Introduction

There are approximately 300,000 maternal deaths in the world annually [1], and in a recent population-based study, we have highlighted 18 episodes of severe maternal morbidity (i.e., a potentially life-threatening condition) per 1000 deliveries in Canada, increasing 1.3% annually for the last one decade [2]. However, severe maternal morbidity can often be anticipated according to known maternal risk factors, and outcomes possibly improved with early recognition and timely supportive acute and intensive care [3,4,5,6,7,8].

Variability in both recognition of critical illness and access to intensive care units (ICUs) influences the risk of death in many critical care syndromes [9, 10]. For example, sociodemographic-based variation and health care provider volume of experience have been shown to be important determinants of outcome for patients with acute lung injury, ischemic cardiac disease, and traumatic injury [11,12,13,14]. We therefore aimed to describe variation in care for acutely ill pregnant and postpartum women in Canada—a country with a vast geography and provincially administered health care—and to investigate the relationship between patient factors and hospital-based variability in ICU admission practices and clinical outcomes.

Methods

Study design, data sources, and population

We conducted a nationwide population-based cohort study in Canada spanning the years from 2004 to 2015, using data from the Discharge Abstract Database (DAD) of Canadian Institute of Health Information (CIHI), which includes information from every admission to acute care hospitals across Canada, except the province of Quebec, which maintains a separate database.

Patients in this cohort were hospitalized pregnant and postpartum women within 6 weeks after the end of pregnancy, including hospitalized women with abortion or ectopic pregnancy. Available variables from the DAD included the following: age, DAD-derived variables (socioeconomic status, comorbid conditions as classified by the International Classification of Diseases and Related Health Problems-Tenth Revision, Canada [ICD-10-CA], interventions and procedures occurring in hospital as determined by the Canadian Classification of Health Interventions [CCI]), pregnancy-specific variables, hospital-level variables, and patient outcomes such as ICU admission, vital status at hospital discharge, and events allowing determination of severe maternal morbidity (Additional file 1: Table S1).

The dataset (the DAD of CIHI) has been well-validated in a number of prior studies [15,16,17,18,19,20]. In addition, the institute annually reports data quality documentation, to have transparency of data quality. An episode of pregnancy was created as previously described [2], so that all related admissions over one pregnancy course could be summarized into one pregnancy episode.

Outcomes

The primary outcome was admission to ICU which included all admissions to CIHI-derived intensive (special) care units, including “step-up”/“step-down” ICUs [21]. Secondary outcomes included a modified Canadian definition of severe maternal morbidity and maternal death [22, 23]. In order to improve the specificity of the outcome of severe maternal morbidity for this analysis, the definition was modified to exclude events consisting only of a blood transfusion and diagnoses reflecting only pre-existing conditions such as HIV, chronic congestive failure, chronic hypertension, and those with an admission-related (as opposed to outcome-related) specific diagnosis [e.g., asthma, placenta previa, severe pre-eclampsia, or sickle cell anemia] [2, 8, 22]. Maternal death was defined as death occurring during pregnancy or within 6 weeks after the end of the pregnancy.

Predictors

A hospital-level predictor was created to represent pregnancy-related admission volumes at each hospital. First, hospitals with fewer than 55 deliveries over the study period (approximately corresponding to fewer than 5 deliveries per year) were excluded using a commonly applied principle of limiting analyses and inferences about data containing fewer than 5 observations in a particular time period (in this case, 1 year). Next, hospital quintiles were created according to pregnancy admission volume at each hospital. Other ordinal predictors (e.g., 5 groups according to ICU admission rate and 5 groups according to the annual number of pregnancy admissions at each hospital) were created in order to perform hospital-related sensitivity analyses.

Comorbid conditions necessary for determining the maternal comorbidity index (MCI) were recorded for each patient, as we have previously determined it to be the most valid risk adjustment tool using health administrative data for hospitalized pregnant and postpartum women [24, 25]. The index ranges from 0 to 45, excluding the age-related variable. Age was removed from the index because the age of the index does not include age younger than 35 years, but age and the index were separately entered into the model. Since six conditions of the maternal comorbidity index had potential overlap with severe maternal morbidity outcome conditions, only the pre-existing conditions were included in the maternal comorbidity index, and those occurring after pregnancy onset were considered as outcomes for severe maternal morbidity, to ensure a valid temporal relation [2].

