Risk factors and high-risk subgroups of severe acute maternal morbidity in twin pregnancy: A population-based study

Objective To determine risk factors of severe acute maternal morbidity in women with twin pregnancies and identify subgroups at high risk. Methods In a prospective, population-based study of twin deliveries, the JUMODA cohort, all women with twin pregnancies at or after 22 weeks of gestation were recruited in 176 French hospitals. Severe acute maternal morbidity was a composite criterion. We determined its risk factors by multilevel multivariate Poisson regression modeling and identified high-risk subgroups by classification and regression tree (CART) analysis, in two steps: first considering only characteristics known at the beginning of pregnancy and then adding factors arising during its course. Results Among the 8,823 women with twin pregnancies, 542 (6.1%, 95% confidence interval (CI) 5.6–6.6) developed severe acute maternal morbidity. Risk factors for severe maternal morbidity identified at the beginning of pregnancy were maternal birth in sub-Saharan Africa (adjusted relative risk (aRR) 1.6, 95% CI 1.1–2.3), preexisting insulin-treated diabetes (aRR 2.2, 95% CI 1.1–4.4), nulliparity (aRR 1.6, 95% CI 1.3–2.0), IVF with autologous oocytes (aRR, 1.3, 95% CI, 1.0–1.6), and oocyte donation (aRR 2.0, 95% CI 1.4–2.8); CART analysis identified nulliparous women with oocyte donation as the subgroup at highest risk (SAMM rate: 14.7%, 95% CI, 10.3–19.1). At the end of pregnancy, additional risk factors identified were placenta praevia (aRR 3.5, 95% CI 2.3–5.3), non-severe preeclampsia (aRR 2.5, 95% CI 1.9–3.2), and macrosomia for either twin (aRR 1.7, 95% CI 1.3–2.1); CART analysis identified women with both oocyte donation and non-severe preeclampsia (SAMM rate: 28.9%, 95% CI, 19.9–37.9) and sub-Saharan nulliparous women with non-severe preeclampsia (SAMM rate: 26.9%, 95% CI, 9.9–43.9) as the two subgroups at highest risk. Conclusion In woman with twin pregnancy, rates of severe acute maternal morbidity vary between subgroups from 4.6% to 14.7% and from 3.8% to 28.9% at the beginning and at the end of pregnancy respectively, depending on the combined presence of risk factors.


Introduction
Twin pregnancies account for about 3% of all births in the United States and France [1,2]. Compared with women with singleton pregnancies, women with twin pregnancies have a fourfold increased risk of developing severe acute maternal complications, mainly during the intra and postpartum periods [3]. However this overall risk augmentation may actually vary by subgroups of women.
Risk factors of severe acute maternal morbidity in women carrying twins have been poorly characterized because previous studies, sparse and old, used variable definitions of maternal morbidity or lacked a specific control group, adequate sample size, and individual data for adjustment [4,5]. Moreover, extrapolation to twin pregnancies of risk factors identified in singleton ones might be inappropriate. Indeed, since twin pregnancy itself constitutes a high risk clinical context, its presence could modify the profile of other risk factors. Therefore, studies of the risk factors of severe acute maternal morbidity in the specific population of twin pregnancies are needed. Understanding which subgroups are at highest risk of severe acute maternal morbidity would be useful in both counseling women and for clinicians to be alert for early recognition and treatment should an event occur to limit its severity.
Because the JUmeaux MODe d'Accouchement (JUMODA) prospective cohort collected detailed individual data in a large population of twin pregnancies, it offered the opportunity to determine risk factors of severe acute maternal morbidity in women with twin pregnancies and identify subgroups at high risk [6].

