Methods
Study design and data
collection.
All adult (18 years and more) admissions to a tertiary care university
hospital (Centro Hospitalar Universitário de São João), in Porto,
Portugal, between January 1, 2013, and December 31, 2015, were
considered for this study (Figure 1). Admissions to maternity and
gynaecological departments, patients with a history of renal transplant,
on chronic renal replacement therapy (RRT) or hospitalised due to
end-stage renal disease were excluded. We also eliminated admissions
with no inpatient serum creatinine (SCr) measured. For patients with
multiple hospitalisations during the study period, we included only one
randomly selected admission.
Using institutional electronic medical records, we retrieved patients’
demographics, admission and discharge data, discharge diagnoses and
procedures coded according to the International Classification of
Diseases, Ninth Revision, Clinical Modification, and laboratory results,
including SCr with the date and time of biological specimen sampling.
All clinical routine SCr measurements were isotope-dilution mass
spectrometry-aligned. Comorbidities were identified, and Charlson Index
was automatically calculated on the basis of patients’ diagnoses of
hospitalisations in the previous five years and certain secondary
diagnoses of the index hospitalisation. Patients’ vital status after
discharge was ascertained based on electronic medical records.
This study was approved by the institutional Ethics Committee (Comissão
de Ética para a Saúde do Centro Hospitalar Universitário de São João,
reference number 365-15 dated: 29-12-2015) with a waiver of informed
consents obtained because of the observational nature of the study.
Definitions
The baseline SCr was defined as the median of ambulatory measurements at
the same hospital between 7 and 365 days before admission(11). When no preadmission SCr was available, missing
values were imputed using random-forest model controlled for patients’
age, sex, admission unit and type (medical, surgical and intensive care;
emergency and elective), comorbid conditions (history of myocardial
infarction, chronic heart failure, cardiovascular disease, dementia,
chronic liver disease, pulmonary disease, diabetes mellitus, chronic
kidney disease (CKD), hypertension, cancer, peripheral vascular disease
and rheumatologic disease), total Charlson Index and first inpatient
SCr. To calculate the estimated glomerular filtration rate (eGFR), we
used Chronic Kidney Disease Epidemiology Collaboration
formula(12).
AKI was identified using the Kidney Disease: Improving Global Outcomes
definition (13, 14) if at least one of the criteria
was met: i) SCr ≥ 1.5 times higher than the baseline SCr within first 7
days after admission; ii) SCr ≥1.5 times higher than the lowest
inpatient SCr creatinine within 7 days and iii) SCr ≥0.3 mg /dL higher
than the lowest value within 48 hours, and with the increase sustained
for more than 24 hours. Urine output data were not available. We
considered one AKI episode per admission and the time of first SCr
measurement meeting the criteria was recorded as AKI onset. Patients
with AKI apparent within 24 hours of admission were designated
community-acquired AKI (CA-AKI), while patients in whom the syndrome
developed afterwards were denoted as hospital-acquired AKI (HA-AKI).
Based on the ratio of peak inpatient SCr relative to the baseline value,
we categorised AKI severity into three stages: Stage 1 - ratio 1.5 to 2
or increase in SCr of 0.3 mg/dL; Stage 2 - ratio 2 to 3 and Stage 3 -
ratio ≥3 or increase in SCr above ≥4.0 mg/dL or receipt of RRT.
Statistical analysis
Patients with CA-AKI and HA-AKI were compared with respect to baseline
characteristics, clinical presentation and outcomes. Normally
distributed continuous variables are reported as means and their
standard deviations (SD) or as medians with 25th and
75th percentiles (P25, P75) otherwise. Categorical
variables are presented as counts with percentages.
We estimated cumulative incidence functions for the length of hospital
stay, in-hospital and 6th-month mortalities and
compared them among groups of patients with No AKI, CA-AKI and HA-AKI,
and across severity stages, with differences being evaluated by log-rank
tests. Discharge and in-hospital death were considered to be the
competing risk outcomes, and time to these events was presented in days
counted from the syndrome onset. We used multivariable Cox proportional
hazard regression to examine the effect of the presence of CA- or HA-AKI
and its’ severity on the risk of outcomes. There were significant
interactions between AKI type (CA- and HA-AKI) and AKI stage, therefore
in the Cox model, we created dummies to treat each combination AKI type
- AKI stage separately.
To estimate the hazard function for AKI incidence over time in patients
free of AKI at admission, we tested two parametric survival regression
models with:
exponential distribution of time to an event, T~exp
(λ), assuming constant hazard function over time and given by
\(h\left(t\right)=\ \lambda\)
where hazard function h(t) denotes the probability of AKI onset
on day t , given that the patient remains AKI free to the
beginning of day t and λ is a scale parameter,
Weibull distribution of time to an event, T~Weibull
(λ, p), allowing monotonic increase or decrease in hazards over time
and given by
\begin{equation}
h\left(t\right)=\text{λpt}^{p-1}\nonumber \\
\end{equation}where p is the shape parameter (the hazard function is increasing
when p > 1, and decreasing otherwise).
We plotted the survival functions of the two models against the
Kaplan-Meier curve to determine the most suited distribution of time to
AKI occurrence. We used Akaike’s information criterion (AIC) to evaluate
which model better fitted to our data.
All analyses were performed using R version 3.5.3 (R Foundation for
Statistical Computing, Vienna, Austria) and packages: ‘random
forest’ , ‘survival’ , ‘survminer’ and ‘cmprsk’ .