This study was approved by the Research Ethics Board at the University of Alberta, Edmonton, Canada (File # Pro00046184). The need for written informed consent was waived.
Study Design, Setting, and Population
This was a multi-centre retrospective cohort including all adult patients (n=10,463) surviving to ICU discharge between June 2012 and December 2014, the period of time during which a large provincial data extraction for all patients admitted to all ICUs in nine acute care hospitals in Edmonton and Calgary, Alberta.16 For patients with multiple ICU admissions, only the first admission was analyzed. The included ICUs were nine mixed medical/surgical units with a median [IQR] of 24 [10-28] beds. Two units were classified as academic/quaternary, two were academic/tertiary units, and five were community/metropolitan units (Table 1). All included ICUs operated with a “closed” model and were staffed by certified intensivists who are present and available during daytime hours, and available afterhours on-call.
Data were analyzed from eCritical Alberta, a provincial clinical information system, data warehouse and clinical analytics system.17eCritical is a bedside interdisciplinary electronic documentation system (MetaVisionTM, iMDsoft) which captures demographic, diagnostic/case-mix (i.e., comorbidity, diagnostic classification, surgical status, Acute Physiology and Chronic Health Evaluation [APACHE] II and III score), Sequential Organ Failure Assessment (SOFA) score, Therapeutic Intervention Scoring System (TISS) 28 score, laboratory, and device data (physiologic monitors, ventilators, renal replacement therapy, use of vasoactive medications). TRACER is a comprehensive, multi-modal and integrated data repository and clinical analytics system. The eCritical Alberta system provides high-quality data due to rigorous data quality assurance and auditing processes and has been used by our team previously in health services research.4,18-21 There were no missing data elements for our primary exposure and outcome (afterhours discharge and hospital mortality, respectively). Missing data for individual variables were <3% and were not replaced or imputed.
Main Exposure and Outcomes
The primary exposure was afterhours ICU discharge, defined as occurring between 19:00-07:59h.5 The primary outcome was all-cause post-ICU hospital mortality. Secondary outcomes included in-hospital mortality within 3 days, 7 days, and 14 days of discharge, ICU readmission at 48- and 72-hours, and total hospital and post-ICU hospital length of stay. Organ dysfunction at the time of ICU discharge was quantified by the average SOFA score in the 72 hours preceding ICU discharge. Additional variables included: age, sex, case-mix (e.g., diagnostic classification, surgical status, comorbidities, and Charlson comorbidity index, use of renal replacement therapy or vasoactive infusions), ICU-care variables (e.g., TISS-28 score 72 hours prior to discharge, percentage of days receiving DVT prophylaxis or stress ulcer prophylaxis, duration of intubation and mechanical ventilation, unplanned extubation) and ICU site (e.g., location, hospital type).
Data were initially explored descriptively. Normally or near-normally distributed data were reported as mean (SD) while non-normally distributed continuous data were reported as median (IQR). Comparisons were by Student’s T-test or Wilcoxon-Mann-Whitney U-test, respectively. Categorical variables were reported as frequency (percentage) and compared by Chi-squared test.
Path-analysis modelling: We proposed a two-stage modelling strategy to evaluate the direct, indirect, and integrated associations between afterhours discharge, illness severity (e.g., organ dysfunction at ICU discharge), and hospital mortality (Additional File 1). We used a path-analysis strategy to better explore hypothesized relationships that may both directly and indirectly modify the potential causal association between afterhours discharge and in-hospital mortality.4,22-24 First, we used random-effects multi-variable linear regression to quantify the association between discharge organ dysfunction and afterhours discharge (Additional File 2). Second, we used random-effects multi-variable logistic regression to evaluate the association between discharge organ dysfunction and afterhours discharge on hospital mortality. We assumed that discharge organ dysfunction varied by ICU, so ICU site was used as a random-effects predictor. A sparse model was created by stepwise variable selection for multi-variable modelling analyses.25 SAS (release 9.4; SAS Institute, Cary, NC) was used for all modelling analyses. Finally, we proposed to estimate the integrated effect by combining the direct and indirect effects of afterhours discharge on hospital mortality, conditional on finding significant association between discharge organ dysfunction and afterhours discharge, in simulation experiments (1 million replicates) via R Core26.4,22 The association between afterhours discharge, discharge organ dysfunction, and lengths of stay were further evaluated using random-effects multi-variable Poisson regression (Additional Files 3 and 4).
Exploratory Subgroup and Sensitivity Analysis: Sensitivity analysis utilized the same modelling process as our primary analysis, and examined afterhours discharge utilizing different definitions of afterhours discharge (i.e., 19:00-06:59, 19:00-08:59, 19:00-09:59, 20:00-06:59, 21:00-06:59, and 22:00-06:59 and by ICU site), as well as indicators of strained ICU capacity, afterhours discharge on non-work day, afterhours discharge when occupancy ≥90%, and afterhours discharge with clustering of admissions ≥2. Clustering of admissions was defined as the number of admissions in the two hours before or after the index admission, divided by the number of funded ICU beds.16