Statistical analyses

We used the mean and standard deviation or median and interquartile range to summarize continuous variables, and counts and proportions to describe categorical variables, as appropriate. Standardized differences were employed to compare those who required ICU admission with those who did not, where 0.2, 0.4, and 0.6 standardized differences were generally equated to small, medium, and large effect sizes, respectively [26]. All variables were screened on the basis of standardized difference and selected for regression models if the standardized difference was greater than 0.1, in addition, to an a priori decision to include the urban or rural location of patient residence and hospital location. To develop random intercept logistic regression models with patient and hospital levels, first, we explored an intercept-only model predicting the outcome of ICU admission with each hospital considered as a cluster (representing differences in ICU admission rates among various hospitals). Then, we added patient-level variables to the intercept-only model and finally hospital-level variables [27, 28].

We used the median odds ratio (MOR) to represent the variability of the outcome (e.g., ICU admission) among the hospital clusters of interest [27, 29]. For example, when the MOR = 2, it means that for two pregnant women with the same characteristics, the odds of being admitted to ICU may be twofold higher in one randomly selected hospital compared to another randomly selected hospital from a different cluster when the clusters were ordered by risk. The MOR was computed from the following equation: \( \mathrm{MOR}=\exp \Big(0.675\times \sqrt{2\times \mathrm{hospital}-\mathrm{level}\kern0.34em \mathrm{variance}} \), where the hospital-level variance of the parameter estimate (i.e., β) is derived from the random intercept model. Interval odds ratio-80 (IOR-80) was used to quantify the impact of a variable at the hospital cluster-level on the outcome (e.g., ICU admission) by incorporating between-cluster variability [27, 29]. IOR-80 covers 80% of the distribution of the odds when comparing two subjects with the same value for all other covariates, but in two different hospitals where hospital-level characteristics may differ. We calculated the lower and upper bounds of the IOR-80 from the following equations: \( {\mathrm{IOR}}_{\mathrm{lower}}=\exp \left(\beta -1.2816\times \sqrt{2\times \mathrm{hospital}-\mathrm{level}\kern0.17em \mathrm{variance}}\right) \) and \( \mathrm{IO}{\mathrm{R}}_{\mathrm{upper}}=\exp \left(\beta +1.2816\times \sqrt{2\times \mathrm{hospital}-\mathrm{level}\kern0.17em \mathrm{variance}}\right) \), where β is the regression coefficient for the hospital level. Note that the IOR does not describe the interval around a MOR; instead, the IOR reflects the size and variation of the odds ratios at the hospital level. If the IOR-80 is wide, it implies high variability in the outcome among hospitals compared to the effect of the cluster level variable. If the interval of IOR-80 crosses 1, it infers that a variable at the hospital cluster level is likely not significantly associated with the outcome. Finally, the variance partitioning coefficient (VPC) represents the proportion of total variability in the outcome (e.g., ICU admission rates) due to the differences among hospital clusters. For multi-level logistic regression models, VPC is computed from the following equation: VPC = hospital-level variance/(hospital-level variance + 3.29) [30].

Sensitivity analyses

Two pre-planned sensitivity analyses were conducted. First, we explored the association of hospital volume of pregnancy-related admissions as a hospital-level ordinal variable comprised of five categories, on the outcome of ICU admission. In the second sensitivity analysis, we restricted the primary outcome of ICU admission to only admissions to the highest acuity ICUs, excluding step-up/down units.

The third, post hoc, sensitivity analysis was performed where all analyses were redone using multiple imputation datasets, given the rate of missing data in variables reflecting socioeconomic status (Table 1). Missing variables were assessed based on an algorithm reported elsewhere, to determine characteristics of missingness (i.e., missing completely at random, missing at random, or missing not at random) [31]. We used listwise deletion (entire record excluded from analysis if any single value is missing) for the primary analysis. This method is valid if the data is missing completely at random [32]. As a sensitivity analysis, to investigate the influence of data missing not completely at random, we performed multiple imputation using the fully conditional method with 20 imputed datasets [33, 34]. We imputed missing variables using a fully conditional specification and the discriminant function method to impute the categorical variables: income quintile (14.0% missing), residence rurality (0.7% missing), and hospital rurality (0.6% missing). We created the 20 multiply imputed datasets as this generally provides good efficiency of the parameters (i.e., smaller standard error of the point estimates), in addition to sufficient power to detect differences [35,36,37].

