Pregnancy and Delivery Outcomes Following Benzodiazepine Exposure: A Systematic Review and Meta-analysis

Understanding the effects of benzodiazepines (BZDs) on maternal/fetal health remains incomplete despite their frequent use. This article quantifies the effects of antenatal BZD exposure on delivery outcomes.

Medline, PsycINFO, CINAHL, Embase, and the Cochrane Library were searched till June 30, 2018.

English-language cohort studies comparing antenatal BZD exposure to an unexposed group on any delivery outcome were eligible. In all, 23,909 records were screened, 56 studies were assessed, and 14 studies were included.

Two reviewers independently assessed quality and extracted data. Estimates were pooled using random effects meta-analysis. Sub-analyses examined several potential moderators including timing of exposure.

There were 9 outcomes with sufficient data for meta-analysis. Antenatal BZD exposure was significantly associated with increased risk of 6 outcomes initially: spontaneous abortion (pooled odds ratio = 1.86; 95% confidence interval [CI], 1.43 to 2.42), preterm birth (1.96; 95% CI, 1.25 to 3.08), low birth weight (2.24; 95% CI, 1.41 to 3.88), low Apgar score (2.19; 95% CI, 1.94 to 2.47), Neonatal Intensive Care Unit (NICU) admission (2.61; 95% CI, 1.64 to 4.14), and induced abortion (2.04; 95% CI, 1.23 to 3.40). There was significant heterogeneity between studies for most outcomes without consistent moderators. Birth weight (mean difference [MD]: −151.35 g; 95% CI, −329.73 to 27.03), gestational age (−0.49 weeks; 95% CI, −1.18 to 0.19), and small for gestational age (SGA; 1.42; 95% CI, 1.00 to 2.01) did not show significant associations although after adjusting for publication bias, gestational age, and SGA became significant, totaling 8 significant outcomes.

Antenatal BZD exposure appears to be statistically associated with increased risk of several adverse perinatal outcomes. Although confounds cannot be ruled out, NICU admission does appear clinically relevant and consistent with the antidepressant literature.

Comprendre les effets des benzodiazépines sur la santé de la mère et du fœtus demeure incomplet malgré l’usage fréquent de ces médicaments. Le présent article quantifie les effets de l’exposition prénatale aux benzodiazépines sur les résultats de l’accouchement.

Une recherche a été menée dans Medline, PsycINFO, CINAHL, Embase, et la Cochrane Library jusqu’au 30 juin 2018.

Les études de cohortes en anglais comparant l’exposition prénatale aux benzodiazépines avec un groupe non exposé selon le résultat de l’accouchement étaient admissibles. La recherche a repéré 23 909 études, 56 ont été évaluées, et 14 études ont été incluses.

Deux réviseurs ont évalué indépendamment la qualité et les données extraites. Les estimations ont été regroupées à l’aide d’une méta-analyse à effets aléatoires. Des sous-analyses ont examiné plusieurs modérateurs potentiels, notamment la durée d’exposition.

Il y avait 9 résultats dont les données suffisaient pour la méta-analyse. L’exposition prénatale aux benzodiazépines était significativement associée à un risque accru de six résultats initialement: avortement spontané (rapport de cotes regroupées = 1.86; intervalle de confiance à 95% CI, 1.43 à 2.42), naissance prématurée (1.96; 1.25 à 3.08), faible poids à la naissance (2.24; 1.41 à 3.88), faible indice d’Agpar (2.19; 1.94 à 2.47), hospitalisation à l’unité de soins intensifs néonatals (USIN; 2.61; 1.64 à 4.14), et avortement provoqué (2.04; 1.23 à 3.40). Il y avait une hétérogénéité significative entre les études pour la majorité des résultats sans modérateur constant. Le poids à la naissance (différence moyenne (DM): −151.35 g; −329.73 à 27.03), l’âge gestationnel (DM = −0.49 semaines; −1.18 à 0.19), et nourrisson petit pour l’âge gestationnel (1.42; 1.00 à 2.01) n’indiquaient pas d’associations significatives bien qu’après ajustement pour biais de publication, l’âge gestationnel et nourrisson petit pour l’âge gestationnel devenaient significatifs, totalisant 8 résultats significatifs.

L’exposition prénatale aux benzodiazépines semble être statistiquement associée à un risque accru de plusieurs résultats périnatals indésirables. Bien que des facteurs de confusion ne puissent être exclus, l’hospitalisation à l’USIN semble cliniquement pertinente et compatible avec la littérature sur les antidépresseurs.

