Selective serotonin reuptake inhibitor use during early pregnancy and congenital malformations: a systematic review and meta-analysis of cohort studies of more than 9 million births

PubMed, Embase, Web of Science, and The Cochrane Library were searched from database inception to 17 January 2018. The search strategy combined medical subject heading (MeSH) and Embase subject heading (EMTREE) terms with other unindexed or free-text terms. Details of the full search strategy are provided in Additional file 2. Reference lists of retrieved articles and previous systematic and narrative reviews were searched manually to retrieve all relevant documents. No language restrictions were imposed.

Cohort studies or randomized controlled trials that reported original data were eligible for inclusion if they reported any congenital malformations in infants born to mothers with any exposure to SSRIs or individual SSRIs (citalopram, fluoxetine, paroxetine, sertraline, escitalopram, or fluvoxamine) during the first trimester, had a comparison group that included pregnant women who were not exposed to any antidepressants and/or teratogens (folic acid antagonists, angiotensin-converting enzyme inhibitors, anticonvulsants, coumarin derivatives, and retinoids), and, if a risk estimate was not reported, provided necessary distribution of exposure, non-exposure, cases, and non-cases, from which a risk estimate could be calculated.

The titles and abstracts of retrieved articles were evaluated by two independent reviewers (S-YG and CS). The full texts of potentially eligible studies that seemed to meet the inclusion criteria were then obtained and independently reviewed by the two reviewers. Any disagreements were identified and resolved by discussion or by consultation with a third reviewer (Q-JW). If data were duplicated in more than one study, we included the study with the largest number of cases.

A standardized, pre-designed spreadsheet was used for data extraction from the included studies. The study quality and synthesis of evidence were assessed. The following data were extracted into the spreadsheet: first author, publication year, geographic location, study period, data source, sample size (cases and cohorts), types of birth, definition of outcome, outcome with their risk estimates and 95% confidence intervals (CIs), and adjusted confounders. Congenital malformations were identified and defined according to the European Surveillance of Congenital Anomalies (EUROCAT) Guide 1.3 and ICD-10 and ICD-9 codes (Additional file 3: Table S1). The primary outcomes of interest were overall MCAs and specific CHD. The secondary outcomes of interest were other system-specific malformations (nervous system defects; eye defects; ear, face, and neck defects; respiratory system defects; orofacial cleft; digestive system defects; urogenital system defects; urinary system defects; genital system defects; musculoskeletal system defects; limb; and abdominal wall defects. Two reviewers (T-NZ and Z-QS) extracted data independently; any disagreements were resolved by discussion with a third reviewer (S-YG) where necessary.

For a study [36] that reported a different follow-up duration, the estimate of the follow-up duration during the first 6 years of life was extracted. For studies [31, 40, 44, 45, 47, 55‐61] that did not report any adjusted risk estimate, we used the crude risk estimate. If a study lacked required data, they were requested by contacting the study authors by email.

We used the Newcastle-Ottawa scale [62] to assess the risk of bias of cohort studies, which included studies based on the selection of study participant groups (four stars), the comparability of study groups (two stars), and the ascertainment of outcome (three stars). Studies were considered to have low risk of bias if they achieved a full rating in at least two categories of selection, comparability, or outcome assessment [63].

For a study [64] that separately reported the risk estimates of SSRIs but did not report combined estimates, the effective count method proposed by Hamling et al. [65] was used to recalculate the effect estimate. Another study [56] reported results separately (but not combined) for CHD (bulbus cordis anomalies and anomalies of cardiac septal closure and other congenital anomalies of heart), nervous system malformations (spina bifida and other congenital anomalies of nervous system), digestive system defects (cleft palate and cleft lip, other congenital anomalies of upper alimentary tract, and other congenital anomalies of digestive system), and musculoskeletal system defects (certain congenital musculoskeletal deformities, other congenital musculoskeletal anomalies, and other congenital anomalies of limbs); here, the results were pooled using a fixed-effect model to obtain an overall combined estimate before combining these estimates with the remaining studies (Additional file 3: Table S1). Similar analyses also were performed for limb defects (limb reduction and clubfoot) [31, 34]. If the selected study did not include a risk estimate, the unadjusted risk estimate and the 95% CI were calculated from the raw data for simplicity [31, 38, 44, 45, 57, 59, 61, 66]. Because the odds ratio is an excellent approximation of the risk ratio in the case of rare outcomes, the results were referred to as relative risks (RRs) [67]; therefore, all results were reported as RR for simplicity. Estimates were pooled using the DerSimonian and Laird random-effects model to calculate summarized RRs and 95% CI [68].

