If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Requests for reprints should be addressed to Benjamin Griffin, MD, Internal Medicine, University of Iowa Hospitals and Clinics, 200 Hawkins Drive, Iowa City, IA 52242.
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Center for Access Delivery & Research and Evaluation (CADRE) Center, Iowa VA Health Care System, Iowa CityDepartment of Medicine, University of Iowa Carver College of Medicine, Iowa City
Acute kidney injury is prevalent among hospitalized veterans, and associated with increased risk of death following discharge. However, risk factors for death following acute kidney injury have not been well defined. We developed a mortality prediction model using Veterans Health Administration data.
Methods
This retrospective cohort study included inpatients from 2013 through 2018 with a creatinine increase of ≥0.3 mg/dL. We evaluated 45 variables for inclusion in our final model, with a primary outcome of 1-year mortality. Bootstrap sampling with replacement was used to identify variables selected in >60% of models using stepwise selection. Best sub-sets regression using Akaike information criteria was used to identify the best-fitting parsimonious model.
Results
A total of 182,683 patients were included, and 38,940 (21.3%) died within 1 year of discharge. The 10-variable model to predict mortality included age, chronic lung disease, cancer within 5 years, unexplained weight loss, dementia, congestive heart failure, hematocrit, blood urea nitrogen, bilirubin, and albumin. Notably, acute kidney injury stage, chronic kidney disease, discharge creatinine, and proteinuria were not selected for inclusion. C-statistics in the primary validation cohorts were 0.77 for the final parsimonious model, compared with 0.52 for acute kidney injury stage alone.
Conclusion
We identified risk factors for long-term mortality following acute kidney injury. Our 10-variable model did not include traditional renal variables, suggesting that non-kidney factors contribute to the risk of death more than measures of kidney disease in this population, a finding that may have implications for post-acute kidney injury care.
Acute kidney injury affects a quarter of all hospital admissions and is associated with increases in long-term mortality.
•
Given disease incidence, universal postdischarge care is not feasible, and a system to effectively triage high-risk patients is necessary.
•
This analysis makes an important contribution by outlining the risk-factors for 1-year mortality in a post-acute kidney injury population. Disease severity, commonly used for triage, was not effective at predicting 1-year mortality.
Introduction
Acute kidney injury is a common and widely recognized source of morbidity and mortality among hospitalized patients that develops in 20%-25% of admissions,
and it is associated with substantially worsened short-term outcomes, including prolonged length of stay, decreased functional capacity, increased discharge to short- or long-term care facilities, and in-hospital mortality.
In addition to its poor short-term prognosis, acute kidney injury has been increasingly associated with poor long-term outcomes, including higher rates of chronic kidney disease and mortality.
Long-term risk of adverse outcomes after acute kidney injury: a systematic review and meta-analysis of cohort studies using consensus definitions of exposure.
While acute kidney injury may be simply a marker of poor prognosis, it is also plausible that it may play a causal role in mortality, as appears to be the case in chronic kidney disease development.
Acute kidney injury is a systemic disease that impacts a number of other organ systems including the heart, lungs, liver, brain, and immune system, and these impacts can last long after hospital discharge,
Long-term risk of adverse outcomes after acute kidney injury: a systematic review and meta-analysis of cohort studies using consensus definitions of exposure.
which could be the mediators of higher mortality risk.
Assessments of risk factors for long-term mortality in acute kidney injury are generally lacking. To date, studies outlining risk factors for mortality following acute kidney injury have been conducted in select populations such as those undergoing cardiac surgery,
The incidence, risk factors and in-hospital mortality of acute kidney injury in patients after surgery for acute type A aortic dissection: a single-center retrospective analysis of 335 patients.
and therefore, applicability to the broader postdischarge population is limited. In this study, we evaluated the risk factors for death in a broad post-acute kidney injury population. We used Veterans Health Administration (VHA) hospital and laboratory data to identify important predictors of mortality and develop a model to predict mortality with a focus on maximizing parsimony and model fit.
Methods
This work was approved by the Institutional Review Board and the Research and Development Committee at the Iowa City Veterans Affairs (VA) Health Care System [IRB 200910772] as part of a larger study with previous publications.
A waiver of informed consent was granted for this retrospective study. The study follows the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) recommendations.