This systematic literature review was guided by an Advisory Committee of key stakeholders and followed the guidelines of the Meta-Analyses and Systematic Reviews of Observational Studies.¹² Details of our methods have been previously described as this is one study in a program of research.¹³ Two professional librarians with expertise in psychiatry and psychopharmacology conducted independent literature searches as part of a broader study of the impact of antenatal anxiety and antenatal use of anxiety-related medications. One librarian searched Ovid Medline, Medline In-Process & Other Non-Indexed Citations, PsycINFO, CINAHL, Embase, and the Cochrane Library from their start date to June 30, 2018. Keywords included anti-anxiety agents, anxiolytic, benzodiazepine, chlordiazepoxide, bromazepam, clonazepam, diazepam, lorazepam, nitrazepam, oxazepam, temazepam, pregnancy, prenatal or antenatal, preterm birth or premature delivery, low birth weight, birth weight, gestational age, Apgar, NICU, neonatal/infant outcomes, delivery outcomes, among others (available upon request). The second librarian consulted with the review team and used an iterative process to develop and test the search strategy. This strategy was reviewed by another senior information specialist using the Peer Review for Electronic Search Strategies Checklist¹⁴ prior to execution. Strategies utilized a combination of keywords (e.g., antenatal, anxiolytics) and controlled vocabulary (e.g., “pregnancy,” “anxiety disorders,” “anti-anxiety agents”), adjusting vocabulary and syntax as needed across databases (full search strategy available from the authors upon request). Results were limited to the English language, human studies, and nonopinion pieces. Reference lists from reviews, meta-analyses, and the final selection of included articles were also searched. The preferred reporting items for systematic review and meta-analyses (PRISMA) were followed.¹⁵

Studies with original, prospectively collected data were eligible if they included a BZD-exposed group and an unexposed comparison group. Acceptable assessment of BZD exposure included either filling a prescription for a BZD or notation of BZD use in maternal/clinical charts. Studies that examined outcomes in the mother or infant in the gestational, delivery, and/or postpartum period were accepted. For cases in which a sample was repeated in more than one publication, we chose the article that most closely addressed our research question or the one with larger sample size if that differed. Studies presenting only data aggregating BZDs with other psychotropic medication(s) were excluded. Case-control studies and one study where BZD was administered during an in-hospital stay for the treatment of severe hyperemesis gravidarum were excluded. Due to the volume of potentially eligible studies, abstracts, conference proceedings, and unpublished data were also excluded. Meta-analysis was conducted if there were at least 3 eligible studies.

We used methods previously developed by our team for quality assessment and data extraction.¹⁶ Two trained reviewers independently conducted quality assessments, and disagreements were resolved on a case-by-case basis by discussion with the study leads (SG, LG). The Systematic Assessment of Quality in Observational Research (SAQOR),¹⁶ adapted from the Downs and Black¹⁷ Checklist and the Newcastle–Ottawa Scale,¹⁸ was used to assess article quality.

The studies were assessed on 19 aspects under the categories of sample (one example included if it was representative), control group (i.e., if one was included), quality of exposure/outcome measure (i.e., if the assessment was adequate), follow-up (i.e., if attrition rate was stated), and distorting influences/control for confounders. This last category, distorting influences, included whether the analyses controlled for confounders such as anxiety diagnosis, other psychotropic medications, and identification then control for additional confounders such as alcohol, smoking, or illicit substances. Assessments using the SAQOR criteria were combined with a modification of the Grading of Recommendations Assessment, Development, and Evaluation system,¹⁹ to assign each article with a quality rating of high, moderate, low, or very low quality. This rating was further classified as either “above quality threshold” (high, moderate, and low ratings) or “below quality threshold” (very low rating).

The Strengthening the Reporting of Observational Studies in Epidemiology criteria²⁰ were used as the foundation for data extraction procedures. Extracted data included authors, year of publication, source country, details of study design, participants (sample, control, demographics, and clinical characteristics), inclusion/exclusion criteria, details of BZD use (timing and indication of use in pregnancy, BZD type, and dose), use of other psychotropics in pregnancy, outcomes and their assessment methods and definitions, statistical adjustment for confounders, and loss to follow-up. Adjusted estimates with their variances were extracted when available; when adjusted estimates were not provided in the published data, we calculated crude odds ratios (OR) or mean differences (MDs) and sample variances. Before calculating the OR for studies that included cells with a 0 count, we added 0.5 to these cells. A request for data analysis clarification was sent to one study author and a response was received, allowing for inclusion of the study in our analyses. Data were extracted by one reviewer and checked by another and the primary author (SG); disagreements were resolved by discussion. Outcomes were as defined by the authors of the original publication.

