Associations between endometriosis and adverse pregnancy and perinatal outcomes: a population-based cohort study

We conducted a population-based, longitudinal cohort study including all women 15 to 49 years of age with a singleton pregnancy in the period of 1980–2015 in WA.

We obtained maternal, infant and birth information from the Midwives Notification System, a validated database [26] that includes > 99% of births in WA of at least 20 weeks’ gestation or birthweight of 400 g or more if the gestational age was unknown [27]. We sourced hospitalization records from the Hospital Morbidity Data Collection, which includes information on all hospitalizations from public, private and day procedure facilities in the state with International Classification of Diseases (ICD-9/10th revision-Australian Modification) coded diagnoses [28]. Data sources have been described in detail elsewhere [29]. Data were probabilistically linked using best practice protocols through the WA Data Linkage Branch [30].

From a total of 487,297 women (964,015 births) during the study period, we sequentially excluded multiple gestations, stillbirths, and pregnancies with missing information for gestational age, outcomes, maternal age, and socioeconomic status (SES). This resulted in 468,778 eligible women who contributed to 912,747 singleton pregnancies included in the analytic cohort (Fig. S1).

We identified all women with a principal or additional diagnosis of endometriosis from the hospital separation data using the International Classification of Diseases (ICD)-AM (Australian Modification) diagnostic codes consistent with ICD-9: 617.0–617.9; ICD-10: N80.0-N80.9 and Australian Classification of Health Interventions (ACHI) for endometriosis-related procedures (codes are shown in Table S5). Women were categorized as having endometriosis if they had hospital admission or surgical procedure coded as a diagnosis of endometriosis. We included women diagnosed before and after pregnancy in the primary analysis because recent studies documented a diagnosis delay of 8–12 years [4‐6]. This approach has been adopted by other recent studies [15, 19].

The outcomes of interest were ascertained from the Midwives’ Notifications System and hospital separation data in the state, with the diagnostic codes consistent with preeclampsia (ICD-9/ICD-9-CM: 642.4, 642.5, 642.7, ICD-10-AM: O14, O11) and placenta previa, with or without haemorrhage (ICD-9/ICD-9-CM: 641.0–641.1, ICD-10-AM: O44.-). The onset of preeclampsia at the gestational age between 20 and 34 weeks and after 34 weeks of gestation was classified as early or late-onset preeclampsia, respectively. Preterm birth was defined as birth before 37 completed weeks of gestation, categorized into moderate preterm birth (gestational week 32–36) and very preterm birth (prior to 32 gestational weeks). Further, we also categorized preterm birth into spontaneous (due to spontaneous onset of labour) and medically indicated (due to elective caesarean section, or induction of labour). The details of ICD codes used to define variables for analysis are presented in Table S5.

Information on potential confounding factors including the calendar year of birth of the child (categorical variable), maternal age group (15–24, 25–29, 30–34, 35–39, 40–49 years), parity (0, 1, 2,  ≥ 3), smoking during pregnancy (Yes vs No), race/ethnicity (Caucasian versus non-Caucasian), and socioeconomic status (SES) was obtained from the databases. SES was measured using Socio-Economic Indexes for Areas (SEIFA). Specifically, we used the Index of Relative Socio-economic Disadvantage level at the time of birth of the child. These scores were obtained from the Australian Bureau of Statistics [31] and categorized into quintiles.

To assess the potential impact of Medically Assisted Reproduction (MAR) or infertility treatment on perinatal outcomes, we identified pregnancies with MAR procedure (using ACHI) or the following ICD diagnostic codes; ICD-9: 628.0-628.9, V26.1-V26.9, and ICD-10-AM: N97.0-97.9; Z31.1-Z31.9. These codes cover assisted reproductive technology (ART) techniques, intrauterine insemination, and ovulation induction and might include some spontaneously conceived pregnancies in couples with fertility issues [32].