Table 1 General characteristics of pregnant and postpartum women according to whether admitted to ICU or not
Full size table

All analyses were performed using SAS statistical software, version 9.4 (SAS Institute Inc., Cary, NC) and Excel for Macintosh, version 15.3.9 (Microsoft Corporation, Redmond, WA). All analyses were completed on September 13, 2018. This study was conducted under the data security and privacy policy of CIHI and approved by the institutional review board at Mount Sinai Hospital in Toronto, Canada (10-0305-C) and University of Toronto (#34468).

Results

Incidence and clinical characteristics of pregnancy-related ICU admissions in Canada

During the study period, there were 3,162,303 pregnancies among 2,035,453 mothers, resulting in 10,204 ICU admissions (3.2 cases per 1000 pregnancies) with regional variation (Fig. 1, Table 1). There were 50,012 severe maternal morbidity events (15.8 women with a severe maternal morbidity event per 1000 pregnancies), with 5017 (0.16% overall and 10.0% of those with severe maternal morbidity) pregnant women admitted to ICU (Tables 1 and 2.). A minority of patients experiencing severe maternal morbidity were admitted to ICU (Additional file 2: Table S2, Additional file 3: Table S3). During the study period, there were 195 maternal deaths (6.2 per 100,000 pregnancies (maternal mortality rate 0.01%)); however, only 130 (67%) of these patients were admitted to an ICU prior to death (Table 2). After excluding hospitals with, on average, fewer than 5 deliveries per year, the study cohort consisted of 3,157,248 pregnancies among 342 hospitals (Additional file 4: Table S4, Additional file 5: Table S5).

Fig. 1
figure 1

Incidence of ICU admission by province in Canada. Figure legend: territories cover Yukon, Northwest Territories, and Nunavut

Full size image
Table 2 Severe maternal morbidity diagnoses: all patients with severe maternal morbidity, those admitted to ICU, and those dying
Full size table

Patients admitted to an ICU had a higher pre-existing maternal comorbidity index and a resulting greater burden of severe maternal morbidity (Table 1). Among all severe maternal morbidity events, postpartum hemorrhage requiring transfusion and sepsis were the most common (Table 2). Among patients admitted with acute respiratory distress syndrome, cerebral edema or coma, eclampsia, or sepsis, only 76.2%, 57.8%, 21.5%, and 4.7%, respectively, were admitted to ICU. Patients dying in hospital, without admission to ICU, included among others those with some form of cardiovascular compromise (56 patients: obstetric embolism, myocardial insufficiency, or shock), hemorrhage (8 patients), and sepsis (3 patients) (Additional file 6: Table S6).

Hospital volume of pregnancy admission

We created 5 hospital groupings according to the volume of pregnancy admissions during the study period. Hospitals admitting the fewest number of pregnant patients had the highest incidence of severe maternal morbidity (20 cases per 1000 pregnancies) and had the highest maternal mortality (40 cases per 100,000 pregnancies) (Additional file 4: Table S4), despite having similar pre-existing comorbidities. Hospitals admitting the fewest number of pregnant patients were also more likely to be in rural areas and had a higher proportion of patients in lower income quintiles. There were geographic differences in the proportion of pregnancies admitted at low-to-high-volume hospitals, with Ontario having the greatest proportion of pregnant women admitted to high-volume hospitals (Additional file 4: Table S4, Additional file 5: Table S5).

Multi-level logistic regression models

The following were entered into the multi-level logistic regression models: patient-level variables (age, maternal comorbidity index, parity, residence location, transfer status for admission) and hospital-level variables (pregnancy-related admission volume, province and territories, hospital location) (Table 3, Additional file 1: Table S1, Additional file 7: Table S7). Independent patient-level factors associated with ICU admission included maternal comorbidity index (1.88, 95%CI 1.86–1.99), urban residence (1.09, 95%CI 1.02–1.16), and lower incomes. The lower income they had, the higher risk of admitting to ICU (i.e., the lowest income had the highest risk of ICU admission: 1.43 with 95%CI 1.33–1.54). Hospital-level factors associated with ICU admission included specific provinces (Table 3, Additional file 1: Table S1, Additional file 7: Table S7).