MD, OR, 95% confidence intervals (CIs), and heterogeneity were determined using Review Manager (RevMan) Version 5.3. Random effects models were used (as heterogeneity was expected across studies) for pooled estimates of the weighted MD (for continuous outcomes) and for OR (for binary outcomes). Relative risks were treated as ORs when events were rare (i.e., rate of <10%) and converted to ORs. Between-study heterogeneity was assessed with Cochrane Q and quantified with I ², which represents the percentage of the between-study variance impacting quantitative results more than expected by chance, low (I ² = 25%), moderate (I ² = 50%), or high (I ² = 75%).²¹ The intended quantitative and qualitative analyses for publication bias included Egger’s test and visual inspection of funnel plots, respectively. In instances where publication bias was identified in analyses with 5 or more studies, the trim-and-fill method²² was conducted to compute an adjusted OR or MD.

For outcomes with ≥4 studies, we examined potential sources of heterogeneity through subgroup analyses, regardless of the Q statistic as they were planned a priori. Subgroup analyses examined sources of heterogeneity: (a) study quality (above or below quality threshold based on the quality assessment), (b) timing of BZD exposure (first trimester, up to 22 weeks, any time in pregnancy), (c) adjustments (data adjusted/not adjusted for confounders), (d) use of other psychotropic medications (exclusion was not specified or assessed) or BZD monotherapy, (e) controlling for psychiatric diagnoses (adjusted data, excluded, not specified), and (f) geographic location (North America, Europe, other).

This study was conducted without patient involvement, although there was involvement at the beginning of the program of research and the advisory committee steered the team.

Publication bias was assessed visually with funnel plot and quantitatively with Egger’s test when there were >5 studies. For preterm birth (6 studies) and low birth weight (6 studies), Egger’s test did not indicate significant publication bias. Visual inspection of the preterm birth funnel plot based on <37 weeks (5 studies) did indicate slight bias. Applying the trim-and-fill method resulted in a slightly higher and still significant OR (OR =2.51, 95% CI, 1.49 to 4.24, P < 0.001). For spontaneous abortion, birth weight, and gestational age, visual inspection of the funnel plots indicated some publication bias. The revised OR using the trim-and-fill method for spontaneous abortion was slightly attenuated but remained significant (OR = 1.43; 95% CI, 1.10 to 1.86, P < 0.001). The effects of antenatal BZD exposure on birth weight remained insignificant, after imputing one missing study for potential publication bias (MD = −92.48 g; 95% CI, −267.09 to 82.13, P = 0.30). Interestingly, gestational age became significant after imputing two missing studies to account for publication bias (MD = −0.88 weeks; 95% CI, −1.60 to −0.16, P = 0.02). Visual inspection of the Apgar at 5 min outcome indicated slight publication bias, and the revised OR using trim-and-fill remained significant (OR = 2.19, 95% CI, 1.94 to 2.46, P < 0.001). The NICU admission outcome also appeared biased, but the revised OR using trim and fill, while slightly reduced, remained significant (OR = 2.54, 95% CI, 1.71 to 3.78, P < 0.001). SGA also appeared to be positive for slight bias after visual inspection of the funnel plot. Interestingly, using trim and fill, the adjusted OR was of the same magnitude but reached significance (OR = 1.38, 95% CI, 1.04 to 1.85, P = 0.03). The induced abortion outcome had too few studies for generation of a funnel plot.

The study had several strengths, the most important of which was the rigorous methodology. Our search was comprehensive, and a broad range of perinatal outcomes was examined. A priori, we set out to explore potential sources of heterogeneity, regardless of significance. Interestingly, although there was significant heterogeneity, our a priori determined potential factors were not consistently significant but do suggest that we need to keep looking for other factors. We assessed study quality to ensure we based our conclusions on the best evidence. The fact that study quality was not a significant source of heterogeneity increases our confidence in the results with the majority of studies being above our quality threshold. Moreover, the findings are consistent with the antidepressant literature, but the magnitude of the ORs generally was higher, suggesting potentially stronger effects.