Materials and methods
The JUMODA national, observational, prospective, population-based cohort study of twin deliveries took place in France from February 10, 2014, through March 1, 2015 [6]. All French maternity units performing more than 1500 annual deliveries were invited to participate, regardless of their academic, public, or private status or level of care, and 176 of the 191 eligible units (92%) agreed. Women who gave birth at or after 22 weeks of gestation were included (n = 8823). Enrolment took place prospectively immediately after delivery.
Detailed information about the participating women and maternity units has been already reported elsewhere [6]. Research nurses collected data about maternal characteristics, medical history, pregnancy complications, maternal complications, neonatal health, and maternity unit characteristics.
The primary outcome was a composite of severe acute maternal morbidity. This multicriteria definition was developed in a formal national Delphi expert consensus process for another study specifically conducted to define and study it (S1 Table) [3,7]. To include conditions involving severe health impairments, it combined diagnoses, organ dysfunctions, and interventions, as recommended by WHO [8]. Severe acute maternal morbidity was therefore defined as one or more of the following: maternal death; severe haemorrhage, defined by need for second line therapy, transfusion � 4 units of packed red blood cells, uterine artery embolization, vascular ligation, compressive uterine suture, emergency peripartum hysterectomy; eclampsia; preeclampsia responsible for induced preterm delivery before 32 gestational weeks mainly for the mother's health; pulmonary embolism; stroke or cerebral transient ischemic attack; severe psychiatric disorder; cardiovascular or respiratory dysfunction, renal dysfunction (creatininemia >1.47 mg/dL or oliguria <500 mL/24 h), neurological dysfunction (coma of any stage and duration), or hematological dysfunction (thrombocytopenia <50 000/ mm 3 or acute anemia <7 g/dL, in the absence of a chronic disorder); emergency surgery in addition to the childbirth procedure, e.g., secondary hysterectomy, laparotomy for a postdelivery complication other than hematoma or wound infection; admission to an intensive care unit (ICU). This primary outcome was treated as a binary variable.
The incidence of severe acute maternal morbidity was calculated, with its 95% confidence interval, as the number of women with a severe acute maternal morbidity event, divided by the total number of pregnancies ending in still-or live birth at or after 22 weeks of gestation in the JUMODA cohort. Among women with such an event, we described the distribution of the underlying causal conditions.
The characteristics of women and of their pregnancies that we tested as potential risk factors for severe acute maternal morbidity were selected from the literature and analyzed in two steps. First, we included only characteristics known at the beginning of pregnancy that might identify women at high risk of severe acute maternal morbidity and thus potentially improve their orientation and initial care. Second, because clinical situations may change significantly during pregnancy, we integrated the information collected over its course about potential complications that might constitute additional risk factors.
The characteristics analyzed as risk factors at the beginning of pregnancy were maternal age, maternal country of birth, prepregnancy body mass index, preexisting insulin-treated diabetes, preexisting hypertension, other preexisting chronic conditions, parity, previous caesarean delivery, mode of conception, and chorionicity. To identify risk factors for severe acute maternal morbidity and calculate adjusted relative risks (RR) with their 95% confidence intervals (CI), we used a multivariate Poisson regression model with a random intercept to take variability between centers into account. We then used the classification and regression tree (CART) descriptive and non-explanatory approach, [9,10] that is, we performed a CART analysis of the risk factors identified in the multivariate analysis to define and rank the factors most predictive of the risk of severe acute maternal morbidity and to individualize high-risk clinical subgroups. CART is a recursive partitioning statistical method that examines the dataset to find the best variables for grouping the women with and without severe acute maternal morbidity. Factors that are both frequent and discriminating rise in importance and result in groupings that bear resemblance and relevance to clinical practice. Among all the variables considered, CART selected the single factor that best separated the women with and without severe acute maternal morbidity to form the first node. The same procedure was then applied to each "child" node, which found the next most discriminating factor. For each node of the tree, we calculated the confidence interval of the severe acute maternal morbidity rate.
To identify the risk factors arising during pregnancy and therefore high-risk subgroups at the end of pregnancy, we repeated these statistical analyses (multivariate Poisson regression and CART analysis), adding the following variables: gestational insulin-treated diabetes, gestational hypertension, non-severe preeclampsia, placenta praevia, twin-to-twin transfusion syndrome, premature rupture of membranes, macrosomia for either twin (birth weight > 95 th percentile of the distribution of birth weights in this cohort), and hospital characteristics, including annual volume of twin deliveries and level of care.
Two sensitivity analyses were performed. First, in order to evaluate if non-severe preeclampsia is a risk factor of severe acute maternal morbidity events other than severe hypertensive complications and to evaluate if it is a discriminant factor in CART analysis not only because severe preeclampsia is part of the severe acute maternal morbidity definition, we excluded cases of severe acute maternal morbidity only due to hypertensive complications. Second, to explore whether the associations found for characteristics present at the end of pregnancy may be due to differences in subsequent delivery context, we conducted an additional analysis of risk factors also including delivery-related characteristics, i.e gestational age at and mode of delivery. For this analysis, we excluded the severe acute maternal morbidity events before labor (antepartum, n = 32) or at an unknown time (n = 2), since those could not have been caused by delivery.
The proportion of women with missing data for any covariate included in the main multivariate model ranged from 0% to 11.4%. There were 7438 (84.3%) women with no missing data, 1201 (13.6%) with only one missing data item, and 184 (2.1%) with at least two missing items for covariates included in the multivariate model. Characteristics of the women with full data were similar to those with missing data (data not shown). We used multiple imputationchained equations to impute missing data and generated 16 independent imputation data sets.
STATA 13 software (StataCorp LP, College Station, TX) was used for the descriptive and multivariate analyses. R Software Package (The R Foundation for Statistical Computing) was used for the CART analysis, in particular, the "rpart" R package.
The national data protection authority (CNIL, DR-2013-528), the consultative committee on the treatment of information on personal health data for research purposes , and the committee for the protection of people participating in biomedical research of Paris Ile-de-France 7 (PP-13-014) all approved this study. They approved that this observational study waived the need to obtain written informed consent according to the French law.
The main underlying causal condition of severe acute maternal morbidity was severe postpartum haemorrhage, accounting for 77.5% (n = 420, 4.76/100 twin pregnancies) of these cases (S2 Table). Admission to an ICU occurred in 22.3% (n = 121, 1.37/100 twin pregnancies). One woman in the cohort died of acute cardiac arrhythmia before labor, at 32 weeks of gestation.
Patient characteristics are presented in Table 1, with the severe acute maternal morbidity rate by maternal characteristics.
In the sensitivity analysis excluding severe acute maternal morbidity related to hypertensive complications, results were similar and notably non-severe preeclampsia remained a risk factor of severe acute maternal morbidity with a similar estimate for adjusted relative risk (S3    Table). In addition, CART analysis showed that non-severe preeclampsia remained the most discriminating factor for identifying women at higher risk of severe acute maternal morbidity (S1 Fig).  Risk factors for severe maternal morbidity in twin pregnancy The sensitivity analysis including delivery-related characteristics identified the same risk factors as above. In addition, caesarean delivery was associated with a significantly higher risk of intra-or postpartum severe acute maternal morbidity, whether performed before (aRR, 1.3, 95% CI, 1.0-1.6) or during labor (aRR, 1.6, 95% CI, 1.2-1.9). Finally a gestational age at birth less than 37 weeks of gestation was associated with a significantly lower risk of intra-or postpartum severe acute maternal morbidity (aRR, 0.8, 95% CI, 0.7-0.9) (S4 Table).

