Study
Data:
In this retrospective analysis, we obtained data from the 2016 and 2017
National Inpatient Sample (NIS), which is sponsored by the Agency for
Healthcare Research and Quality as a part of the Healthcare Cost and
Utilization Project (HCUP). NIS is the largest publicly available
all-payer administrative database, containing data on more than 7
million hospitalizations (unweighted); when weighted, it represents
about 35 million hospitalizations nationally. It provides information on
clinical and resource utilization with safeguards to protect data for
individual patients, physicians, and hospitals. Beginning in October
2015, the NIS started using the International Classification of
Diseases, Tenth Edition, Clinical Modification/Procedure Coding System
(ICD-10-CM/PCS) to reflect the implementation of ICD-10-CM/PCS by
hospital systems. Using the Agency for Healthcare Research and Quality
sampling and weighting method, national estimates of the entire U.S.
hospitalized population were calculated. [16]
Study
design:
Given the de-identified nature of the NIS data, our study was exempt
from approval from the Institutional Review Board. We identified all
patients (≥18 years of age) who had a discharge diagnosis of ESRD (n
=2187954), using their respective ICD-10-CM/PCS codes. We divided the
total sample into two groups: ESRD with LGIB (n=242075) and ESRD only
(n=1945879). We identified patients with LGIB using appropriate
diagnosis codes. The ICD-10-CM/PCS codes used in this study are
displayed in supplementary tables 1 and 2.
For baseline characteristics, we used patient demographics (age, race,
and sex), the Charlson Comorbidity Index, insurance status, hospital
characteristics, and relevant comorbidities coronary artery disease
(CAD) or CAD equivalent, hypertension (HTN), obesity, dyslipidemia,
diabetes mellitus (D.M.), chronic lung disease, tobacco smoking, alcohol
use, peripheral vascular disease , blood thinner use
(anticoagulants/antithrombotics/antiplatelets) and congestive heart
failure (CHF) (Table 1). Comorbidities were identified using their
respective ICD-10-CM/PCS codes (supplementary table 1)).
Outcomes:
The primary outcome of interest was all-cause in-hospital mortality and
predictors of mortality. The secondary outcomes included the incidence
of sepsis, acute coronary syndrome, shock requiring vasopressors, acute
respiratory failure, disseminated intravascular coagulation, and
mechanical ventilation. Complications were identified using their
respective ICD-10-CM/PCS (supplementary tables 1 and 2). We also studied
the length of hospital stay (LOS) and average hospital costs.
Statistical
Analysis:
We conducted all statistical analyses as per the recommended methods
accounting for the intricate survey design of the NIS database. [17]
Categorical data are reported as frequency and percentage, and
continuous data as mean with standard deviation and standard error.
Categorical variables were analyzed using Pearson’s Chi-square test, and
continuous variables were analyzed using the Student’s t-test.
Unadjusted odds ratios for the primary and secondary outcomes were
calculated using univariate logistic regression. Multivariable logistic
regression was used to adjust for potential confounders in the final
model. Statistical significance was set at a two-sided p-value of
<0.05. STATA/ MP 15.10 (Stata Corp LLC) was used for
statistical analysis. All analyses in our study were weighted using
provided discharge weights to produce national estimates. Hospital costs
were inflation-adjusted for 2018 using the Consumer Price Index
(provided by the U.S. Department of Labor).
To account for the differences in baseline characteristics, we used
propensity score matching. [18 19] To establish a propensity-matched
cohort, we used the treatment outcomes as the outcome variable and
potential confounders as covariates. A 1:1 propensity score-match was
performed using a caliper width of 0.1 using the psmatch2 command.
Appropriate Caliper was calculated by multiplying 0.2 with a standard
deviation of the logit of the propensity score. Pstest was used
to generate the unmatched and matched variable. The standardized
difference of <10% checked with pbalchk command suggested
adequacy of the match between two groups among the measured covariates.
Regression analysis was done using a generalized linear model using all
covariates in succession in the final model, including patient-level
discharge weights. We added covariates one by one in the model, and if
the coefficient changed by more than 20%, we included that covariate in
the final model. A full list of covariates used in the regression
analysis and confounders in the multivariable regression model is shown
in the supplemental table 2.