Introduction

Sepsis is defined as organ dysfunction resulting from infection, and is a critical condition with high mortality rates.1 Global estimates, based mainly on data from selected high-income countries, and extrapolated for low-income countries, suggest that there are approximately 32 million sepsis cases annually, 19 million of which are severe. These cases result in an estimated 5 million deaths worldwide.2 The Canadian Institute for Health Information (CIHI) has estimated that there are approximately 30,000 sepsis hospitalizations annually in Canada,34 while a 2011 study of vital statistics data showed that 1 in 18 deaths among Canadians involved sepsis as the main or underlying cause.5 Survivors of sepsis are often left with long-term physical and cognitive deficits, and reduced quality of life,6 contributing to increased cost, extended hospitalization, and intensive post-discharge healthcare utilization.7 Despite the burden of sepsis, public and political recognition of the disease is limited.8,9 As a result, the World Health Organization recognized sepsis as a global health priority in 2017 and adopted a resolution on improving the prevention, diagnosis, and management of sepsis.10 They recommended that the development of monitoring and evaluation tools as well as epidemiological surveillance systems should be the cornerstones of the sepsis strategy.

There are a number of important challenges with sepsis surveillance, even in high-income countries such as Canada.2 At the population level, a practical and cost-efficient method to collect information about sepsis hospitalizations and mortality is through retrospective review of hospital records using International Classification of Disease (ICD) coding strategies. Nevertheless, hospital records alone contain incomplete socio-demographic information, limiting their utility for epidemiological surveillance.11,12,13,14

Another method has been to use ICD coded death certificates to investigate clustering of sepsis mortality by geographic area.15,16 While these studies have revealed geographic patterns in sepsis mortality in the United States, they were ecological in nature and could not directly investigate any associations between socio-demographic exposures and outcomes.15,16

An alternative approach to sepsis surveillance is to utilize linked population-based data. Specifically, population-based data sources, including the 2006 Canadian Census, have been linked to the CIHI Discharge Abstract Data (DAD) at Statistics Canada, to facilitate population health and health services research.17,18,19 The data collected in the long-form Canadian Census provide a rich multidimensional look at the Canadian population, including many important social determinants of health. Combined with clinical information available in the DAD, these data afford a unique opportunity to examine detailed socio-demographic characteristics of patients admitted to hospital with sepsis. Using Census-linked data provides antecedent information about sepsis patients in line with a paradigm where sepsis is conceptualized as a preventable public health problem.20 This is an important step towards identifying modifiable risk factors and high-risk groups to target future primary and secondary sepsis prevention strategies and includes considering population risk factors available through sources such as Census data.

The primary purpose of this study, achieved by linking data from the 2006 long-form Canadian Census with the CIHI DAD, was to identify socio-demographic factors in adults aged 18 yr and older, associated with admission to a Canadian hospital with sepsis. A secondary objective was to compare the outcomes in patients hospitalized for sepsis with those hospitalized for other reasons.

Methods

Study design

This study was a retrospective linked cohort design utilizing the 2006 long-form Canadian Census linked to the DAD from 2006/2007 to 2008/2009. This was the most current Census-DAD linkage completed at Statistics Canada. Approval for this linkage was granted by Statistics Canada’s Executive Management Board.21

Data sources

Census of Canada

The 2006 Census of Canada was administered by Statistics Canada on May 16th 2006 and collected information using both short- and long-form questionnaires.22 Eighty percent of households received the short-form, containing eight questions about birth date, sex, marital status, and language spoken by all household members. Twenty percent of households received the long-form, containing 53 questions on topics including income, education, employment, ethnicity, visible minority status, immigrant status, and functional limitations. The 2006 Census of Canada represented 95–97% of the non-institutionalized population of the Canadian provinces and 93–94% of the non-institutionalized population in the territories.23

Discharge abstract database

CIHI compiles the DAD, which is a national database containing demographic, administrative, and clinical data on hospital discharges across Canada, excluding Quebec. CIHI receives data directly from all public hospitals in every province and territory covered. Re-abstraction studies have found that non-clinical data elements are of a high quality and that the clinical data captured by the DAD are consistently reported.24,25 The DAD clinical data include 25 fields of ICD version 10 coded diagnoses and 20 fields of Canadian Classification of Interventions coded procedures.

