GLP-1 Exposure Around Conception: Birth Defect Meta-Analyses

Pooled observational studies have not found a clear increase in major birth defects after GLP-1 exposure around conception, but they do not show the drugs are safe in early pregnancy. The analysis most careful about timing combined four studies with verified periconceptional or first-trimester exposure and found an odds ratio of 0.99 (95% CI 0.77 to 1.27), then graded its certainty as very low^[1]. That leaves room for a modest increase or decrease in risk. The figures are most useful for someone counseling an accidental exposure, and they do not support continuing a GLP-1 while trying to conceive^[4][12].

What the 2025 and 2026 meta-analyses reported

Six recent reviews reached similar headline conclusions by different routes: Mavigök 2026^[1], Liu 2026^[2], Uysal 2026^[3], Hakim 2026^[4], D’Antonio 2026^[5] and Armstrong 2025^[6]. The number of exposed pregnancies swings from 8,325^[5] to 49,395^[4], partly because the reviews set different exposure windows, from three months before conception^[2] to six months before a positive test^[5].

One cohort supplies most of the apparent precision

The tightest confidence intervals in that table mostly reflect a single US study. Lewin 2026 used aggregate Epic Cosmos records from 143,593 patients with pregestational diabetes, including 35,083 who used a GLP-1 at some point in pregnancy. No individual-level data were available, so its comparisons were not adjusted for confounders^[7]. The full text of Mavigök 2026 identifies this as the cohort that contributed 93.5% of the weight in its secondary estimate of 1.00 (0.94 to 1.06), and warns against reading that figure as a first-trimester risk estimate because the study did not resolve exposure by trimester^[1].

Liu 2026 counts the same study as 40,929 of its 43,577 exposed pregnancies, a count of pregnancies rather than patients. Removing it left the point estimate at 1.02, but the interval widened from 0.96 to 1.08 to 0.81 to 1.30^[2]. A reader who sees a sample of more than 40,000 should therefore picture a much smaller, timing-verified evidence base beside one large unadjusted dataset. Multiple reviews also draw on overlapping studies, so several agreeing results are not several independent replications^[1].

The kidney and urinary signal

Two reviews reported a statistically significant association with one malformation category: renal defects in Hakim 2026 (OR 1.23; 95% CI 1.09 to 1.39)^[4] and urinary defects in Uysal 2026 (OR 1.24; 95% CI 1.05 to 1.47)^[3]. Hakim said the finding was largely driven by one large cohort with substantial baseline imbalances, and Uysal said it rested only on unadjusted data. Lewin’s raw figures show how small the absolute gap is: 0.5% versus 0.4% for urinary malformations and 0.7% versus 0.6% for genital malformations, while any malformation occurred in 3.3% of both groups^[7]. Mavigök 2026 concludes that these renal and urinary estimates trace back to that dominant cohort, so they do not count as independent replication^[1].

Confounding is a plausible explanation, though it is not proof that the drug plays no role. The cardiac picture is mixed as well. Armstrong 2025 reported fewer congenital heart defects with exposure^[6], while a French national study found no overall increase in its main analyses and raised the possibility of a small increase in cardiac malformations when stricter exposure definitions were used^[11]. A protective effect on the heart should not be inferred from observational data of this kind, and neither should a confirmed harm.

Studies built to reduce confounding

The design choice that matters most is the comparison group. A GLP-1 user differs from the average pregnant patient, so several cohorts compared against insulin users^[8][11][12], women with similar conditions^[9] or people who stopped the same drug^[10]. Their intervals, shown below, stay wide, and that imprecision is the main finding. Brown 2026 stands out because everyone in it had a GLP-1 dispensed in the 90 days before the last menstrual period and 41.1% had type 2 diabetes. Its authors said the malformation estimate was compatible with both no increase and a clinically relevant one^[10].