We first estimated the unadjusted relative risks (RRs) with 95% confidence intervals (CI) using generalized linear models (GLM) fitted using a Poisson distribution and a log link function. Next, we estimated the causal effect (average treatment effect) of endometriosis on adverse pregnancy outcomes using the potential outcome approach, which allows for the estimation of causal effects in large observational data [33]. This was done by combining the inverse probability of treatment weighting (IPTW) and the outcome regression model in a doubly robust estimation [34]. IPTW weights each person by the inverse of their propensity score. The doubly robust estimation allows us to estimate the unbiased average causal effect when either the outcome regression model (traditional way of obtaining treatment effect) or the propensity score model (treatment selection model) is correctly specified [24, 25]. We estimated the adjusted RRs with 95% CI for each outcome derived using modified Poisson regression with a robust error variance to account for the effects of repeated pregnancies per mother [35]. We fitted the exposure model with maternal age, birth year, SES, ethnicity/race, and MAR treatment, and the outcome model with maternal age, birth year, SES, ethnicity/race and parity. As preeclampsia and placenta previa may influence the risk of preterm birth, gestational age (< 32, 32–36, > 37 weeks of gestation) was also included in the exposure model as a covariate for all outcomes included except for preterm birth. To examine the influence of MAR on the association between endometriosis and adverse pregnancy outcomes, we included a sub-analysis stratified by MAR status (Table 3). To check the covariate balance after propensity score matching using IPTW, we performed diagnostics including standardized differences and % of bias in means of all covariates (Fig. S2, Table S7). The standardized differences describe between-group differences in units of standard deviation. The confounders included in the treatment weighting were decided based on prior knowledge as well as consideration of Directed Acyclic Graphs (DAGs) (Fig. S3).

For the main results, we conducted a complete case analysis as the proportion of missing data was small (< 3%, range 0.6% for gestational age to 1.8% for SES).

To check the robustness of our findings, we conducted several sensitivity analyses. Firstly, to ascertain the sensitivity of our result to higher-order parity, we restricted the analysis to primiparous women. Secondly, to limit the possibility of misclassification bias, we conducted an analysis restricted to women (i) with a principal diagnosis of endometriosis, a diagnosis established to be chiefly responsible for occasioning an episode [28]; (ii) with any diagnosis of endometriosis prior to the birth of the child to ensure endometriosis was present during pregnancy; (iii) with endometriosis diagnosis before delivery and up to five years after delivery; and (iv) considering endometriosis diagnosis at more than one-time point during five years look-up period. Thirdly, to explore the potential influence of maternal smoking during pregnancy, which was routinely collected in the Midwifery notification from 1997 onwards, we conducted a separate analysis adjusting for smoking. Next, we compared the effect of endometriosis on preterm birth (very preterm vs moderate). Fifth, we undertake a causal mediation analysis based on the counterfactual framework using a parametric regression approach [36] to estimate the natural direct effect of endometriosis compared with the natural indirect effect through MAR. Finally, to assess the extent of unmeasured confounding, we calculated E-values, which represent the minimum strength of association on the risk ratio scale, that any unmeasured confounder would need to have with both endometriosis and each outcome to fully explain away the observed association, conditional on the measured covariates. [37]

All analyses were performed using Stata version 16.1 (Stata Corporation, College Station, Texas, USA).

In total, we included 912,747 eligible singleton births with a gestational age of 20–44 weeks from women (n = 468,778) aged 15–49 years in the study period between 1980 and 2015 in WA. In these pregnancies 8874 women (1.9%) had a diagnosis of endometriosis, corresponding to 19,476 pregnancies (2.1%). Women with endometriosis were on average of advanced age at the time of birth (> 35 years), Caucasian, and had a higher proportion of medically assisted reproduction compared to women without endometriosis. Socio-economic status, parity, and ethnicity were similar among exposed and non-exposed groups (Table 1).

The prevalence of pregnancy complications was higher among pregnancies of women with endometriosis compared to women without endometriosis (preeclampsia, 7.5% vs. 6.1%; preterm birth, 10.2% vs. 7.0%; placenta previa, 1.9% vs. 1.0%; Table 2).