Data Source
We obtained data from the VA Informatics and Computing Infrastructure, an integrated infrastructure system from the VA's electronic health records. We identified admissions in VA hospitals in the Corporate Data Warehouse Inpatient files. These files include information on all hospitalizations in VA facilities nationally and contain patient demographics, diagnosis and procedure codes during the stay, admission and discharge dates and times, and discharge status. We retrieved laboratory and vital sign data from Corporate Data Warehouse files, and mortality data from the VA Vital Status File.
Study Population
We included all adult patients >18 years of age with an acute hospital admission within the VHA system from January 1, 2013 to December 31, 2018. Eligible patients were required to have at least 1 year of data within the VA system prior to the index admission, and at least one creatinine value prior to the index admission. Patients were excluded if they did not have creatinine or proteinuria data available at admission, had severe or end-stage liver disease, estimated glomerular filtration rate <30, or metastatic malignancy at the time of admission, or died during the index hospitalization (Figure).
wherein stage 1 disease is defined as an increase in creatinine of ≥0.3 mg/dL from baseline, stage 2 is a doubling of creatinine from baseline, and stage 3 is a tripling of creatinine from baseline or an absolute creatinine value ≥4.0. Baseline creatinine was the median outpatient creatinine value measured between 3 months prior and 7 days prior to the index hospitalization. If prehospitalization creatinine was not available, baseline creatinine was the admission creatinine value. Variables evaluated for inclusion were either based on known pathophysiology or epidemiologic links to the primary outcome, or were factors associated with severity and duration of in-hospital acute kidney injury. Comorbid conditions defined by Elixhauser and Quan
were identified using inpatient and outpatient claims incurred during the 12 months prior to admission. Additional laboratory values were estimated at baseline using the most abnormal value within ±24 hours of admission. We estimated glomerular filtration rate using the CKD-EPI (Chronic Kidney Disease Epidemiology Collaboration) formula.
Proteinuria was categorized as none (albumin-creatinine ratio [ACR] <30 mg/g or negative dipstick protein), mild (ACR 30-300 or dipstick protein trace ≤1+), and severe (ACR >300 mg/g or dipstick protein ≥2+) according to prior literature.
A total of 45 potential variables were evaluated for inclusion in the final model.
Primary Outcome
The primary outcome was 1-year mortality.
Derivation and Validation Cohorts
The total cohort was divided into one derivation cohort and 3 validation cohorts. The derivation cohort was composed of half of the patients admitted in the years 2014-2017, and the validation cohorts were 1) the remaining half of admissions from 2014-2017, 2) 2013 admissions, and 3) 2018 admissions.
Derivation of Models
To find an appropriate multivariable model, bootstrap resampling with a forward stepwise selection was used with a significance level of 0.01 for variable retention. A bootstrap inclusion fraction of at least 60% after 100 bootstrap replications was used to determine the potential covariates for a nonparsimonious model.
We developed a parsimonious model using best subsets method to select the best fitting set of variables using Akaike Information Criterion (AIC) as the criterion. The overall goodness of fit of the full and final model was evaluated in the derivation cohort and 3 validation cohorts using the c-statistic. We also evaluated the c-statistic for disease stage as a lone predictor of mortality.
Results
Population Characteristics
From January 2013 to December 2018, a total of 182,683 met inclusion and exclusion criteria (Figure). The primary outcome of 1-year mortality was observed in 43,414 (21.3%) eligible patients.
Baseline differences between those with mortality at 1 year and survivors are given in Table 1. Patients with mortality at 1 year were older and had a higher comorbidity burden, with greater likelihood of abnormal laboratory values at admission. The differences between the derivation and the 2014-2017 validation groups were clinically negligible. In general, the 2013 cohort had a slightly lower comorbidity burden, and the 2018 cohort had a slightly higher comorbidity burden.
Table 1Baseline Characteristics of AKI Patients by 1-Year Mortality Outcome
The reduced 10-variable model included age, chronic lung disease, non-metastatic cancer, unexplained weight loss, dementia, congestive heart failure, hematocrit, blood urea nitrogen, bilirubin, and albumin (Table 3). Notably, multiple kidney variables including proteinuria, baseline creatinine, and chronic kidney disease were not selected for the final model.
