Population and setting

This is a retrospective population-based cohort study of adults (≥ 18 years old) admitted to any of the 31 ICUs within 14 hospitals in the province of Alberta, Canada, between January 2015 and April 2018. During the study period, Alberta had a population of about 4.4 million. All ICUs in Alberta are governed by a single healthcare service provider, Alberta Health Services, within a publicly funded healthcare system. All patients admitted to any ICU (general, cardiac, cardiovascular or neurosciences) in Alberta, as indicated by an ICU electronic medical record, were included unless their ICU admission lasted less than 24 h, and they were not residents of Alberta (have a primary healthcare number from another province) or did not have follow-up data for at least 180 days after index hospital discharge.

Patients were excluded if they were not residents of Alberta because our case definition of individuals experiencing homelessness required Alberta postal codes. Homelessness is complex, nuanced and fluid; however, for the purposes of analyzing quantitative data, in this study individuals were considered to be experiencing homelessness if, during any of their hospital admissions during the study period, they did not have a fixed address (had postal code T1T1T1, which is used to identify individuals experiencing homelessness, had an International Classification of Disease, 10th revision (Canadian edition; ICD-10CA) code Z59.0 indicating the patient is experiencing homelessness) or had precarious housing (their postal code was that of established housing shelters in Alberta). This definition was adopted from previous work; a case definition using postal code-based (including those of shelters) algorithms found this method was 33% sensitive and 99% specific [31], while a more recent study also included ICD-10CA codes to increase sensitivity [32].

Data sources

All data used in this study were previously collected for administrative and clinical purposes by the data custodian, Alberta Health Services. Our cohort was identified using eCritical Alberta, a clinical information system that captures and delivers multimodal patient data to the bedside and is a repository for these data. These data include patient demographic, clinical and outcome data [33,34,35]. There are continuous data quality auditing to ensure data from eCritical TRACER is valid. [33]

The cohort was deterministically linked to additional data sources using a unique personal healthcare number and date of birth. These additional data sources were used to create a complete profile of participants during their hospital admission and after any hospital discharge that was associated with the ICU admission. These data sources included:

1.

The Discharge Abstract Database (DAD) contains demographic, diagnostic (up to 25 International Classification of Disease version 10, Canadian codes; ICD-10-CA, with an associated diagnosis type) [36], administrative and procedural data on patients discharged from the hospital.

2.

The National Ambulatory Care Reporting System (NACRS) collects and stores demographic, administrative, clinical and service-specific data from emergency departments (and other ambulatory care visits) including a complaint lists and emergency department discharge diagnoses using the Canadian Emergency Department Diagnoses Shortlist (800 diagnoses) mapped to ICD-10 codes are also collected [37].

Variables

The exposure variable was housing status (experiencing homelessness versus stable housing) using the definitions outlined above. The proportion of patients experiencing homelessness were measured across the study period (quarters: 1 = January to March, 2 = April to June, 3 = July to September, 4 = October to December) to assess temporal trends.

The primary outcome was healthcare resource utilization. In the absence of a single measure of healthcare resource utilization, several individual variables were used to measure healthcare resource utilization: (1) processes of care, (2) ICU and hospital length of stay (number of days or part of day from admission to discharge), (3) hospital readmission (binary variable within 30 days of hospital discharge) and (4) emergency room visits (within 30 days of hospital discharge). Processes of care in the ICU included: receipt of advanced life support interventions including invasive and noninvasive mechanical ventilation, vasoactive medications and continuous renal replacement therapy (CRRT), which were binary variables (received intervention or not) and time on advanced life support were continuous variables measured in minutes or hours for those receiving the interventions. While readmission variables were calculated using all available data, only data from the index ICU and hospital admission were used to calculate lengths of stay and processes of care variables.

The secondary variables were hospital adverse events, ICU mortality and hospital mortality. Hospital adverse events were measured using validated ICD-10-CA algorithms to identify 18 patient safety indicators [38]. Each of the 18 patient safety indicators was dichotomous—ever having a hospital adverse event or not—and the overall adverse event variable was created to dichotomously indicate the presence of any adverse event.

Patient variables included age (continuous variable), sex (dichotomized as male or female), ICU admission diagnosis (categorized as medical, surgical, neurological, trauma, cardiac surgical or non-surgical or unknown) and Charlson comorbidities (categorized as none, one comorbidity or two or more comorbidities). The reason for hospital admission (the most responsible diagnosis as defined by diagnosis type and categorized by the ICD-10-CA chapters) and any mental health-specific diagnosis related to the hospital admission (ICD-10-CA codes for depression, anxiety, substance use, psychosis, suicide or severe psychiatric disorders that were noted as the reason for hospital admission, comorbidities or that occurred in hospital) were identified.

Statistical analysis

The demographic characteristics of the cohort were explored using descriptive statistics. The exposure variable, patients experiencing homelessness, was described as a frequency with proportion. Seasonal and temporal trends in homelessness were explored using quarters (Q1 = January to March, Q2 = April to June, Q3 = July to September, Q4 = October to December). The demographic characteristics, healthcare resource utilization and outcomes of patients experiencing homelessness and those who had stable housing were compared using Chi-squared tests for categorical variables and Wilcoxon rank-sum tests for continuous variables.

Regression models (linear for continuous outcome variables and logistic for dichotomous outcome variables) were developed to address each of the research questions. Pre-specified patient variables were included in multivariable regression models to control for potential effect measure modifiers and/or confounders. Specifically, age, sex, Charlson comorbidity index and type of ICU admission were included in each model.

Data analysis was conducted using Stata version 17 [39]. Statistical significance was set at α = 0.05 for both univariate analysis and multivariable regression.