Linkage methods

An hierarchical deterministic linkage approach was used to link Census respondents to the DAD from 2006/2007 to 2008/2009. The linkage process has been described in detail and validated by Rotermann et al., who found that the linked file represented approximately 80% of hospitalizations during the time period.17 The 2006 Census linked to the DAD cohort has previously been used for research.18,19

Study sample

The study sample consisted of adult patients 18 yr and older admitted to Canadian hospitals between 2006/2007 and 2008/2009 (excluding Quebec), who successfully linked to the 2006 long-form Canadian Census. Patients admitted with sepsis (defined below) were compared with those admitted for all other diagnoses, excluding patients admitted to give birth. Patients in the study sample could have been hospitalized more than once; for analysis, only one record per patient was retained. For those admitted with sepsis, the first hospital record containing a diagnosis of sepsis was retained. For those admitted for reasons other than sepsis, the first hospital record was retained. The eFigure available as Electronic Supplementary Material (ESM) shows the study flow chart.

Defining sepsis and severe sepsis

Sepsis was defined using an existing administrative data definition developed by CIHI3,4 and based on previous studies by Angus et al. and others.26,27,28 A patient was identified as having an admission for sepsis if any one of the ICD-10-CA codes listed in eTable 1, available as ESM, were present in any of the 25 diagnosis coding fields of the DAD record. Severe sepsis was indicated by the combination of a sepsis code with at least one organ dysfunction code (see eTable 1 available as ESM).

Primary outcome

Hospital admission with a diagnosis of sepsis was the main outcome of interest in this analysis.

Secondary outcome

In-hospital death among patients admitted with sepsis was the secondary outcome of interest. Other secondary outcomes of interest included admission to a special care unit, special care unit length of stay, hospital length of stay and discharge disposition (see eTable 2 available as ESM for a more detailed description).

Exposures

The covariates are described in detail in eTable 2 (available as ESM). The exposures studied in relation to hospital admission with a diagnosis of sepsis were Census characteristics, including age, sex, income, education, marital status, visible minority status, immigrant status, employment, and location of residence. We also examined Indigenous identity. The 2006 Census gathered information on whether a person self-identified with Aboriginal peoples of Canada. In this document, the term “Indigenous” is used in place of “Aboriginal”. We further categorized identity by a previously used breakdown18 in descriptive analysis (First Nations people living on reserve, First Nations people living off reserve, Métis, Inuit, and other or multiple Indigenous identity). For the calculation of rates and multivariable modelling “Indigenous identity” was used as a dichotomous variable because of small sample sizes in some of the groups (see eTable 2 available as ESM for more details).

These Census variables were chosen for study because they provide a broad description of the socio-demographic characteristics of the Canadian population and are not typically available in health administrative data. Many of these variables have been identified as important social determinants of health for the Canadian population. In addition, the population groups identified here, such as visible minorities, immigrants, and Indigenous populations, may experience a greater burden of health problems because of inequalities in the social determinants of health.29,30

Statistical analysis

Missing values in the Census data were imputed during processing.23 The DAD records were screened for missing postal code, birthdate, and sex in advance of being made available for linkage, and approximately 0.8% of records were excluded. None of the DAD records linked to the Census data were missing the main diagnosis field. Descriptive statistics (proportion with 95% confidence interval [CI] and medians [interquartile ranges (IQR)], as appropriate) were used to compare patients hospitalized with and without a sepsis diagnosis. Crude and age-standardized hospitalization rates for sepsis were calculated by Census characteristics. Estimates were directly age-standardized to the 2006 population.

Multivariable logistic regression models were used to investigate the association between Census characteristics and hospital admission for sepsis. Variables were not individually screened for inclusion in the multivariable models; all Census variables considered important a priori were entered into the model. Variables were not eliminated from the model if they were found to be non-significant. Initially sex-specific models were constructed but showed very similar patterns of associations among men and women, so sex was included as a covariate in the models. A model with interaction terms between age and Census characteristics found strong evidence of effect modification by age, and as a result three age-specific models were run: model 1 for 18–34 yr old; model 2 for 35–64 yr old; and model 3 for 65 yr and older. The age groups were chosen to enable the clinical interpretability of the results. A detailed description of the other variables included in the models can be found in eTable 2 (available as ESM).

Multivariable-adjusted age-specific mortality rates were also calculated for patients hospitalized with and without a sepsis diagnosis. These models were adjusted for additional socio-demographic and clinical variables including sex, income, education, marital status, visible minority status, Indigenous status, immigrant status, difficulties with activities of daily living, labour force participation, Charlson score, severe sepsis status, and special care unit stay.

