The predictive value of early acute kidney injury for long-term survival and quality of life of critically ill patients

All patients admitted consecutively to the mixed ICU of the University Medical Center Utrecht from July 2009 to April 2013, without chronic dialysis prior to ICU admission, were eligible for inclusion. In patients with multiple ICU admissions within this period, only the first ICU admission was used in the analyses. Patients under 16 years of age and those with a length of stay shorter than 24 hours were excluded. Patients admitted to the ICU after cardiac surgery were excluded because of the low incidence of AKI [18], and low risk of poor outcome in these patients in general [16]. The Institutional Review Board (IRB) of the University Medical Center Utrecht approved the study protocol and waived the need for informed consent when working with anonymised patient and follow-up data (UMC Utrecht IRB protocol number 10/006).

The following data were collected: serum creatinine levels, urine output per hour within the first 24 hours, age, sex, pre-ICU hospital length of stay, admission type, Charlson Comorbidity Index [19, 20], the need for mechanical ventilation, confirmed infection and the acute physiology score (APS, as part of the APACHE II score) within the first 24 hours of admission. These variables were prospectively collected according to strict definitions, as part of a national registry used for benchmarking [7]. Data on the Charlson Comorbidity Index were obtained from the electronic patient files, as was described previously [21].

After hospital discharge, patient survival was tracked using the municipal registry. All patients surviving one year after ICU admission were sent the EuroQoL 5D-3L™ (EQ-5D) HRQoL questionnaire [22]. If this questionnaire had not been returned within six weeks, the questionnaire was resent and patients were reminded by telephone to return the questionnaire. More details on the definitions of the Charlson Comorbidity Index and the EQ-5D can be found in Additional file 1.

The presence of eAKI was determined according to the risk, injury, failure, loss, end-stage renal failure (RIFLE) classification [5]. This classification is based on measurements of serum creatinine, urine output per hour and the use of renal replacement therapy. A risk; injury; failure (RIF) classification of renal impairment (analogous with the RIFLE system) was used. An algorithm was developed to determine the presence of these acute RIFLE classes within the first 24 hours of ICU stay, based on routinely collected data. RIFLE based on serum creatinine was scored by calculating the factor of change in serum creatinine from baseline. The baseline was defined as the lowest serum creatinine value in the 6 months prior to ICU admission, when available in the hospital laboratory registry. If this was unavailable, the lowest serum creatinine during the first day of admission was used as baseline. RIFLE based on urine output was scored per hour, where 6, 12 or 24 hour stretches of oliguria or anuria were scored according to the RIFLE classification. Hours with missing urine output were replaced by dividing the first known urine output over the stretch of missing hours. This was done for missing periods up to 6 hours, including the hour with the known urine output measurement. Renal replacement therapy was scored based on parameters indicating a running dialysis and invasive therapy registration. The highest (i.e., worst) acute RIFLE class based on serum creatinine or urine output attained during the first 24 hours of admission was used to classify eAKI in included subjects. Subjects with renal replacement therapy were scored as “failure” regardless of urine or serum creatinine. More details on this algorithm can be found in Additional file 2.

The primary outcome measure was one-year “poor outcome”. This outcome was defined as a composite of death or low HRQoL at one year follow up. To study the contribution to the composite endpoint, one-year survival, and HRQoL in the one-year survivors were analyzed as separate secondary outcomes.

HRQoL was measured using the EQ-5D. This questionnaire consists of five questions each representing a dimension of HRQoL (mobility, self-care, usual activities, pain or discomfort, anxiety or depression). Patients assigned a score of no, little or many problems to each of these dimensions. The results were indexed on a scale between 1 (full health) and 0 (dead) according to the weighting scheme for the Dutch population [22].

We defined low HRQoL as an EQ-5D index of 0.4 or below. Patients who qualified their health state as such are on par with those with moderate to severe amyotrophic lateral sclerosis (mean EQ-5D index 0.56–0.27) [23], patients suffering from dementia with depression (mean EQ-5D index 0.37) [24], or patients with a severe to extreme depressive episode (median EQ-5D index 0.57–0.29) [25].

Based on the constituents of most benchmark prediction models (e.g., APACHE and SAPS) additional predictors were selected. These were age, gender, pre-ICU hospital length of stay, admission type (medical/elective surgical/urgent surgical), Charlson Comorbidity Index, the need for mechanical ventilation, confirmed infection and the APACHE II acute physiology score in the first 24 hours of admission [26] (excluding the points for creatinine). The association between eAKI and outcomes was adjusted for overlapping information on the predictive value of these eight additional predictors.

Missing EQ-5D data were expected to occur in non-responding one-year survivors. Multiple imputation was used to replace the missing EQ-5D dimension scores of non-responding survivors [27‐30]. A total of 35 imputation datasets were created. Further details can be found in Additional file 1.