Using doubly robust estimation, pregnancies in women with a diagnosis of endometriosis were associated with a higher risk of preeclampsia (RR 1.18, 95% CI 1.11–1.26), placenta previa (RR 1.59, 95% CI 1.42–1.79), and preterm birth (RR 1.45, 95% CI 1.37–1.54). This risk associated with endometriosis was higher for medically indicated preterm birth (RR 1.74, 95% CI 1.58–1.93) compared to spontaneous preterm birth (RR 1.40, 95% CI 1.27–1.56) (Table 2). Furthermore, endometriosis was associated with both moderate and very preterm birth with a stronger association observed for very preterm birth (Table S3). In a stratified analysis based on MAR status, the higher risk of placenta previa and preterm birth persisted regardless of conception mode with the strongest effect estimates for the non-MAR group (RR 1.77, 95% CI 1.50–2.08 for placenta previa and RR 1.67, 95% CI 1.55–1.80 for preterm birth) and slightly attenuated effect estimates among the MAR group. However, for preeclampsia, the observed association disappeared when stratified by MAR status (Table 3).

The results restricted to nulliparous women were consistent with the main finding with a slight attenuation (Table S2; Model 2). Findings from the analyses restricted to a subset of the study population with different exposure definitions were very similar to those reported in the main analyses (Table S2; Model 3–6). Analyses restricted to pregnancies from women with endometriosis diagnosed before delivery also resulted in slightly higher risk estimates for all adverse pregnancy outcomes evaluated (Table S2; Model 4). Additionally, the pattern of the association between endometriosis and adverse pregnancy outcomes was similar when further adjusted to smoking status (Table S2; Model 7). Our mediation analyses suggest that the percentage of endometriosis effect on preterm birth and placenta previa that was mediated through MAR was 8% and 3% respectively. (Table S4). The E-values for the observed RRs varied from 1.64 to 2.87 for these three adverse pregnancy outcomes (Table S6).

To our knowledge, this is the first population-based retrospective cohort study to examine the association between endometriosis with adverse pregnancy outcomes using the potential outcome framework. Using a large (~ 1 million births) cohort in WA, we observed a higher risk of preeclampsia, placenta previa, and preterm birth among pregnancies in women with endometriosis as compared to women without endometriosis. The associations persisted after stratification for conception mode (MAR or natural conception, non-MAR), with an elevated risk among the non-MAR group meaning the risks observed were attenuated among pregnancies conceived by MAR. The risk for adverse pregnancy outcomes was higher (approximately 24%, 56%, and 85% of increased risk of preeclampsia, preterm birth, and placenta previa, respectively) when we restricted our sample to women with endometriosis as the principal diagnosis code, suggesting probably more severe disease. These observed associations were not mediated through MAR.

Our cohort was based on longitudinally linked, highly reliable sources of population-based perinatal information ascertained from hospital separations and midwives’ notifications. We also included sensitivity analysis to check the robustness of our result. Our cohort is less prone to exposure misclassification bias since the hospital morbidity data collection (source data for our exposure) contains records for all hospital separations of admitted patients from all public and private hospitals in WA. Furthermore, our study restricted the analysis to singleton pregnancies, which improved generalizability to other similar cohorts.

We only had information on the diagnosis of endometriosis for women who have been hospitalized during the study period. Such data may likely represent more severe stages of endometriosis. The clinical routines and obstetric care have changed through the years and the diagnosis and awareness regarding endometriosis have evolved. However, to minimize this bias our model included the birth year of the child as a covariate. In our main analysis, we included all women with any diagnosis of endometriosis (principal and additional diagnosis). This could have introduced non-differential misclassification bias (i.e., independent of the outcome) and, therefore, will potentially bias the results towards the null. To limit this possible misclassification, we included a sensitivity analysis restricted to an exposure defined as a principal diagnosis of endometriosis, which indicated higher risk estimates as compared to the main result. We opted to include women with a diagnosis of endometriosis before and after pregnancy to account for the diagnostic delay [4, 5]. This could induce similar misclassification bias and attenuation of the association. Indeed, our sensitivity analysis restricted to a diagnosis of endometriosis before delivery consistently suggested higher effect estimates as compared to the main analysis. Though the validity of the diagnosis of endometriosis in the hospital separation database remains unknown, previous analysis of the same database suggested that endometriosis is reliably recorded in the hospital separation data [38]. Moreover, while the use of ICD and procedure codes ensures that those classified as having endometriosis are likely true cases, there is the possibility that the comparison group may have undiagnosed endometriosis. Our cohort had a relatively small number of events for stillbirths to be considered as an outcome. We, therefore, excluded pregnancies resulting in stillbirths from our analysis. This may have introduced a livebirth bias in the association between endometriosis and adverse pregnancy outcomes. However, a previous simulation study indicated that the magnitude of this bias is small [39]. Despite the potential susceptibility of doubly robust estimation to the limitations of misspecification bias, our study employs directed acyclic graphs (DAGs) in the selection of variables for the exposure and outcome models which helps to address the limitation and strengthens the robustness of our results.