Table 3 Estimated odds ratios for multi-level logistic regression model for intensive care unit admission
Full size table

Variability

Although the overall ICU admission rate was 3.2 cases per 1000 pregnancies, the range in ICU admission variability among all hospitals was wide—from 0 to 500 cases per 1000 pregnancies (Additional file 4: Table S4). There was a minimal variability in ICU admission rate at the provincial level (Additional file 5: Table S5). Using the median odds ratio, for 2 pregnant women with similar characteristics at different hospitals, the average (median) odds of being admitted to ICU was 1.92 in 1 hospital compared to another (Additional file 7: Table S7). The proportion of variability in ICU admission due to hospital-level factors, as denoted by the variance partitioning coefficient, was 33.3%, with substantial variation in the odds of ICU admission across hospitals (Additional file 7: Table S7). Furthermore, ICU admission rate for postpartum readmissions was relatively high (36.5 per 1000 pregnancies) in comparison with other periods of admissions, and ICU admission among abortion-related admissions had the highest variability (median odds ratio, 2.21) (Additional file 8: Table S8).

Sensitivity analyses

All primary and secondary results were similar in sensitivity analyses, employing the number of pregnancy-related hospital admission in continuous variable (Additional file 9: Table S9), considering only the highest acuity ICUs (Additional file 10: Table S10), and using multiple imputation datasets to account for missing data (Additional file 11: Table S11).

Discussion

In this study, we found that only 10% of the 15.8 women per 1000 pregnancies in Canada who experience a severe maternal morbidity event and only two thirds of the 6.2 women who die per 100,000 pregnancies are admitted to ICU. We also found wide hospital-level variability in ICU admission practices. Patients living in urban settings and of lower socioeconomic status were more likely to be admitted to ICU (Table 4).

Table 4 Box summary of key points
Full size table

Discovering variability in clinical practice for pregnant and postpartum women is important as it presents an opportunity for improved quality of care. For example, identifying large variability in preoperative medical consultation has led to efforts to improve perioperative anesthetic care [38]. Systematic literature search has revealed that ours is the first population-based study to examine the variability of the care of critically ill pregnant and postpartum Canadian women and highlights modifiable factors in ICU admission practices (Table 5, Additional file 12: Table S12). Prior population-based studies of pregnancy-related ICU admissions in addition to ours show variability in ICU admission rates among identified countries, ranging from 3.2 to 39 per 1000 pregnancies (Table 5) [3, 39,40,41,42]. While greater maternal comorbidity was associated with increasing likelihood of ICU admission, a minority of patients with sepsis and eclampsia were admitted to ICU. It may be due to early recovery from a sick condition after the delivery in case delivery was the first step of management and/or different pathway of patient care for pregnant women. In other words, first being admitted to obstetrical general wards may delay an acute care. Although only two thirds of all women who subsequently died were admitted to ICU, some conditions of severe maternal morbidity (e.g., obstetric embolism) and subsequent death could occur very quickly and limit the potential for ICU admission (Additional file 6: Table S6), which were indicated by an additional analysis of maternal death without ICU admission (Additional file 13: Table S13). Maternal death without ICU admission happened to those who were less sick (i.e., less maternal comorbidity index) at the time of hospital admission and less likely to transfer to another institution for acute care, but more likely to develop a life-threatening condition (i.e., severe maternal morbidity) in urban hospitals, compared to maternal death with ICU admission. There were 15 maternal deaths in operating rooms without ICU admission, related to postpartum hemorrhage and obstetric embolism. On the other hand, our results found that other conditions that progress more slowly may inherently have a greater time period for clinicians to consider admission to ICU, and hence a greater potential for intervention and improved outcomes (e.g., sepsis) through quality improvements of implementation of sepsis bundles, for example [43]. These findings highlight a potential opportunity for earlier recognition by clinicians for women at risk and earlier recognition of the onset of critical illness. More practically, in comparison with MCI = 0, increasing maternal comorbidity index has higher odds of admitting to ICU (index = 1, odds ratio (OR) 1.90, 95%CI 1.79–2.02; index = 2 or higher, OR 8.67, 95%CI 8.24–9.12) (Additional file 14: Table S14). Our findings also have relevance to policymakers. Women living in rural and resource-limited locations may experience the greatest risk [2], highlighting additional opportunities for identification of those at risk, with potential solutions of a lower threshold for consultation, referral and increased frequency of follow-up care, outreach of tele-medicine [44, 45], ease of transportation for mothers to specialist care, and/or consideration of a mobile ICU to resource-limited hospitals [46].