We used the I² statistic to assess heterogeneity in effect measures between the studies. I² values of 25, 50, and 75% were considered to represent low, moderate, and high heterogeneity, respectively [69]. If ≥ 8 studies were available, potential sources of heterogeneity were explored by conducting subgroup analyses according to the following parameters: study quality (high risk vs. low risk), geographic location (Europe vs. Northern America or other regions), and adjustment for potential confounders (adjusted vs. unadjusted) including maternal age, socioeconomic status, smoking, alcohol drinking, body mass index (BMI) during pregnancy, pregnancy complications, and parity. Heterogeneity between subgroups was evaluated by meta-regression analysis. The potential for publication bias was examined through Begg’s and Egger’s tests [70, 71]. To determine the influence of an individual study in each main analysis of the estimated RR, we conducted a sensitivity analysis that recalculated the pooled effect by omitting one study at a time. Analyses were performed with Stata version 11.0 (StataCorp, College Station, TX). A two-tailed P value less than 0.05 was considered as statistically significant.

We identified 10,919 potentially eligible articles in PubMed, Embase, Web of Science, and The Cochrane Library. Two additional studies were identified in a manual search of the reference lists. The titles and abstracts were screened, and 79 articles qualified for full-text review (Fig. 1). The authors of two studies failed to respond to requests for additional data. Finally, 29 cohort studies (published between 1996 and 2017) providing 649 data points that contributed to the quantitative synthesis met all inclusion and exclusion criteria, which included a total of 9,085,954 individuals for analysis. These included 25 studies focused on women in the general population, 8 studies focused on women with a psychiatric disorder, and 6 studies focused on both; 7,926,215 untreated pregnant women without psychiatric disorders, 1,916,076 SSRI-untreated women with psychiatric disorders, and 59,894 SSRI-treated women with psychiatric disorders; 7,590,399 individuals from Europe (15 studies), 1,206,094 from North America (10 studies), and 289,461 from Japan and Israel (4 studies). The key characteristics of the included studies are presented in Additional file 3: Table S2.

Analysis of the included studies using Newcastle-Ottawa criteria indicated that 23 studies were low risk and 6 were high risk for bias. All studies achieved a total score of 4 to 9 (median = 8) (Additional file 3: Table S3).

The results of subgroup and meta-regression analyses are presented in Additional file 3: Table S11-S15. Subgroup analyses indicated that the low risk of bias studies and European studies were generally consistent with the main results; however, they were not all statistically significant. No statistically significant source of heterogeneity was identified in meta-regression analyses. The sensitivity analysis omitted one study at a time, which showed the results appeared to be robust to the influence of individual studies. By contrast, the pooled RR of MCAs was 1.06 with SSRIs (95% CI 0.85 to 1.32, I² = 0, P = 0.67) after excluding the study by Huybrechts et al. [55].

We identified a small but significant association between maternal use of SSRIs during the first trimester and MCAs in infants. This observed increase was consistent with the results of previous meta-analyses [15, 19] However, the association was attenuated after controlling for psychiatric diagnosis. Although the effects of SSRIs could never be completely separated from a psychiatric illness itself, such estimates are likely to be influenced by potential confounding by indication. Jimenez-Solem and colleagues [37] first differentiated between the consequences of SSRI use and the underlying disease, and found an increased risk of MCAs among infants born to women who took SSRIs during the first trimester, whereas the correlation was not significant for women who paused their use of SSRIs during the first trimester in a nationwide cohort study. Ban and colleagues [72] compared risks between pregnant women with medicated and unmedicated depression based on a population-based cohort study. Although results for MCAs were not statistically significant, the point estimate slightly decreased in comparison with pregnant women treated with SSRIs.

There was an association between first-trimester exposure to SSRIs and the risk of CHD. Some biological evidence possibly supports the observed increase [8, 73], but the results were inconsistent in three meta-analyses [14, 15, 19]. This difference is probably due to variations in sample size, study design, and time of exposure. The meta-analysis by Wang and colleagues [14] included four population-based cohort studies enrolling 1,996,519 participants and found no significant associations between the use of SSRIs and heart defects (Additional file 1: Table S1). In recent years, fourfold population-based cohort studies including nearly eight million participants have been conducted to examine the aforementioned relationship, and the results were consistent with our primary results (our unreported data). Myles and colleagues [15] synthesized evidence from nine cohort studies combined with case-control studies; however, high heterogeneity might have been inherent in those data. Nikfar and colleagues [19] assessed the risk of cardiovascular malformations with the use of SSRIs during pregnancy, but not during the first trimester. Similarly, the association was markedly attenuated after controlling for psychiatric diagnosis.

Potential biological mechanisms of SSRI use and the increased risk of congenital malformations is based on studies of drug metabolite levels in cord blood in human [74]. In vitro, a growing body of evidence has suggested that the neurotransmitter serotonin (5-HT) plays a crucial role as a signaling molecule in cardiogenesis [8, 75]. Consequently, disruption of the 5-HT signaling caused by the use of SSRIs may result in several different types of CHD [73, 76].