Our study observed a modest association between endometriosis and preeclampsia, which is consistent with previous studies [12, 15, 20, 21]. However, other studies have reported a lower risk of preeclampsia [13, 17] with some suggesting no association [14]. These controversial results may be related to sample size, heterogeneity in exposure or outcome definition, selection bias, not taking MAR into account, diagnostic methods, and disease heterogeneity.

The attenuated risk in women who received MAR treatment is supported by other studies that did not find an association or a reduced association between endometriosis and preeclampsia or hypertension in pregnancy in women with an endometriosis and MAR procedure. [40, 41] Our study also found an increased risk of placenta previa in pregnancies among women with endometriosis, which is consistent with other research [20]. It has been suggested that the association may be confounded by the increased use of MAR in women with endometriosis. In our study, the association persisted even after stratification by MAR status, with a stronger association observed in non-MAR women which is consistent with other studies [40]. In our study, for women using MAR, the precision of the effect estimates was reduced likely because of the small sample size. For preterm birth, we observed higher risk estimates for very preterm deliveries compared to moderate preterm deliveries, and an association between endometriosis and both spontaneous and medically indicated preterm birth, with a stronger association for medically indicated preterm birth. This could imply that pregnancies from women with endometriosis are more likely to be induced or delivered through a caesarean section before gestational week 37. This finding is consistent with previous research and the association seems independent of MAR. [20, 41, 42] A smaller protective effect was also observed in a Canadian study [12].

Endometriosis may be associated with adverse pregnancy outcomes through various mechanisms, including effects on the uterine environment, progesterone signalling, and the remodelling of the spiral artery [18, 43, 44]. These factors may play a role in the association with preeclampsia, preterm birth and intrauterine growth restriction [45‐47]. Endometriotic lesions in the uterus may also reduce uterine contractility and cause abnormal implantation, leading to placenta previa [48]. In our sub-analysis of the timing of preeclampsia, we found an elevated risk for early onset compared to late-onset preeclampsia, which can be accounted for inadequate and incomplete trophoblast invasion of maternal spiral arteries [43]. MAR treatment itself has shown to be a risk factor for adverse pregnancy outcomes, with mixed results for women with endometriosis [47]. In MAR treatment, the effects caused by endometriosis such as inflammatory processes and regulatory disbalances are suppressed offering a better pregnancy environment and could explain the attenuation in the risks seen in our study in the MAR pregnancies group [49]. Women conceiving following MAR might also have support from better obstetric care and closer screening for adverse outcomes.

In this study, a stronger association between endometriosis and adverse pregnancy outcomes was observed, but residual and/or unmeasured confounding could not be completely ruled out. Nevertheless, the E-values for the observed RRs (ranging from 1.64 to 2.87) indicated that substantial confounding would need to explain away these associations (Table S6). Systematic reviews that examined the risk factors for endometriosis, for example, reported RRs ranging from 1.63 for smoking to 1.87 for overweight—lower than that of the E-values. [50, 51] In general, findings from our sensitivity analyses were remarkably similar to those reported for the main analysis and collectively support the hypothesis that endometriosis is associated with adverse pregnancy outcomes independent of MAR. Therefore, knowledge of a patient's endometriosis history may inform targeted prenatal care and reduce unfavourable pregnancy outcomes. Future studies would benefit from elucidating the potential mechanism that might explain how endometriosis affects implantation, placentation, and fetal growth and identifying potential interventions to decrease the risk of adverse perinatal outcomes.