Oxygenation/non-invasive ventilation strategy and risk for intubation in immunocompromised patients with hypoxemic acute respiratory failure

We investigated how the initial ventilation/oxygenation management may influence the need for intubation on the coming day in a cohort of immunocompromised patients with acute hypoxemic respiratory failure (ARF). Data from 847 immunocompromised patients with ARF were used to estimate the probability of intubation at day+1 within the first 3 days of ICU admission, according to oxygenation management. First, noninvasive ventilation (NIV) was compared to oxygen therapy whatever the administration device; then standard oxygen was compared to High Flow Nasal Cannula therapy alone (HFNC), NIV alone or NIV+HFNC. To take into account the oxygenation regimens over time and to handle confounders, propensity score weighting models were used. In the original sample, the probability of intubation at day+1 was higher in the NIV group vs oxygenation therapy (OR = 1.64, 95CI, 1.09–2.48) or vs the standard oxygen group (OR = 2.05, 95CI: 1.29–3.29); it was also increased in the HFNC group compared to standard oxygen (OR = 2.85, 95CI: 1.37–5.67). However, all these differences disappeared by handling confounding-by-indication in the weighted samples, as well as in the pooled model. Note that adjusted OR for day-28 mortality increased with the day of intubation. In this large cohort of immunocompromised patients, ventilation/oxygenation management had no impact on the probability of intubation on the coming day.


Estimation of IPTW
In order to estimate and compare the causal effect of daily respiratory management strategy on the probability of intubation in the coming day, we computed inverse probability of treatment weights (IPTW) using propensityscore (PS) [4]. This approach aims at providing causal estimate of treatment effect, by creating a new sample in which the distribution of measured baseline covariates is independent of treatment assignment [5]. Our quantity of interest was the Average Treatment Effect on the Treated (ATT) that addresses the question of how outcomes would differ if the subjects who were actually treated were given the other choice. We used unstabilized weights given benefits of stabilized weights for dynamic marginal structural model (MSM) has been considered uncertain [6]. Thus, treated subjects receive a weight of 1, while untreated subjects receive a weight of PS/(1-PS).
Two treatment exposures (that is, of non-invasive oxygenation strategies) were considered over time successively. First, each day, we only distinguished NIV versus oxygen therapy regardless the device (standard oxygen or HFNC), where NIV defined the exposure of interest, using logistic regression to predict treatment assignment. Secondly, we considered four groups of noninvasive strategies, distinguishing among NIV patients those administered NIV alone and those receiving NIV associated with continuous administration of oxygen through HFNC, patients receiving HFNC alone and those with standard oxygen therapy alone. As proposed by McCaffrey et al. [7], we used Generalized Boosted Model (GBM) for estimation of the IPWT. GBM estimation is a non parametric machine learning technique which used an iterative process with multiple regression trees to take into account nonlinear relationships between pretreatment covariates and treatment assignment [8,9]. Standard oxygen strategy was defined as the reference group, so that the other three strategies (NIV, HFNC alone or HFNC+NIV) appeared those of potential interest to be compared with the standard. For each comparison, the PS algorithm was run separately with a number of 8,000 iterations in order to optimize the balances.

Assumption checking
We computed for each covariate standardized mean difference, also referred as to the absolute standardized mean difference or the population absolute standardized bias (PSB) [10], considering that a standardized difference below 10% or 20% [7] was an acceptable threshold indicative of negligible imbalance. As an overall balance measure, we then computed the C-statistic derived from the ROC curve from the PS model on the weighted sample for the two-treatment exposure model [11], and the Kolmogorov-Smirnov (KS) statistic for the four-treatment exposure model [10]. We also assessed the positivity assumption by an examination of the weights values and distribution. Otherwise, to avoid extreme weights due to near violations of the positivity assumption, we truncated weights at the 90th percentile (i.e., any weight larger than the 90th percentile was assigned to the 90th percentile) [12].

Propensity score and inverse probability weighted based analysis
Covariates balance before and after weighting is reported in Supplementary Figure 5A. As expected, the imbalances before weighting as measured by the standardized mean differences were decreased after weighting, all below the threshold of 10%. The reduced balance in covariates was also illustrated by the values of the C-statistic for the three final models that were 0.566, 0.558 and 0.526, respectively. The mean of the weights in the final MSM was 0.78 with a maximum value of 5.8 and a minimum value of 0.06 ( Supplementary Figures 6 and 7).

Model 2: Standard oxygen therapy alone vs. Non invasive ventilation and high flow nasal cannula Propensity score analysis and Inverse Probability Weighted based analysis
Supplementary Figure 5B displays the standardized mean differences on the original sample and after weighting. Graphical examination suggests that the IPTW has created a similar distribution of measured covariates between treated and control subjects, though some imbalances were still above 0.2. Supplementary Figure 8 shows the different weights according to the oxygenation strategy in the MSM. The mean value of weights was 1.14 with extreme values ranging from 0.02 to 15.7. Although the mean value was close to one, extreme values likely reflect the limited sample size of the NIV with HFNC group.