In the present study, all estimates were calculated using the Census weights. The long-form Census is a 20% sample of the Canadian population and as such may be subject to sampling error, therefore use of the Census weights is recommended for analyses of these data. Using the Census weights allows us to make inferences about the Canadian population by compensating for unequal probabilities of sampling and non-response and adjusting the weighted sample distributions for key variables of interest (for example age, sex, and geography) to make it conform to a known population distribution.31,32 The Census weights were not adjusted for linkage eligibility. All analyses were conducted using SAS 9.3 (SAS Institute, Cary, NC, USA) and Stata 13 (StataCorp, College Station, TX, USA).

Results

Creating the population-based cohort

Of the 4,652,700 individuals that had a completed 2006 long-form Canadian Census and were linked to the DAD, 3,433,900 individuals remained in our population-based cohort after excluding Census and hospitalization records in individuals less than 18 yr of age, hospitalizations occurring outside the study period, hospitalizations for births, and non-unique hospitalizations. Of the 3,433,900 individuals aged 18 yr and older, 10,400 were hospitalized with sepsis between 2006/2007 and 2008/2009 (see the eFigure available as ESM).

Socio-demographic and clinical characteristics

These respondents represented a weighted count of approximately 49,000 Canadians from all provinces and territories in Canada, excluding Quebec. Forty-four percent of patients admitted to hospital with sepsis were women compared with 52.71% admitted to hospital for all other reasons (excluding birth-related admissions). The median age of patients admitted with sepsis was 68 yr compared with 60 yr for patients admitted for other reasons. Further socio-demographic and clinical characteristics of respondents admitted to hospital with and without sepsis, as well the overall population-based cohort, are detailed in Table 1 and eTable 3 (available as ESM).

Table 1 A summary of socio-demographic and clinical characteristics of 2006 long-form Census respondents hospitalized for sepsis, compared with those hospitalized for other reasons, as well as the overall population-based cohort (including those not hospitalized), 2006 Census-Discharge Abstract Database study cohort, Canada excluding Quebec, 2006/2007 through 2008/2009
Full size table

Rates of sepsis hospitalization per 100,000 population

The overall average annual age-standardized rate of sepsis hospitalization in Canada was 96 cases per 100,000 population in the years 2006/2007–2008/2009. Severe sepsis cases comprised almost 40% of those cases with an average annual rate of 37 cases/100,000 population. Overall, rates of sepsis hospitalization increased with age and were higher among men than women, while age-standardized rates varied across Census characteristics; notably, rates increased with decreasing income quintile and lower level of education. Sepsis hospitalization rates were higher among Indigenous individuals, as well as those reporting functional limitations and those not in the labour force (Table 2).

Table 2 Crude and age-standardized rates of sepsis per 100,000 population by Census characteristics, 2006 Census-Discharge Abstract Database study cohort, Canada excluding Quebec, 2006/2007 through 2008/2009
Full size table

Census characteristics associated with sepsis hospitalization

The association between Census characteristics and sepsis hospitalization was examined in a series of multivariable models. Sex-specific models showed very similar patterns of associations among men and women, and therefore were not presented separately; however, sex was included as a covariate. A model with interaction terms between age and Census characteristics found strong evidence of effect modification by age, and as a result three age-specific models were run; model 1 for 18–34 yr old; model 2 for 35–64 yr old; and model 3 for 65 yr and older. Within all the age-specific models, the term for age continued to be statistically significant, indicating that with increasing age the odds of being hospitalized with sepsis increased. In all three age-specific models, men were more likely than women to be hospitalized for sepsis, with the strongest association observed in model 3 (≥ 65 yr; odds ratio, 1.52; 95% CI, 1.48 to 1.56) (Table 3).

Table 3 Socio-demographic factors associated with sepsis hospitalization compared with hospitalization for other reasons. Age-stratified logistic multivariable adjusted regression models, 2006 Census-Discharge Abstract Database study cohort, Canada excluding Quebec, 2006/2007 through 2008/2009
Full size table

In terms of Census characteristics, different associations emerged as statistically significant by age group. Among 18–34 yr olds, marital status, visible minority status, having functional limitations, and not being in the labour force were all factors associated with an increased odds of sepsis hospitalization (Table 3). Among 35–64 year olds, the same factors, in addition to income, education, Indigenous identity, and living in an urban location were associated with an increased odds of sepsis hospitalization (Table 3). In the oldest age group, all Census variables except income were associated with an increased odds of sepsis hospitalization (Table 3). Notably, being a member of a visible minority and having functional limitations had a significant impact on the odds of sepsis hospitalization in all three models with the effect size attenuated with increasing age. A similar pattern was seen for employment.

