The scale of potential exposure is staggering. In an Australian cohort of more than 1.6 million women, 90.5 percent of those who started GLP-1 therapy in 2022 did not have type 2 diabetes, and 2.2 percent became pregnant within six months of their first prescription. Clinical development programs covering 32,598 women documented 164 unintended pregnancies. Unintended exposure is therefore not a hypothetical edge case but a foreseeable event on a population scale, and clinicians are already fielding calls from patients who discovered they were pregnant while taking semaglutide or tirzepatide.

Interpreting pregnancy safety data for this class is fiendishly complicated by confounding by indication. Women with type 2 diabetes already carry a higher baseline risk of major congenital malformations than the general population, and exposed women often have substantially higher body mass index than insulin-treated or unexposed comparators. Previous syntheses have reached discordant conclusions, ranging from lower maternal or preterm-birth risks to renal or urinary malformation signals driven largely by one unadjusted cohort. The new review was designed specifically to disentangle independent replication from repeated pooling of the same evidence.

The methodological innovation is a time-window-stratified framework. Exposure was classified into four clinically defined windows: distant pre-pregnancy exposure ending more than 90 days before conception, recent pre-pregnancy exposure ending 1 to 89 days before conception, periconceptional and first-trimester exposure from day −90 through the end of the first trimester, and continuation beyond the first trimester. The non-interchangeability of the pre-pregnancy windows was demonstrated within a single data source, where the gestational diabetes estimate rose from a risk ratio of 0.97 to 1.54 depending on discontinuation timing. Trimester-unspecified gestational exposure was coded separately rather than forced into a single interval.

The searches identified 2,312 records, and after duplicate removal and full-text assessment, 14 comparative data sources were included, 13 of which were quantitatively eligible. Eight contributed to at least one pooled estimate. Reported exposed-group sizes summed to at least 46,341 pregnancies, though the authors caution that this should not be interpreted as unique women because data sources may overlap. Each independent study or data source contributed one prespecified primary comparator, selected by a hierarchy that prioritized diabetes-based active comparators, then same-drug discontinuers, then unexposed or general-population controls. Odds ratios were pooled only with odds ratios, and adjusted risk ratios were retained separately as study-level estimates rather than relabeled.

The headline result concerns major congenital malformations. In the timing-specific primary analysis, four independent studies with verified periconceptional or first-trimester exposure yielded an odds ratio of 0.99 with a 95 percent confidence interval of 0.77 to 1.27 and zero heterogeneity. A secondary pregnancy-period sensitivity analysis that added one large unadjusted cohort with unresolved trimester timing produced an apparently precise estimate of 1.00 (0.94 to 1.06), but that cohort alone contributed 93.5 percent of the weight. The authors explicitly refuse to interpret the near-null secondary estimate as a first-trimester risk figure, because the dominant study lacked covariate adjustment and its exposure window could not be resolved to a trimester.

Comparator choice materially changed the estimates, illustrating why clinically incompatible control groups should not be collapsed into a single causal interpretation. One study reported an adjusted risk ratio of 0.95 versus insulin and a crude odds ratio of 1.02. Another reported adjusted odds ratios of 0.98 versus a diabetes reference and 0.54 versus an overweight or obese reference. A third yielded 0.64 versus insulin but 1.24 versus an unexposed population. A fourth yielded 3.52 versus a drug-free population but 0.74 versus prepregnancy discontinuers. Alternative comparator selections produced a pregnancy-period odds ratio of 1.14 (0.64 to 2.02), and without the dominant cohort, 1.37 (0.42 to 4.49). These are clinically different contrasts, not one uniform causal effect.

Pooled preterm birth, small-for-gestational-age birth, and stillbirth analyses showed no statistically significant differences. Preterm birth yielded an odds ratio of 1.16 (0.47 to 2.82) with substantial heterogeneity of 73.6 percent. Small-for-gestational-age birth yielded 0.99 (0.77 to 1.26), though removal of the dominant study, which contributed 88.1 percent of the weight, shifted the estimate to 0.93 (0.39 to 2.24). Stillbirth, pooled across three studies, produced an odds ratio of 0.93 with a confidence interval of 0.04 to 22.9, an interval so wide that it is essentially uninformative. Preeclampsia was reported by two independent studies and was not pooled; one yielded an unadjusted odds ratio of 1.12 (1.08 to 1.16), while the other produced a within-study combined estimate of 1.08 (0.83 to 1.42). The signal was considered exploratory because it was not independently reproduced and residual confounding remained.

The renal and urinary malformation signal that has circulated through recent meta-analyses traces to a single dominant unadjusted cohort, which reported an unadjusted odds ratio of 1.25 (1.05 to 1.49). Rounding scenarios ranged from 1.00 to 1.57, with three of nine intervals including the null. Similar estimates in previous syntheses were driven by the same cohort and therefore do not constitute independent replication. The authors describe this as repetition of the same estimate across meta-analyses being mistaken for biological replication, a distinction that has profound implications for how safety signals propagate through the literature and into clinical guidance.

Certainty was graded as very low for all key outcomes using GRADE, reflecting observational design, residual confounding, comparator and timing indirectness, sparse events, dominance by high-risk unadjusted data, and imprecision. The authors conclude that current evidence does not demonstrate a clear or consistent increase in major congenital malformation risk, but that very low certainty precludes conclusions of safety. Inadvertent exposure alone is not evidence of teratogenicity and is not, by itself, an indication for pregnancy termination. Conversely, lack of statistical significance should not be equated with established safety. Counseling should incorporate exposure timing, agent, diabetes and obesity severity, concomitant therapy, and patient preferences, while prospective, timing-specific, comparator-aligned, and agent-specific studies are generated. Agent-specific evidence remains sparse, with extractable comparative tirzepatide-specific pregnancy outcomes unavailable, and class-level estimates should not be interpreted as establishing safety for any individual molecule, given differences in structure, half-life, placental transfer, indication, and GIP co-agonism.

Subject of Research: Pregnancy safety of GLP-1 and dual GLP-1/GIP receptor agonists under the statistical microscope

Article Title: Pregnancy and perinatal outcomes following exposure to GLP-1 and dual GLP-1/GIP receptor agonists: a time-window-stratified systematic review and meta-analysis

Article References: Mavigök, E., & Mavigök, A. N. (2026). Pregnancy and perinatal outcomes following exposure to GLP-1 and dual GLP-1/GIP receptor agonists: a time-window-stratified systematic review and meta-analysis. Frontiers in Pharmacology, 17, Article 1994366. https://doi.org/10.3389/fphar.2026.1994366

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1994366

Keywords: Pregnancy, perinatal, outcomes, following, exposure, GLP-1, dual, receptor, agonists, time-window-stratified, systematic, review