What these studies cannot see

Pregnancies that did not end in a live birth

Several malformation estimates come only from live births. Cesta followed live-born infants^[8], and Brown measured malformations among live-birth pregnancies linked to an infant^[10]. A pregnancy lost or terminated because of an anomaly may never enter that count, which could hide some harm. Pregnancy loss itself is not clearly raised in the adjusted data: Brown’s weighted risk of nonlive birth was 29.7% with continuation and 27.1% without (adjusted RR 1.09; 95% CI 0.98 to 1.23)^[10]. The unadjusted Lewin cohort did report more first-trimester loss or termination (9.8% vs 7.5%)^[7]. In the teratology cohort, the cumulative incidence of termination was 18% in exposed pregnancies against 6% and 8% in the reference groups, while cause-specific models found no increase in pregnancy loss^[9]. The abstract does not report why people chose termination.

When the drug was actually in the body

Most exposure definitions rest on a filled prescription. The windows vary from 30 days before through 91 days after conception in France^[11] to 8 weeks either side of the last menstrual period in Denmark^[13]. The Danish authors note they had no data on whether a redeemed prescription was taken^[13]. Counting non-users as exposed can blur real differences toward the null. Semaglutide’s long half-life cuts the other way, because a dose stopped before conception may still count as relevant exposure; that is the reason its labels specify a two-month washout before a planned pregnancy^[15][17].

Diabetes as the competing explanation

Maternal diabetes carries its own malformation risk. In Cesta 2024, the standardized prevalence of major malformations was 3.7% across all infants and 5.3% among infants of women with type 2 diabetes^[8]. The Ozempic label cites a background risk of 6 to 10% with pregestational diabetes and a periconceptional HbA1c above 7, and as high as 20 to 25% above 10^[17]. The French study could not measure BMI or HbA1c directly and used proxies for diabetes severity^[11], and Brown 2026 names prior glycemic control as a possible source of residual confounding^[10].

The Danish cohort shows why indication matters. Among 529 exposed pregnancies, preterm birth was elevated for women using these drugs for diabetes (liraglutide aOR 1.70; 95% CI 1.17 to 2.48; semaglutide aOR 1.84; 95% CI 1.24 to 2.71) but not for weight management (liraglutide aOR 1.01; 95% CI 0.58 to 1.76; semaglutide aOR 0.71; 95% CI 0.30 to 1.70)^[13]. The authors read that as compatible with diabetes-related factors driving the risk, and the design cannot establish cause. The paper discloses Novo Nordisk Foundation funding for several authors and industry ties for some. It studied obstetric outcomes rather than malformations, but the same split complicates any pooled estimate that mixes the two indications.

People using a GLP-1 only for weight loss are underrepresented in the malformation data shown here, since many of the cohorts enrolled people with diabetes^{[7][8][11][12]}. Evidence for individual drugs is sparse^[1]. The Zepbound label states that available tirzepatide data are insufficient to evaluate a drug-related risk of major birth defects^[16], and pooled reviews do not resolve that gap.

Putting the confidence intervals in absolute terms

FDA labels put the US general-population background risk of major birth defects at 2 to 4%^[16][17]. Take 3% as a round illustration. The upper bound of the timing-verified estimate, 1.27^[1], would mean about 3.8%, and the upper bound in Brown 2026, 1.82 for continuing versus not continuing^[10], about 5.5%. This is our arithmetic and it treats odds and prevalence ratios as approximate risk ratios, which holds reasonably at low baseline rates. These are the upper bounds of those two estimates, not expected outcomes, and the smaller studies in the tables above have wider intervals still. Both intervals also include no difference at all. A 3% starting point understates the absolute numbers for people with diabetes, whose baseline risk is higher^[17].

Questions to raise after an accidental exposure

• How does my diabetes or HbA1c affect my baseline risk? For some patients, that figure may matter more than the drug estimate^[17].
• Should my anatomy scan or follow-up plan change? None of these reviews sets a screening protocol, so this is a question for your obstetric clinician.
• Is there a pregnancy registry I should join? The Wegovy and Zepbound labels both describe exposure registries, which help fill the data gap^[15][16].

A 2026 BJOG narrative review advises individualized counseling that treats exposure as neither established harm nor proven safety^[14]. The Hakim review offers “cautious reassurance” after inadvertent exposure and does not support routine use in pregnancy^[4]. For anyone not yet pregnant, contraception is the practical gap to close, and tirzepatide users on the pill should read about the backup contraception warning. Hakim also calls for robust studies with long-term follow-up^[4].