Table 5 Research in context. Population-based studies of pregnancy-related ICU admissions
Full size table

Strengths and weaknesses of the study

Our study has several strengths. First, this is the first study based upon a nationwide cohort to explore hospital variability of pregnancy-related ICU admission using multi-level logistic regression models accounting for hospitals as clusters. Second, we adopted the perspective of an episode of care, to account for the entire pregnancy and postpartum course rather than one admission allowing use to determine outcomes from the entire pregnancy instead of outcomes relating to individual hospital admissions. Third, the maternal comorbidity index was well-validated risk adjustment tool in the DAD of CIHI, to account for known risk factors of maternal death and organ injury [24, 25, 47]. Fourth, maternal age younger than 20 years old and over 35 years old was associated with maternal mortality, and hence properly adjusted in our analyses [8]. Fifth, we employed the best risk adjustment tool for our dataset among existing risk prediction models including ICU admission scores [24, 48] and ensured a valid temporal relationship between the MCI and severe maternal morbidity/ICU admission. Last, we have used existing and well-validated health administrative datasets to describe pregnancies as well as ICU practice across diverse regions and over time [15,16,17,18,19,20].

This study also has limitations. First, certain variables that may be influential for maternal outcomes were unavailable in our dataset, for example, patient characteristics such as body mass index and pregnancy-associated factors such as the use of assisted reproductive technology, among others. However, obesity did not remain in the maternal comorbidity index to account for maternal death and organ injury during the original development [47]. Even among available variables, and considering patient and hospital factors, our predictive models for ICU admission do not explain all variability, highlighting the existence of other potentially important factors. We had a high rate of missing data for certain variables, particularly postal code markers of socioeconomic status. However, there were few differences among known characteristics between groups with missing and non-missing data, and our results did not differ when we used a multiply imputed dataset in our sensitivity analyses. CIHI did not provide structural characteristics for each of all 342 hospitals enrolled in this study such as detailed type of ICUs and existence of neonatal intensive care unit. And hence, groups according to the annual volume of their delivery admission were created, to characterize enrolled hospitals from an obstetric practice point of view. Additionally, ICU utilization and transfer rate in each group of hospitals were also calculated (Additional file 4: Table S4). It is possible that some diagnostic coding error exists, for example, over-coding of ARDS among patients managed in post-anesthesia care unit or labor and delivery room and not the ICU (only 76.2% of patients with ARDS were admitted to ICU). Canadian Coding Standards Evolution Chronicle reports that post-procedural respiratory disorder was used to be coded as ARDS [49]. Over-coding of ARDS may be attributed to this previous coding procedure. Probably, contemporary validation studies of ARDS by utilizing audits of hospital chart is required for future large epidemiological studies, since there was no such specific validation study of ARDS using CIHI DAD after the Berlin definition in 2012. Although the DAD does not contain specific information about the number of ICU beds in each acute care hospital, a previous study estimated the number of ICU beds per 100,000 population, by provinces, in 2009, ranging from 5.5 to 19.3 beds per 100,000 population [50]. Given the provincial ICU beds per 100,000 population in Canada, our study showed that there was a mostly linear relationship between provincial ICU capacity and maternal ICU admission rate (r2 = 0.74, p = 0.002) (Additional file 15: Figure S1), which implies that ICU capacity may influence the admission patterns of pregnancy-related ICU admission as well in Canada. Also, in this analysis, we did not explore neonatal outcomes—this will be an important next step. Finally, although CIHI collects data on each admission at all acute care hospitals across Canada, our findings may not generalize to the province of Quebec which maintains a unique database.

Unanswered questions and future research

We found that only a minority of women who experience severe maternal morbidity are admitted to an intensive care unit in Canada, that there is a wide variability in ICU admission practice across hospitals and provinces, and that the greatest vulnerability exists for women at low socioeconomic status. These represent potentially modifiable factors for patients and should be the focus of quality of care improvement initiatives for clinicians and health systems.

Conclusions

In conclusion of this study, we found that a minority of women who experience a severe maternal morbidity event and only two thirds of those who die are admitted to ICU. There is a wide hospital-level variability in ICU admission practices and that those living in urban locations and those of the lowest socioeconomic status were most likely to be admitted to ICU.