Main findings
In women with twin pregnancies, the overall increased risk of developing severe acute maternal complications varies by subgroups of women. At the beginning of twin pregnancy, nulliparous women with oocyte donation were identified as those with the highest risk of severe acute maternal morbidity. At the end of pregnancy, two subgroups of women with a risk exceeding 25% of developing severe acute maternal morbidity were identified: those with both oocyte donation and non-severe preeclampsia, and nulliparas born in sub-Saharan Africa with nonsevere preeclampsia.

Strengths and limitations
Our study has several strengths. It was a large population-based study including a substantial number of both twin pregnancies and of cases of severe acute maternal morbidity. The analysis of severe acute maternal morbidity was planned during the design of the JUMODA study, so that the data to characterize it were defined in advance and collected prospectively. The data collection method involving a manual review of the medical records of each woman provided detailed and accurate information on maternal and pregnancy characteristics, unlike data extracted from routine hospital databases in retrospective studies.
One limitation of this study is that the JUMODA cohort included only maternity units with more than 1500 annual deliveries. Although the twin deliveries included in the JUMODA study accounted for 75% of all twin deliveries in France over the study period, this could potentially mean that our results cannot be fully generalized to women giving birth in the smallest hospitals. However, the incidence of severe acute maternal morbidity we found here is similar to the one reported in another French population-based study including units of all sizes. 3 Furthermore, despite more than 8800 women were included in this study, its statistical power for covering rare pathologies that might be severe acute maternal morbidity risk factors remains limited. Finally, the results of CART analyses would benefit from a validation in another cohort of twin pregnancies and are limited by the fact that not all clinical situations are represented in the final tree.