Table 3Final 10-Variable Model to Predict Mortality Within 1 Year in Patients Following an Episode of In-Hospital Acute Kidney Injury
Model fit statistics are given in Table 4. The differences in c-statistics between the nonparsimonious model and the 10-variable model were small. Stage as a univariable model had c-statistic values around 0.5 in the derivation and validation cohorts.
In this analysis that included data derived from more than 180,000 veterans, we evaluated risk factors for mortality in a broad veteran population following an episode of inpatient acute kidney injury. We found that 30 variables predicted mortality in a majority of bootstrap samples, while a 10-variable model of those variables provided a parsimonious model that performed nearly as well as the model with all 30 variables. Notably, the parsimonious model did not include acute kidney injury stage, baseline creatinine, chronic kidney disease status, discharge creatinine, or proteinuria. These findings give insight into the nature of high-risk post-acute kidney injury populations, and may have implications when looking at which patients should receive dedicated postdischarge care.
The association between acute kidney injury and long-term mortality has been established in multiple large studies in a variety of patient populations.
In a study of over 850,000 veterans who survived for 90 days following discharge, approximately 82,000 had acute kidney injury while hospitalized. Over an average follow-up time of 2.3 years, the mortality rate was 30%, compared with 16% in the control group, with an adjusted mortality risk of 1.41 (confidence interval, 1.39-1.43).
Similarly, in a meta-analysis of over 2 million patients, the association with in-hospital acute kidney injury and long-term mortality was seen in patients following angiography, cardiac surgery, and intensive care unit admission. While numerous studies have evaluated the risk factors for in-hospital mortality,
little is known about long-term risk factors for mortality after these patients are discharged from the hospital.
Here, we utilized an innovative approach to evaluate a large number of variables in order to develop a predictive model of mortality. Specifically, the parsimonious 10-variable model that was chosen through a robust method of bootstrapping, forward selection, and best subsets regression to minimize AIC included age, weight loss, chronic lung disease, dementia, congestive heart failure, hematocrit, blood urea nitrogen (BUN), bilirubin, and albumin. These findings raise several interesting questions when looking at postacute kidney injury care. Due to the very high rates of morbidity and mortality in this population, we project that dedicated postdischarge follow-up in some capacity will become part of standard care. However, given the extremely high incidence of this disease, it is not currently practical to see all such patients, and so a method of triage is necessary. Our findings suggest that within this population, markers of disease involving multiple other organ systems were more useful than acute kidney injury severity alone in identifying patients at high risk for 1-year mortality. This is a notable finding because severity is currently the most often utilized criteria for postdischarge follow-up. Two recent examples include the FUSION trial,
that also enrolled patients based solely on disease stage.
Our results may appear surprising at first because kidney factors have proven to be powerful predictors of death in the general nonhospitalized population,
Long-term risk of adverse outcomes after acute kidney injury: a systematic review and meta-analysis of cohort studies using consensus definitions of exposure.
are independently associated with higher postdischarge mortality. However, in this study we focused on the subset of patients with acute kidney injury, not the general population, and it appears within this stratification that individual kidney function parameters may become relatively less important than other markers of acute/chronic illness severity. The ASSESS-AKI trial found similar findings when looking at chronic kidney disease progression in this population.
Post-acute kidney injury proteinuria and subsequent kidney disease progression: the Assessment, Serial Evaluation, and Subsequent Sequelae in Acute Kidney Injury (ASSESS-AKI) study.
The purpose of this analysis was not necessarily to develop a postdischarge triage system or to determine which patients are most likely to benefit from postdischarge follow-up. Indeed, the inclusion of several nonmodifiable factors including dementia, weight loss, and malignancy history suggest the model is not fully optimized for these purposes. However, an evaluation of the risk factors, which were selected in an objective fashion, provide a number of important insights into the nature of the high-risk postacute kidney injury population.
Bilirubin elevations are classically associated with hepatobiliary and hemolytic diseases. Bilirubin also has antioxidant properties, and has been shown to have a U-shaped relationship with all-cause mortality, cardiovascular mortality, and cancer mortality.
The relationship between total bilirubin levels and total mortality in older adults: the United States National Health and Nutrition Examination Survey (NHANES) 1999-2004.