Length of stay and hospital outcomes among sepsis and non-sepsis patients

Patients hospitalized with sepsis were more often admitted to special care units (45.61% vs 11.97%) and had longer median lengths of stay in both special care units (5.42 vs 2.08 days) and hospital (10 vs 3 days) (Table 4).

Table 4 Hospital length of stay and outcomes of 2006 long-form Census respondents hospitalized for sepsis compared with those hospitalized for other reasons, 2006 Census-Discharge Abstract Database study cohort, Canada excluding Quebec, 2006/2007 through 2008/2009
Full size table

Among patients hospitalized with sepsis, there were 3,200 deaths in hospital, representing a weighted count of 15,665 (95% CI, 15,064 to 16,265) deaths across Canada, except Quebec. The estimated overall crude hospital mortality rate among those hospitalized with sepsis was 31.95% (95% CI, 31.54 to 32.37), compared with 3.03% (95% CI, 3.00 to 3.05) for those hospitalized for all other reasons (excluding birth-related admissions) (see Table 4). Crude mortality rates varied by age group, from 9.48% (95% CI, 6.42 to 13.79) among patients aged 18–34 yr to 39.31% (95% CI, 37.94 to 40.72) in patients aged 65 yr and older (Table 4). Multivariable adjusted mortality rates by age group are presented in eTable 4 (available as ESM) for comparison.

Discussion

We identified more than 10,000 patients aged 18 yr and older who were admitted to hospital with sepsis between fiscal years 2006/2007 and 2008/2009 (excluding Quebec), who were successfully linked to the 2006 long-form Canadian Census. This sample represented a weighted count of approximately 49,000 Canadians from all provinces and territories, excluding Quebec. The overall average annual age-standardized rate of sepsis hospitalization was 96 cases/100,000 population; of these almost 40%, or 37 cases/100,000 population, were classified as severe sepsis. The association of Census characteristics with sepsis hospitalization varied with age. In age-specific hospital admission models several characteristics were consistently associated with an increased odds of hospital admission with sepsis including increasing age, male sex, never being married, visible minority status, having functional limitations, and not being in the labour force. This study also showed that sepsis remains a considerable burden for individual patients and the healthcare system, with in-hospital mortality rates among patients admitted with sepsis approximately ten-times higher than for those admitted for all other reasons (except for birth-related admissions).

Census variables identified as risk factors for admission with sepsis in this analysis, would not otherwise have been identified, given the limited socio-demographic information available in administrative data sets. The breadth of the data collected in the Census allows for analysis across multiple social determinants of health, enabling the identification of risk factors and risk groups in which to target future primary and secondary sepsis prevention strategies.20 These findings lend support to the importance of variables measured in Census data (linked to hospitalization data), not only for population health surveillance of chronic conditions, but also for their potential importance in the surveillance of acute life-threatening critical illnesses like sepsis.

Our findings in relation to overall and severe sepsis rates are consistent with those produced by CIHI and Husak et al.,3,4 using the full DAD (excluding Quebec) over the same time period. Nevertheless, in a recent meta-analysis Fleischmann et al. found considerably higher rates of sepsis and severe sepsis in seven high-income countries (not including Canada).2 This can be explained by differences in study design, in particular the methods used to ascertain cases of sepsis/severe sepsis. Fleischmann noted that studies included in their meta-analysis used both clinical criteria and ICD9/10 code-based criteria to identify cases.2 This led to considerable heterogeneity in the estimates, which was also highlighted recently by both Jolley et al. and Mariansdatter et al.11,33 Both authors noted that, even in code-based analyses, ICD-9/10 coding schemes varied from study to study.11,33 The World Health Assembly also emphasized the issue of ICD coding by urging stakeholders “to apply and improve the use of the International Classification of Diseases system to establish the prevalence and profile of sepsis…”.10 While an expanded ICD-10 definition of sepsis has been developed in Canada,12 we wanted to benchmark our findings to those of CIHI and Husak et al.,3,4 as a validation of the Census-linked data. The benefit of the Census-linked data is that they can be used in combination with evolving ICD-10 (or future ICD versions) code-based definitions of sepsis. Indeed, the ICD-11 coding scheme has been developed by the WHO with the aim of providing far more granular information compared with the ICD-10,34 and will presumably improve case ascertainment of sepsis.