Interpretation
The most original result-and perhaps the most helpful for clinicians-is the characterization among this population of subgroups with various levels of risk of severe acute maternal morbidity, by CART analysis. This analysis is complementary to the standard epidemiologic study of risk factors by multivariate regression, but applies an approach that is more pragmatic than explanatory. By directly estimating the severe acute maternal morbidity rate in subgroups that combine multiple risk factors, it may help clinicians to advise women on the most appropriate place for delivery and to adapt their management throughout pregnancy by anticipating the occurrence of those adverse events.
At the beginning of pregnancy, we found that maternal birth in sub-Saharan Africa, preexisting insulin-treated diabetes, nulliparity and mode of conception were risk factors for severe acute maternal morbidity. Although these results are concordant with previous studies conducted in general populations of parturients or populations of singleton pregnancies, it is important to verify that these risk factors persist in this high-risk population [11][12][13][14][15][16][17]. CART analysis found that at the beginning of pregnancy oocyte donation is the most discriminating factor for severe acute maternal morbidity. This increased maternal risk in women with twin pregnancies after IVF contradicts some previous studies exploring this association [9,[18][19][20]. The latter, however, were limited by their inability to differentiate between spontaneous pregnancy and ovulation-inducing drugs alone and between in vitro fertilization with autologous and donated oocytes.
An unexpected result was the lack of association between severe acute maternal morbidity and maternal age. Wondering whether this result might be due to overadjustment when mode of conception and age were simultaneously included in the model, we conducted an analysis stratified by mode of conception. It confirmed the lack of association, regardless of the mode of conception (data not shown).
At the end of pregnancy, the CART analysis highlighted that even in subgroups of twin pregnancies with the lowest risk, the rate of severe acute maternal morbidity still remains twice the rate usually reported in singleton pregnancies [3,[9][10][11][12]. We found that the most discriminating factor for the occurrence of severe acute maternal morbidity was non-severe preeclampsia, which has been previously reported as a risk factor for severe acute maternal morbidity in general populations of parturients and in singleton pregnancies [21,22]. We confirm here that this association also exists among twin pregnancies. This result might appear trivial, since non-severe preeclampsia could be considered just a step within the morbidity continuum of hypertensive-related complications. Nonetheless, it is noteworthy that when we excluded severe acute maternal morbidity events due to hypertensive complications from the outcomes, results of multivariate and CART analyses remained similar. This suggests that non-severe preeclampsia is a risk factor for other causes of severe acute maternal morbidity than hypertensive complications. Moreover, the CART analysis provided here shows that among women with twin pregnancies and non-severe preeclampsia, additional risk factors boost the risk of a severe acute maternal morbidity up to rates above 25%.
The identification of some modifiable risk factors can help improve the management of women with twin pregnancy. Preventive care can start during the preconceptional period. The high maternal risk in twin pregnancies after in vitro fertilization constitutes an argument for limiting the number of embryos transferred to a single embryo in order to prevent medicallyinduced multiple pregnancies and associated severe maternal morbidity. This is particularly important in case of oocyte donation, which must be a reasoned practice. An increased risk of severe acute maternal morbidity associated with maternal country of birth, another identified risk factor, may also be indirectly modifiable. Some previous studies have reported an increased risk of severe hypertensive complications in migrant women, in particular from sub-Saharan Africa [11,23,24]. Among suggested hypotheses are genetic factors [25] but also differential prenatal care [26,27,28]. Thus if the association between maternal country of birth and severe acute maternal morbidity reflects differential prenatal care, this risk factor could be changed by improving the quality of care for these vulnerable women.

Conclusion
About one in 17 women with a twin pregnancy will develop severe acute maternal morbidity overall, and this proportion rises up to more than a quarter in some particular subgroups of women. These results have implications for clinical practice. They will help identifying modifiable risk factors, personalizing information and improving shared decision regarding prenatal and delivery care for women with twin pregnancies, according to their individual profile.
Supporting information S1