Baseline-characteristics and outcomes were compared across RIFLE classes using the Chi² test for categorical variables and were compared non-parametrically using the Kruskal-Wallis test for continuous variables. The association between RIFLE and poor outcome was analysed using multivariable Poisson regression analysis, modified for binomial outcomes, adjusting for the aforementioned additional risk factors [31]. To adjust for any non-linearity in the association of continuous variables with outcome, fractional polynomial transformations were added to the model [32]. Cox proportional hazard regression analysis was conducted to assess the association between eAKI and one-year survival. Any violation of the proportional hazard assumption was verified and where necessary adjusted for by including an interaction term with time in the regression model. In one-year survivors we investigated the association between eAKI and low HRQoL by using a similar multivariable modified Poisson regression analysis as for the primary outcome. Rubin’s rule for pooling multiple imputation datasets was used to arrive at correct effect estimates and standard errors [27].

All statistical analyses were two-sided using a level of significance of 0.05. Statistical analyses were performed using IBM SPSS statistics software package version 21 (IBM, USA, 2012). Fractional polynomial analyses were performed in R, version 3.2.0 (R Foundation for Statistical Computing, 2015) using the “mfp” package, version 1.5.1 (2015).

During the inclusion period, 2420 out of 2601 eligible ICU patients were included in this study. One-year survival in the entire population was 1549/2420 patients (64.0 %). Of surviving patients, 1020/1549 (65.8 %) responded to the EQ-5D questionnaire. Hence the outcome status of a total of 1891/2420 (78.0 %) patients one year after ICU admission was known (either alive with known HRQoL (n = 1020) or dead (n = 871)) (see Fig. 1).

In the total study population, the median age was 59 years (interquartile range (IQR) 47–69) and 1000/2420 patients (41.3 %) were female. The majority of patients, 1418/2420 (58.6 %), were admitted for medical reasons and 1231/2420 patients (50.9 %) had a Charlson Comorbidity Index of one or higher. The median APACHE II score was 19 (IQR 14–25) (see Table 1).

Within the first 24 hours of admission 832/2420 patients fulfilled the RIFLE criteria for risk (456/2420; 18.8 %), injury (253/2420; 10.4 %) or failure (123/2420; 5.1 %). Out of the 123 patients classified as eAKI failure, 62 had renal replacement therapy initiated on the first day of admission. The median time from ICU admission to renal replacement therapy was 7 hours (IQR 4.4–14.5). Except for gender and mechanical ventilation within the first 24 hours of admission, distribution across RIFLE classes differed significantly for all baseline characteristics, with older age, frequent comorbidities and greater disease severity in the RIFLE injury and failure groups (see Table 1).

Table 2 shows the numbers of patients with one-year poor outcome, mortality, and in one-year survivors, HRQoL. Poor outcome was seen in 1157/2420 subjects (47.8 %) in the total population. In the eAKI subgroups of no eAKI, risk, injury and failure, the incidence of one-year poor outcome was 43.7 % (694/1588), 49.1 % (224/456), 59.7 % (151/253) and 72.4 % (89/123), respectively. Survival rates differed significantly between patients with no AKI and those with increasing severity of eAKI.

When compared to the patients who did not experience eAKI, the unadjusted relative risk (RR) for poor outcome was 1.12 (95 % CI 0.99, 1.26; p = 0.05) for RIFLE class, risk, 1.36 (95 % CI 1.20, 1.55; p < 0.001) for injury and 1.64 (95 % CI 1.34, 2.02; p <0.001) for failure. After adjustment for the aforementioned additional set of predictors, the association with poor outcome of RIFLE classes, injury (RR 1.14; 95 % CI 1.01, 1.29; p = 0.03) and failure (RR 1.25; 95 % CI 1.01, 1.55; p = 0.04) remained statistically significant (see Table 3). Additional file 3 contains separate analyses of the association between eAKI and poor outcome in previously defined ICU subgroups [16]. The association of eAKI with outcome did not differ across subgroups based on admission diagnoses or number of comorbidities. No clear association was seen between eAKI and poor outcome only in patients admitted with traumatic brain injury. eAKI was most frequent in ICU patients with sepsis: 21.6 % of the group of 449 patients (n = 97) experienced eAKI risk, 14.3 % (n = 64) experienced eAKI injury and 12.0 % experienced eAKI failure (n = 54).

In the Cox regression analysis, crude estimates of the hazard ratios for mortality in the eAKI RIFLE classes were 1.23 (95 % CI 1.03, 1.46; p = 0.02), 1.67 (95 % CI 1.37, 2.04; p < 0.001) and 2.45 (95 % CI 1.92, 3.13; p < 0.001) for risk, injury and failure. After adjustment for the additional predictors, the hazard ratios for RIFLE classes, injury and failure, remained statistically significant: 1.35 (95 % CI 1.10, 1.65; p = 0.004) and 1.77 (95 % CI 1.37, 2.28; p < 0.001) (see Table 3). In one-year survivors none of the RIFLE classes were independently associated with low HRQoL (see Table 3).