Study design and Patients
We analysed data from the French and European Outcome Registry in Intensive Care Units (FROG-ICU). The FROG-ICU study (www.clinicaltrials.gov/show/NCT01367093) was a prospective, observational, multicenter cohort study, designed to assess all-cause 1-year mortality after ICU discharge and to identify the mortality risk factors during the year following discharge from the ICU [18]. The study protocol has been previously published [19]. Briefly, the study was conducted in France and in Belgium and was approved by ethical committees of both countries [20]. The study involved ICUs of 21 centers. The study cohort included 2087 consecutive patients, who were admitted to the ICU in any of the participating centers from August 2011–June 2013 when the following inclusion criteria were met: invasive mechanical ventilation support for at least 24 h and/or treatment with a vasoactive agent (norepinephrine, epinephrine, dobutamine, levosimendan, phosphodiesterase inhibitors) for more than 24 h. Non-inclusion keys criteria were: < 18 years old, severe brain injury or brain death or a persistent vegetative state, pregnancy or breastfeeding, transplantation in the past 12 months, not expected to survive or to leave the hospital and/or no social security coverage [18].
Data collection and biological samples
In the FROG-ICU study, the following patient data were collected at the time of inclusion: demographics, past medical history, measure of premorbid status (Mac Cabe score classifies all hospitalised patients into 3 categories: (1) non-fatal disease, (2) fatal disease within 5 year and (3) fatal disease within 1 year), ICU admission diagnosis, hemodynamic, and severity of disease classification scores. The need for organ support (vasopressors, renal replacement therapy) and the number of transfused units (packed RBC, fresh frozen plasma, or platelets) were recorded throughout the ICU stay. In addition, critical parameters and clinical events between admission and inclusion were also recorded (e.g.transfusion status, SAPS 2, antiplatelets treatments, coronary revascularization, heparin treatment…). Biological routine parameters were collected at inclusion and at discharge to study risk factors associated with 1-year survival. Hemoglobin concentration was measured daily from inclusion to day 3, and then bi-weekly until discharge or death. A biobank was created and stored at − 80 °C with blood samples collected within 24 h after patient inclusion and at discharge. Among the 1551 patients discharged alive from ICU and who were included in this study, the following biomarkers were centrally measured a posteriori (1) plasma levels of hs troponin I (Abbott, Abbott Park, IL, USA), N-Terminal pro-Brain Natriuretic Peptide (NT-proBNP, Roche Diagnostics GmbH, Mannheim, Germany), proenkephalin A 119–159 (penKid, Sphingotec GmbH, Hennigsdorf, Germany), neutrophil gelatinase associated lipocalin (NGAL), galectin-3 (Abbott, Abbott Park, IL, USA), haptoglobin (Architect, Abbott Park, IL, USA), interleukin 6 (IL-6, Elecsys, Roche, Penzberg Germany) and (2) urine concentrations of: NGAL (Abbott, Abbott Park, IL, USA), cystacin C (Abbott, Abbott Park, IL, USA), liver fatty acid binding protein (L-FABP, Nordia L-FABP; Sekisui Medical Co., Ltd., Tokyo, Japan). To account the close relationship between haptoglobin expression and IL-6 levels, haptoglobin level was normalised on IL-6 level at discharge time point [21].
Objectives
The primary objective was to describe the association between in-ICU RBC transfusion and 1-year mortality after ICU discharge. Exploratory analyses were also conducted to determine the discharge factors associated with 1-year post-ICU mortality, specifically hemolysis and kidney injury.
Statistical methods
Additional file 1: Figure S1 summarises the statistical analysis performed.
Patients were separated into two groups: those who received RBC transfusion (i.e.: at least one unit of packed RBC) during their ICU stay, and those who did not. Survival was observed over a period of 1 year following ICU discharge.
Of note, to assess the impact of transfusion in patient selection at discharge, 1-year survival curves were additionally also drawn from admission to ICU with the whole FROG cohort population.
