposite was true. In terconazole, both individual restriction approaches improved the overlap, but did not retain the target sample size ratio. For all three drugs, the restriction to either performed comparably to the preferred individual restriction. The montelukast model was bimodal, and semi-automated approaches resulted in poor fit in all scenarios.

Conclusions: Of the three approaches, requiring unexposed pregnant patients to have either a diagnosis or a dispensing is the most general strategy to retain sample size with reasonable PS model fit. Drugs with multiple indications may not be as amenable to this approach and may require more specific attention to create a comparable pregnant population.