Our study has several strengths. First, this meta-analysis of current evidence from cohort studies includes the largest sample size (more than nine million births) analyzed to date and combines the results with the most comprehensive data related to associations between the use of SSRIs and all individual SSRIs during early pregnancy and the risks of 29 categories of congenital malformations. Second, this meta-analysis pays particular attention to the potential confounding by indication. Third, for ethical reasons, there are no randomized controlled trials; therefore, the quality of evidence from cohort studies could provide clinicians and pregnant women with a reference in clinical practice.

There are limitations in our meta-analysis related to evidence synthesis and quality. First, the definition of outcomes varied among studies, particularly the definition of CHD, which could contribute to the high heterogeneity in our study. CHD has a specific definition and coding in the EUROCAT guide, but not in the ICD codes; however, most of the individual studies defining outcomes were based on ICD codes such as ICD-10 and ICD-9 (Additional file 3: Table S1-S2). We also failed to find any specific coding in either the ICD codes or EUROCAT subgroups related to MCAs, septal defects, RVOTD, and left ventricular outflow tract defects. Furthermore, the definitions of outcomes might depend on the authors of individual studies. For example, Ban and colleagues [72] defined septal defects as ASD, VSD, and atrioventricular septal defects, whereas Pedersen and colleagues [43] defined the defects as ASD and VSD. Additionally, the follow-up duration may also be a potential source of heterogeneity. On the one hand, some types of CHD (e.g., VSDs) may be self-healing [77]. On the other hand, due to the serious malformations are usually symptomatic with early detection, whereas milder malformations are sometimes identified at later age [36].

Second, the majority of individual studies only included live births. Stillbirth, spontaneous abortion, or induced abortion caused by severe malformations [78, 79] were not always recorded or observable and would have been missed, which could introduce selection bias and underestimate the strength of the associations between the use of SSRIs during early pregnancy and congenital malformations in infants [80]. The restriction of results according to different data sources also could also result in potential bias. The pooled effects of this study were dominated by record-linkage studies. However, data collected from prescription registries, dispensation registries, or drug reimbursement registries that rely on dispensed prescription information to determine maternal use of SSRIs would lead to misclassification of exposure [81, 82]. The dispensing of SSRIs may not always precisely reflect the specific time of exposure or verify that SSRIs were actually taken as prescribed. Selection bias presents a potential limitation in teratology information service studies, as women recruited during early pregnancy who feel the need for counseling about the teratogenic potential of SSRIs may be at higher risk than those who have no concerns [83].

Third, the event rate of congenital malformations in infants is very low, and individual studies may not have consistently adjusted for potential confounders. Therefore, we included adjusted or unadjusted risk estimates in our meta-analysis. Unadjusted risk estimates should be interpreted with caution, but the main results were still robust after removing crude risk estimates in the sensitivity analysis. Due to the lack of information on other potential confounders such as folic acid supplementation and familial-related factors, we could not fully rule out the possibility of residual confounding. For example, the study by Furu and colleagues [34] reported results from sibling design in addition to the full cohort. The results of sibling-controlled analyses showed attenuated risk compared with the full cohort. Thus, the small observed increased risk could be explained by familial-related factors or other lifestyle-related factors not adjusted for. In addition, we lacked information about the restricted cohorts regarding severity of disease. Eliminating the potential teratogenicity of SSRIs from a potential effect of the underlying psychiatric diagnoses remains a challenge.

Fourth, as we could not obtain an estimate for the incidence of MCAs and CHD events in the general population or in women with a psychiatric diagnosis, we could not provide an absolute risk increase of MCAs and CHD associated with exposure to SSRIs in the general population and in women with a psychiatric diagnosis. However, we obtained some examples from the published studies to give a suggestion of the increased absolute risk. Ban and colleagues [72] reported that children born to women with diagnosed depression unmedicated in early pregnancy had higher absolute risks of overall MCAs than children of mothers with no depression (absolute risk increase: 15 per 10000 births). Futhermore, Alwans et al. and Huybrechts et al. [55, 84] found a small increase in the absolute risk of CHD with exposure to SSRIs. Although the absolute risk of MCAs and CHD were highly likely to remain small, it is still of concern to pregnant women.

Fifth, it should be recognized that the implicated system-specific malformations and controlled psychiatric diagnosis studies are rare. The small number of included studies limited the statistical power of the study, which limited our ability to perform subgroup analyses to further investigate these issues and interpret the results. There was insufficient evidence to estimate fetal outcomes for the dosage of SSRIs use during pregnancy. Thus, we were unable to conduct a dose-response analysis.

Finally, since the study focused on non-exposure, i.e., pregnant women who were not exposed to any antidepressants and/or teratogens, rather than pregnant women who were exposed to other individual SSRIs, we could not determine if any of the individual SSRIs was preferable over others.