Study limitations stem primarily from methodological limitations of the included studies and represent the early state of scientific knowledge in the BZD literature and not the systematic review methods themselves. Specifically, all included studies were observational studies. Of these, 4/14 studies did not describe the indication for BZD use, 2/14 were based on women with epilepsy, and in 1 study, BZDs were used for obstetrical reasons. The studies that did not report indication for use of the BZD were typically the ones with the larger sample sizes. Further, the studies where indication was noted for a psychiatric disorder were based on women with mixed diagnoses, predominately depression and anxiety.^{10,26,27,29,31,32} While we did not find a significant effect for the moderator variable of diagnoses (or anxiety indication, data not shown), these factors indicate the study populations were likely heterogeneous. Psychiatric or other medical disorders (i.e., epilepsy) which may have required BZD use may themselves be associated with poor birth outcomes (e.g., preterm birth, delivery of a low birth weight infant).^13,34,35 A good example is the spontaneous abortion outcome. The subgroups with the “unspecified” groupings pooled to the highest ORs among the significant moderator sub-analyses. This is evidence for sample type explaining the between-study heterogeneity. Further, the Ban et al.¹⁰ study, which controlled for illness, found that the women with illness who stopped using BZDs had a lower risk compared to those who continued, but acknowledged that the severity of the illness (which was not measured) may have affected the results. That is, women with more severe illness may have stayed on the drug, and thus, the illness, and not necessarily the drug, may have affected the risk. The above limitations are a result of confounding, that is, the BZD association is distorted because other factors that the study groups differ on, such as diagnosis or indication, can also be associated with adverse outcomes. Effects of confounding by indication can explain a portion or all of the associations found. Moreover, residual confounding can still affect results even after controlling for certain variables. For example, the Bech et al.⁹ study statistically adjusted for a number of factors, but the authors acknowledged that the increased risk can still be attributed to “unmeasured confounding.” The Calderon-Margalit et al.²³ study acknowledged that both the indication (i.e., anxiety), which was not measured, and residual confounding by unmeasured variables (such as illicit drug use) may have influenced the results. At least 6/14 studies directly discuss confounding by indication and 10/14 residual confounding as limitations to their results.

Another limitation in pharmacoepidemiology studies includes noncompliance with the treatment. Recording in a database when a study drug is prescribed does not necessarily mean that the drug was taken as prescribed. Kallen et al.³⁶ reviewed this literature and concluded that data based on prescriptions do not necessarily reflect accurate use of the drugs. The studies that use prescription databases are the largest ones in terms of sample size and thus typically dominate the meta-analyses. This can be seen in the low Apgar score at 5-minute analysis, where the Kallen et al.⁵ study has a weight of 96.1% in the pooled OR, thereby contributing more to the weighted average than the other studies.³⁷ Moreover, in these large database studies, the time period of exposure may also not be accurate.³⁸

Given the limited data, we grouped the BZDs without examining individual drugs; because they differ at least in half-life, this may also limit our conclusions. Moreover, the pattern of use was not known. Chronic versus intermittent use of BZDs, especially at higher doses and with shorter half-life drugs, is known to confer different consequences in adults such as higher likelihood of dependence.^39–41 BZD dependence itself is linked to social determinants of health; these in turn may affect outcomes and thus confound results. In 6/14 studies (see Supplemental Table 1), the BZDs appear to have been prescribed or found in the medical record at least once, but it is unclear how often the drugs were used (e.g., other than during the first trimester). Furthermore, there is evidence that chronic versus occasional use has differing reliability implications, that is, self-report may be less reliable with occasional use.⁴² Bech et al.⁹ and Kilic et al.¹¹ reported an analysis of high versus low dose of all the drugs they studied (antiepileptics, including clonazepam); there were significant effects regardless of dose, although the magnitude of effects differed. Czeizel et al.²⁵ examined a wide dose range of diazepam from minimal dose up to 460 mg, while Laegreid et al.²⁶ described regular use of BZDs studied. In addition, known confounds for obstetrical outcomes were infrequently addressed. For example, maternal smoking⁴³ was not controlled in studies,^9,25,27,28 even if reported as significantly different between groups.²⁶

We have previously reported that antidepressants are associated with adverse perinatal outcomes.⁴⁴ Although not often investigated, some reports found elevated risks of neonatal outcomes associated with concurrent use of sedatives/hypnotics and antidepressants compared to the risks associated with the individual medication class.^5,45 Whether this increased risk is related to the combination of medications or, perhaps, to more severe psychopathology (necessitating combinations of medications) is unknown. The studies in our meta-analysis were also heterogeneous regarding concurrent use of other psychotropic medications (in addition to BZDs) in pregnancy, including different medications with unknown doses and windows of exposure.

Finally, the Freeman et al.³¹ study (a national registry of women housed in Massachusetts) and the Yonkers et al.³² study (which recruited some women from Massachusetts) may have some overlap. There was an 8-month overlap during their subject recruitment period, and it is possible that some women may be registered in both studies. While this could potentially inflate the results, the results did not change when these 2 studies were excluded.