Data collection and analysis
Data were collected retrospectively from the electronic and archived
hospital medical records. We attempted to specifically identify the
effects of VV-ECMO as a BTT on: posttransplant 30-day mortality and
complications (need for postoperative ECMO, delayed chest closure,
surgical re-exploration, tracheostomy, chest drainage within 24 hours,
chest infection, sepsis, stroke, acute kidney injury [AKI] requiring
renal replacement therapy [RRT]) and 1-year mortality.
Preoperative, intraoperative, and postoperative data are summarized for
BTT and non-BTT patients. In the main analysis, both subgroups were
submitted to optimal full matching based on Mahalanobis distance in
respect to preoperative covariates. Based on their potential relevance
to the observed outcomes and imbalance between the two subgroups,
included covariates were age, gender, body mass index (BMI), serum
creatinine and hemoglobin levels, platelet count
<150x109/L and main diagnosis, with exact
matching on the gender, low platelet count and underlying diagnosis
(cystic fibrosis [CF] or “other”).19,20 We had
no patients that required VV-ECMO as BTT among those with chronic
obstructive pulmonary disease (COPD), emphysema, bronchiolitis,
bronchiectasis, pulmonary hypertension and lymphangioleiomyomatosis.
Therefore, in order to avoid aliasing between potential effects of
VV-ECMO and diagnosis, a sensitivity analysis (using the same
methodology) was performed including only diagnoses where at least one
patient was bridged to LTx with VV-ECMO. To evaluate the effect of
VV-ECMO as a BTT (vs. non-BTT), generalized mixed models (binary
distribution, logit link; subclass as a random effect [cluster])
were fitted to each binary outcome with further adjustment for
unbalanced covariates: frequentist (maximum likelihood estimation with
Gauss-Hermite quadrature approximation; classical [sandwich] robust
estimator) and Bayesian (4 chains, 4000 iterations, 8000 samples of the
posterior, vaguely informative normal priors for ln[odds] and the
intercept [0, 2.5; scaled], and priors on the terms of a
decomposition of the covariance matrices [Gamma shape=1, scale=1; LKJ
for correlation matrix, regularization=1; Dirichlet for the simplex
vectors, concentration=1]). To evaluate the effect of VV-ECMO as a BTT
on the chest drainage within the first 24 hours, data were
ln-transformed (since right-skewed) and the same models, although with
normal distribution and identity link, were fitted. We used packageMatchIt in R for matching,21 SAS 9.4 for
Windows proc glimmix (SAS Inc., Cary, NC) for fitting frequentist
and package rstanarm in R for Bayesian
models.22 We evaluated susceptibility of the observed
effects to unmeasured confounding by determining the E-value (packageEvalue in R).23 Despite a large number of
analyzed outcomes and related formal statistical tests, we considered
more appropriate not to implement multiplicity adjustments as
adjustments of comparison-wise alpha could have resulted in falsely
overlooked adverse effects of VV-ECMO as a BTT.