This is a post hoc analysis derived from a prospective, multicenter cohort study of 1011 critically ill adult patients with hematologic malignancies admitted to 17 ICUs in Belgium and France from January 1, 2010 to May 1, 2011 [3]. Briefly, the study was carried out in university or university-affiliated centers in France and Belgium that belonged to a research network instituted in 2005. In all 17 centers, a senior intensivist and a senior hematologist are available around the clock and make ICU-admission decisions together. During the study period, consecutive patients having hematologic malignancies who were admitted to the participating ICUs for any reason were included. Exclusion criteria were complete cure of the malignancy for more than 5 years, ICU admission only to maximize safety of a procedure, and age younger than 18 years. For this post hoc analysis, patients with missing data on the outcome were also excluded. The study was approved by the appropriate ethics committees in France and Belgium. All patients or relatives were informed and consented to participate in the study.
Data collection and outcome measure
Data abstracted from the study database were: age, sex, underlying malignancy, disease status (newly diagnosed if malignancy was diagnosed in the last month, complete or partial remission, other), autologous or allogeneic bone marrow or hematopoietic stem-cell transplantation (BMT/HSCT), treatment with long-course corticosteroids, Charlson comorbidity index, performance status, existence of organ failure based on the Sequential-related Organ Failure Assessment (SOFA) criteria, SOFA score, main reason for ICU admission, circumstances of ICU admission, time between hospital and ICU first request, and between first request and ICU admission, number of requests before ICU admission, specialty and experience of the physician that requested ICU admission, direct admission from the ED, length of ICU and hospital stay, vital status at ICU, at hospital discharge, and at day 90.
The variable of interest was direct admission from the ED, and the main outcome was in-hospital mortality.
Statistical analysis
Results are reported as medians with interquartile ranges (IQR) for continuous variables and numbers with percentages for binary and categorical variables.
Patient characteristics were compared using the Chi-square test for categorical variables and the Wilcoxon rank sum test or the Student t test, as appropriate, for continuous variables (with normality tested with Shapiro–Wilk test). We investigated the association between the variable of interest and the outcome by multivariable logistic regression to search for potential confounders. Characteristics associated with the outcome on the basis of P-values less than 0.1 by univariable analyses were included in a multivariable logistic model; clinically relevant prognostic characteristics such as SOFA score, Charlson risk index or performance status, were forced in the model regardless of their P value. Then, a backward selection procedure was applied, except for clinically relevant prognostic variables that were not removed from the model.
Missing data were managed with multiple imputation by chained equations [4]. The distribution of the data according to the presence or absence of missing data was checked (plots if continuous or tables if categorical variables) to ensure that missing data were missing completely at random. As recommended [5], variables included in the imputation model were those of the logistic regression prediction model (including the outcome), in addition to auxiliary covariates correlated with the missing variables (i.e., sex, underlying disease, days since diagnosis, days between first call to intensivist and ICU admission, experience of the physician requesting ICU, number of calls before ICU admission, organ failures). Five datasets were imputed with 50 iterations each. Multivariable logistic regression model was applied to the 5 imputed datasets and final estimates were obtained by averaging the 5 estimates according to Rubin’s rules.
Several sensitivity analyses were performed. First, a complete cases analysis was performed. Second, to handle potential residual confounding by indication, a propensity score matching was done, where propensity score of being directly ICU admitted was estimated from a multivariable logistic model, with resulting balances in confounders checked by standardized mean differences and c-index [6], then matching performed without replacement within a caliper of 0.2 standard deviation of the logit of propensity score [7]. Third, we plotted survival Kaplan–Meier curve (from ICU admission to 90 days) according to direct admission status from the ED. Hazards ratio (HR) from Cox proportional-hazards models was used to quantify the association between the direct admission status and the outcome, adjusting for baseline predictors of survival. Underlying assumptions of the Cox model were checked: Proportional hazards (PH) assumption was tested using a formal test based on the Schoenfeld residuals, with time-dependent effect considered for covariates that violated the PH assumption [8]. Log-linearity between non-binary covariates and hazard was assessed through splines; in case of nonlinear effect, covariates were dichotomized according to thresholds derived from the splines.
All P-values were two-sided, with values of 0.05 or less considered as statistically significant.
Data were analyzed with R 3.5.0 software (the R Foundation for Statistical Computing, Vienna, Austria).