Primary outcomes
The primary outcome was 1-year survival after ICU discharge. The average treatment effect of RBC transfusion on survival was estimated from the survival curves of patients with and without RBC transfusion, from the associated hazard ratio, and from the differences in restricted mean survival times (RMST). The latter corresponds to the average number of days gained or lost in terms of 1-year overall survival after ICU discharge between patients transfused and not transfused during their ICU stay. The confidence intervals associated with these estimated values were computed from 100 bootstrap samples.
Data description
Data were expressed as median (inter-quartile range, IQR), mean ± standard deviation (SD), or number (percentage). Numerical data were compared using t-test or Wilcoxon rank test, while categorical variables were compared using χ2 or Fischer’s test, as appropriate. Repeated measures of continuous variables were handled by a linear mixed model tested with Kenward-Roger’s F tests.
Management of missing data
Two approaches were used for handling the missing values: a parametric one with multiple imputations by chained equations (MICE), and a non-parametric one with random forest-missingness incorporated in attributes (MIA). Details of the two methods are provided in the Additional file 2. Number of missing values per variable was also added in Additional file 1: Figure S2.
Models
Two different approaches were considered to estimate the effect of RBC transfusion on 1-year mortality: a semi-parametric approach and a non-parametric approach. In the semi-parametric approach, we used Cox models to model the survival and the censoring. Treatment allocation was modelled with a propensity score calculated from a logistic regression. In the non-parametric approach, we modelled with random survival forests the survival, the censoring, and the treatment allocation.
Under these semi-parametric and non-parametric approaches, two estimators (see estimators performed below) were applied to assess the study’s primary outcomes based on models in which identification of confounding factors was required.
A three round Delphi method including experts in critical care and transfusion was used to identify the confounding variables necessary to build the different models. Additional file 1: Figure S3 shows the causal inference diagram applied in a directed acyclic graph, differentiating variables assessed as predictors of the outcome but unrelated to the treatment assignment and the variables assessed as predictors of both treatment and outcome.
Causal inference estimators
The unweighted curves were estimated with the Kaplan Meier estimator, the unweighted hazard ratio was estimated with a Cox regression with only the transfusion status as a variable. The weighted survival curves were built with the non-parametric doubly robust estimator, i.e.: augmented inverse probability of treatment weighting—augmented inverse probability of censoring weighting (AIPTW-AIPCW using survival and random forests-MIA method for management of missing values, see Additional file 2). The weighted hazard ratio was computed from the weighted survival curves by averaging the hazard ratio at each time point. This estimator was calculated in our main population of interest: ICU-survivors from ICU-discharge but also, to ensure that the result was not driven by patient’s ICU stay, in the whole initial population (including ICU-survivors and non-survivors) from ICU admission.
The three following causal inference estimators were performed to calculate the RMST: (1) the unweighted estimation with no adjustment, then, in parametric and non-parametric approaches of missing values: (2) the inverse probability of treatment weighting with the Kaplan Meier estimator (IPTW), (3) the AIPTW-AIPCW. Details of each estimator are provided in the Additional file 2.
Exploratory analysis
Except for the packed red blood cells unit number threshold associated with 1-year mortality (see below), all exploratory analyses were performed using parametric (with MICE) imputed FROG-ICU cohort.
With non-imputed data, we looked for the number of the packed RBC units for which there was a maximal increase in 1-year mortality after ICU discharge. First, the log linearity assumption was checked using the restricted cubic spline method. Given the lack of log linearity, the number of transfused packed RBC units has been dichotomized according to an optimal level determined using the most significant p value from the log rank test. Subsequently, this threshold has been validated using a univariate Cox model.
A two-tailed p value of less than 0.05 was considered significant. Statistical analyses were performed using R v3.6.3 (R Foundation for Statistical Computing, Vienna, Austria).