Benchmarking against previous national estimates, our analysis showed that the rates of sepsis varied by Census characteristics beyond age and sex. These characteristics were further investigated in age-stratified multivariable logistic models, and the association we found between visible minority status and sepsis hospitalization can be compared with several other studies that have examined sepsis incidence by ethnicity and race in the United States.28,35,36 Both Dombrovskiy et al. and Martin et al. reported variation in the incidence of sepsis hospitalization, with rates among non-whites almost double those of whites in population-based studies.28,35 A population-based study by Barnato et al. obtained mixed findings in relation to race; they found that black race was associated with higher incidence of sepsis, whereas Hispanic ethnicity was associated with lower incidence after adjustment for area-level poverty and rurality.36 In contrast, Moore et al.37 showed that the odds of first sepsis incidence was similar between black and white participants in the REasons for Geographic And Racial Differences in Stroke (REGARDS) cohort study.

In this study, employment, education, and income (albeit among 35–64 yr olds only), were also associated with an increased odds of hospitalization for sepsis. These findings are very similar to those of Storm et al.,38 who found, in a small Danish case-control study, that individuals with lower levels of education and those not in the labour force had significantly higher risks of intensive care unit admission for sepsis. As well, these findings are indirectly supported by those of Moore et al.,15 who showed that sepsis mortality was clustered in three major regions in the southern United States, which were more likely to have lower education levels, higher poverty levels, higher unemployment rates, and a high proportion of medically uninsured individuals. More recently, Donnelly et al.39 showed, with mediation analysis, that physical weakness, individual income, and comorbid diabetes may partially explain differential infection risks by neighbourhood socioeconomic status in the United States.

The association of Indigenous identity with increased odds of sepsis hospitalization in two of our age-specific models is consistent with findings by Carrière et al.18,40 who reported overall higher rates of hospitalization for all causes among those of Indigenous identity in Canada, using the same data source as our study. These authors noted that inequalities in social determinants of health may influence hospitalization rate disparities, because in many First Nations communities, municipal infrastructure is inadequate41; educational and employment opportunities are limited; and the prevalence of low income is high.42 As well, other Canadian studies have shown that both the incidence of critical illness in general, as well as severe trauma and injury are higher among Indigenous Canadians.43,44,45,46 Similarly in Australia, Davis et al.47 have shown higher rates of sepsis incidence, but lower mortality rates among Indigenous Australians in the tropical Northern Territories.

The results of this study, combined with the existing literature discussed above, indicate that socio-demographic characteristics, in addition to already established risk factors, are important determinants of sepsis incidence and hospitalization. These results are relevant for policy makers interested in developing data sources for sepsis surveillance as well as clinicians interested in addressing the distal determinants of sepsis and approaching sepsis as a preventable public health problem.20 This means conceptualizing sepsis as a problem that can be treated with population and system-based solutions, such as those suggested by Kempker et al., including management of risk factors, appropriate antibiotic prophylaxis, public awareness, hygiene, and immunization.20

Limitations

This study should be considered in the context of its limitations. The ascertainment of socio-demographic variables in the Census is cross-sectional. While the study period was relatively short, is it possible that individuals’ socio-demographic information changed between Census day and the date of hospitalization. Validation of the linked files used in this study showed that not all hospitalizations were represented.17 In particular, there was lower coverage of some populations, for instance among 15–24 yr olds, those living in the territories, and Indigenous populations, which may lead to under ascertainment of sepsis in these groups. As well, all Indigenous identities were grouped for analysis because of small sample sizes. This grouping may mask important differences between distinct groups of Indigenous people.

Hospitalization data were not available for the province of Quebec, which is the second most populous province in Canada. As well, the institutionalized population is not represented in the Census. Although the Census data are from 2006, it was the most recent Census-linked file available at the time of analysis; it is unlikely that the results would differ if more recent Census data were available because socio-demographic characteristics, as well as interrelationships between socio-demographic characteristics and health outcomes, are relatively stable over short periods of time. Nevertheless, it is possible that the case-mix of patients admitted to Canadian hospitals has changed over time. In addition, the accuracy of sepsis coding in the hospital administrative data may have changed because of increasing public and professional awareness of sepsis. Future analyses could take advantage of additional population-based data as well as electronic health record data that are becoming available for linkage, for future surveillance of sepsis and other critical illnesses.

Conclusions

Using linked population-based data, we were able to identify a large sample of Canadians who are hospitalized for sepsis. This data source affords a unique opportunity to examine in more detail the socio-demographic characteristics of patients admitted to hospital with sepsis. Such data could be helpful for epidemiological surveillance, as well as primary prevention, and can be used in combination with evolving administrative data definitions of sepsis.