(230) Design Considerations for Using the Tree-Based Scan Statistic in Surveillance of Maternal Outcomes Following Medication Use During Pregnancy

Friday, August 26, 2022

8:00 AM – 6:00 PM CEST

Location: Hall C

Publication Number: 561

Presenting Author(s)

Thuy N. Thai, MPH
– University of Florida Department of Pharmaceutical Outcomes and Policy, Gainesville, Florida, United States

Co-Author(s)

Almut Winterstein, PhD, FISPE
– Distinguished Professor, University of Florida, Gainesville, Florida, United States
Elizabeth A. Suarez
– BWH Div of Pharmacoepidemiology and Pharmacoeconomics, Boston, Massachusetts, United States
Danijela Stojanovic, PharmD, PhD
– U.S. Food and Drug Administration, Silver Spring, MD, United States
Jane Liedtka, MD, Board certified dermatologist
– Medical Officer, US Food and Drug Administration, Silver Spring, MD, Silver Spring, MD, United States
Abby F. Anderson, M.D.
– Medical Officer, US Food and Drug Administration, Silver Spring, MD, Silver Spring, MD, United States
Di Zhang, PhD
– US Food and Drug Addministration, MD, United States
Yueqin Zhao, PhD
– US Food and Drug Administration, United States
Monica A. Munoz, PharmD, PhD
– US Food and Drug Administration, Silver Spring, Maryland, United States
Wei Liu
– US Food and Drug Administration, Silver Spring, MD, Silver Spring, MD, United States
Inna Dashevsky
– Department of Population Medicine, Harvard Pilgrim Health Care Institute and Harvard Medical School, Boston, MA, Boston, MA, United States
Talia Menzin
– Department of Population Medicine, Harvard Pilgrim Health Care Institute and Harvard Medical School, Boston, MA, Boston, MA, United States
Elizabeth Messenger-Jones, MSPH
– Department of Population Medicine, Harvard Medical School & Harvard Pilgrim Health Care Institute, Boston, MA, United States
Elizabeth Siranosian
– Harvard Pilgrim Healthcare Institute, United States
Judith C. Maro, PhD
– Harvard Pilgrim Health Care Institute, Boston, Massachusetts, United States

Background: TreeScan is a data-mining method to screen for thousands of adverse events. Propensity scores (PS) can be used for confounding control, but must be simplified to facilitate broad outcome screening.

Objectives: To assess TreeScan’s signal identification performance to identify signals for maternal and obstetric adverse outcomes from 20 weeks of gestation to 30 days after delivery among patients with livebirths exposed to oral macrolide compared to beta-lactam antibiotics.

Methods: We used the IBM® MarketScan® Research Database from Oct 2015 to Feb 2020 to identify patients with livebirth delivery and exposure to study drugs. We excluded patients diagnosed with preterm premature rupture of membranes, exposed to other antibiotic types and routes or teratogenic drugs. We modeled a PS based on infection indications a week prior to and including index, gestational age at index, and covariates developed for pregnant populations. We used PS stratification with quartiles or deciles to balance covariates between exposure groups. Outcome-specific risk factors were not included in the PS due to the diverse outcomes screened, which included all ICD-10-CM codes related to pregnancy, childbirth and the puerperium (O00-O9A). We used conditional Poisson TreeScan models that adjusted for multiple scenarios to identify any outcome that occurred disproportionately with an alert threshold at p ≤0.05.

Results: Our cohort included 13,460 macrolide and 19,192 beta-lactam users. Indications for antibiotic use only balanced when using PS deciles. We could not identify indications for 67% macrolide and 48% beta-lactam users. Unsurprisingly, preliminary results suggested that several alerts meeting our threshold (e.g. infection of kidney following delivery) were due to confounding by indication. We also noted alerts indicating infection complications, which were unlikely safety issues, suggesting that indication-related outcomes could be removed from the outcome tree to reduce noise.

Conclusions: Signal identification assessments designed to screen thousands of outcomes are unlikely to achieve optimal confounding control. Increased efforts are needed to reduce bias due to unmeasured confounding, especially given different confounders could be associated with different outcomes. Additionally, the outcome tree may need to be tailored to eliminate issues of protopathic bias though retention of complications that imply ineffective treatment may be desirable. Screening analyses should anticipate and minimize noise but should also tolerate potential false alerts to facilitate full capture of safety issues when prioritizing outcomes for targeted pharmacoepidemiology studies.