Center effect in intubation risk in critically ill immunocompromised patients with acute hypoxemic respiratory failure

This study included adult immunocompromised patients with hypoxemic acute respiratory failure (ARF) from the TRIAL-OH cohort and the HIGH randomized controlled trial (RCT). Details of each study have been described previously [3, 16]. Firstly, we derived from the TRIAL-OH study (1011 patients, 17 centers, 2010–2011) [16] a cohort of hematological patients admitted to ICU for ARF. We selected patients with the following criteria: ARF as defined by tachypnea > 30/min, respiratory distress, SpO2 < 90% at ICU admission, and/or labored breathing. Exclusion criteria were admission for another cause, hypercapnia (defined by a PaCO2 > 50 mmHg), and invasive mechanical ventilation before ICU admission. Secondly, we used the HIGH trial [3], a RCT which enrolled 776 immunocompromised patients with hypoxemic ARF in 31 ICUs, in order to compare HNFC to standard oxygen at day-28 mortality. In both studies, each participating center has a senior intensivist and a senior hematologist in site available 24/7 and sharing decisions of ICU admission [17]. Note that 14 centers were common in both the TRIAL-OH cohort and the HIGH trial.

The appropriate ethics committees approved each study.

The primary outcome was the need for intubation through ICU stay as a dichotomous variable. In the TRIAL-OH study, the decision to perform endotracheal intubation was left to clinicians in charge at each ICU. In the HIGH trial, the decision to perform endotracheal intubation was based on predefined criteria, in agreement with current guidelines, including the therapeutic response and the clinical status (SpO2, respiratory rate, signs of respiratory distress, and bronchial secretion volume) [3]. Individual data reported in tables were collected prospectively. ICU-level data (i.e., center characteristics) were also used in the analysis. In the TRIAL-OH cohort, these data were assessed as part of data collection. In the HIGH trial, they were assessed with a secondary survey from all the participant centers.

Continuous variables are described as median and interquartile range (IQR), and categorical variables are summarized by counts (percent).

Patient characteristics at admission were examined at the patient level in both studies, and the need for intubation was first assessed in each center without adjusting on potential confounding factors, which is referred as the “crude need for intubation.” Hierarchical logistic regression models were used to examine the variability on the outcome between ICUs and the association between ICU characteristics and the outcome, adjusting for patient characteristics [18]. To do this, we used the mixed-effect model with a normally distributed random effect for ICU (random intercept). Exchangeability was assumed across all providers [5]. In practice, the effect of a given ICU was modeled through its own regression coefficient, which compares to the crude average need for intubation across all centers [5, 6]. This allows us to model between ICU heterogeneity in the average intubation risk. These models provide an estimation of heterogeneity in the form of the variance of the random effects, where the closer the variance is to zero, the smaller the center effect is. Because it could be hard to interpret, we computed the median odds ratio (MOR) to better understand the importance of the center effect on the mean intubation risk, that is on the same scale as traditional prognostic factors [6, 7, 18, 19]. Briefly, MOR corresponds to the median of all ORs that can be computed between two patients with the same characteristics, but randomly chosen from two different centers, namely in a higher risk center and in a lower risk one [6, 18]. MOR quantifies the differences between ICUs. If the MOR equals 1, then there are no differences in intubation risk between ICUs.

First, the model was built without any adjustment (“empty model”), which allows us to investigate a potential center effect. This unadjusted model contained only ICU-specific random intercepts. Then, we provided adjustment on predefined individual covariates as fixed effects, without ICU characteristics or interventions. These covariates were specified a priori as potential confounders and included age, comorbidities assessed by the Charlson comorbidity index, type of immunosuppression (in three classes, malignant hemopathy, solid cancer, and others), previous allogeneic stem cell transplantation, performance status > 2 (bedridden or dependent), severity of illness with SOFA score without respiratory item, Pa0₂/FiO₂ ratio in four categories (> 300, 200–300, 100–200, ≤ 100 mmHg, with > 300 mmHg as reference), respiratory rate > 30 /min, and ARF diagnosis. This model was used to estimate adjusted ICU random effects, because we were interested in studying heterogeneity in outcomes. ICU were then ranked by their estimated random effect on intubation risk (adjusted for patient characteristics only) and classified into quartiles. For descriptive purposes, patient and ICU characteristics were compared across quartiles of the risk-adjusted intubation rate, using the Cochran–Mantel–Haenszel row mean score test and nonzero correlation test for testing the difference for categorical and continuous variables, respectively. To illustrate the magnitude of the effect of center on intubation risk, we performed conditional standardization of the regression results for a given patient with median and modal values for the covariates in the patient-level adjusted model. Last, hospital- and ICU-level covariates (center volume, annual volume of ID in ICU, annual rate of IMV patients defined by the number of patients with IMV/number of admission, academic status of hospital) were entered in the model as fixed effects and reported in terms of odds ratios, in order to try to explain the discrepancies between center. Because the time since respiratory symptoms onset and ICU admission could reflect local practices, it was considered as an ICU-level characteristic in the analysis.

To test for the significance of the center effects (empty model and model adjusted on patient and center characteristics), we used permutations test, a recommended approach to test for random effect [20‐23]. More details about the methods used are given in the Additional file 1.

These methods were applied in both the observational cohort and the RCT, separately.

Primary analyses were performed on the complete cases, assuming missing completely at random covariates. Then, sensitivity analyses for such assumptions were performed, based on multiple imputation with chained equation [24]. We performed exploratory subset analyses, restricting ourselves to patients with full code status, in patients with malignant diseases in the HIGH trial and after exclusion of ICUs with extreme size (ICUs > 20 beds or ICUs< 8 beds). In the HIGH trial, we also investigated the center heterogeneity in the prognostic effect of the oxygenation strategy which was randomly assigned (HFNC or standard oxygen). This was not possible in the observational cohort due to allocation bias. To do this, we introduced two random effects in the patient-level adjusted model for each center: a random effect on the mean intubation (random intercept) risk as previously described and a random effect on the effect of the oxygenation strategy (with standard oxygen as reference) on intubation risk random slope. This allows us to model centers’ variability not only on the average intubation risk but also on the effect of the oxygenation strategy on the need for intubation. We then applied permutation test to investigate for significance of center effect [20, 22].

All reported p values are two-sided; p < .05 was considered statistically significant. All analyses were performed using R version 3.1.0 (R Foundation for Statistical Computing [http://​www.​R-project.​org/​]).

Overall, 703 patients (age 61.0 years [51.0–71.0], 61.0% male) were included in the observational cohort and 776 patients (age 64.0 years [56.0–71.0], 66.2% male) from the HIGH trial (Additional file 1: Figure S1). Table 1 reports the main characteristics at ICU admission and the respiratory parameters at baseline. As shown, the main cause of immunosuppression was hematological malignancies in both studies (94% in the observational cohort and 45% in the RCT). Most of the patients were admitted from ward with a good performance status (i.e., 0–2) over the 3 months preceding ICU admission (78.8% and 63.7% respectively). Sequential Organ Failure Assessment (SOFA) score at admission was 6 [3, 4, 8‐12, 25] and 6 [4, 8‐10, 25] the in cohort and trial respectively. In both studies, the main diagnosis of ARF was bacterial infection (269 (38.2%) in the Trial-OH cohort, 355 (45.7%) in the HIGH trial) followed by opportunistic germ infections (89 (12.6%) and 93 (12.0%)).

ICU and hospital mortality rates were 32.4% (228 deaths) and 44% (309 deaths) in the cohort and 31.6% (245 deaths) and 41.5% (322 deaths) in the trial, respectively.