In this study, patients with known severe liver disease were excluded from the analysis, and the percentage of patients with mild liver disease were similar in patients who did and did not have 1-year mortality (10.8% vs 10.3%). Bilirubin categories starting at 2.0 mg/dL had a significant impact on mortality (odds ratio 2.11; confidence interval, 1.83-2.44) in the 10-variable model, and about 6% of patients had elevated bilirubin values. While bilirubin has been shown to associate with increased mortality in a nonhospitalized population,
and is a component of multiple illness severity scores including SOFA (Sequential Organ Failure Assessment) and APACHE III, it is generally not present in long-term mortality prediction scores outside of the severe liver disease population. Similarly, albumin has been shown to predict long-term mortality and is a component of some illness severity scores (APACHE III), but frequently does not appear in larger mortality prediction models. One potential explanation for the inclusion of both bilirubin and albumin is that these in combination may detect acute liver disease that does not by itself increase to the level of “severe.” For example, the ratio of albumin to bilirubin (such as in the ALBI [Albumin-Bilirubin] score) has been shown to predict postdischarge morality in patients with mild liver disease,
Abnormal serum bilirubin/albumin concentrations in dementia patients with aβ deposition and the benefit of intravenous albumin infusion for Alzheimer's disease treatment.
which is thought to be due to liver dysfunction. We suggest that our model may be showing that multiorgan dysfunction during a hospitalization, even if the individual organ dysfunction is not severe, is a major contributor to 1-year morality.
BUN was also selected for inclusion in the reduced model rather than baseline creatinine or any other kidney metrics. While the correlation of BUN and creatinine is well known, factors other than glomerular filtration can impact BUN levels, including protein intake, total parenteral nutrition, corticosteroid use, catabolism of endogenous proteins, ketoacidosis, volume status, and gastrointestinal bleeding.
and have also been shown to predict mortality independent of creatinine or other measures of kidney function. For example, in a study of 26,000 patients with creatinine values of 0.8-1.3 mg/dL, those with BUN levels >40 mg/dL had an adjusted odds ratio of 3.5 for 1-year mortality.
Similarly, in a study of older veterans that measured BUN at hospital admission, BUN was a significant predictor of mortality years after the event, even with adjustments for kidney disease.
Finally, BUN was selected for inclusion in the top 10 most predictive subsets according to AIC, while neither baseline creatinine nor acute kidney injury stage were selected in any of these subsets. We therefore think that BUN in the model is capturing illness factors beyond kidney disease severity.
When looking closely at the reduced model, as noted above, it did appear that the number of organ systems involved was more important than the severity of kidney dysfunction. In fact, our model more closely resembles mortality prediction scores for the general postdischarge population, such as the Care Assessment of Needs score,
which includes 9 of the 10 variables in our reduced model (except bilirubin). Our model features variables for multiple organ systems, including pulmonary, cardiovascular, hepatic, and neurologic systems, as well as unexplained weight loss and low albumin, which can be markers of general malnutrition and failure to thrive. The combination of organs involved thus seems to trump disease severity when predicting 1-year mortality in this population.
This study has several limitations. Because of the nature of the analysis, not all relevant variables were available for evaluation in all patients. We evaluated 45 variables for inclusion in our model, and while this is a significant amount, it does not cover the full range of possible inputs. Secondly, the use of VHA data resulted in a population that is overwhelmingly male, limiting generalizability to female patients. Lastly, we were not able to determine the cause of death within this study, which limits some of our ability to draw firm conclusions as to why certain predictors may have been more predictive in our patient population. We assume that the VA population's cause of death will be similar to a large Canadian database, wherein heart disease was the most prominent cause,
but without this information it is hard to determine with certainty why the variables in the reduced model were chosen. Notwithstanding these limitations, the study has several strengths. The model was developed in a large, population-representative cohort of adults hospitalized with acute kidney injury. The predictor variables in these risk models consist of readily available patient demographics and laboratory tests, which could allow them to be implemented into clinical practice using automated approaches within electronic medical record systems. The final model was chosen empirically using bootstrapping and forward selection followed by use of AIC to pick the best-fitting subset of variables, providing a parsimonious model that performed well and may be implemented in clinical settings more easily than a more complicated model.
Conclusions
We report the results of a 10-variable multivariable model evaluating risk factors for 1-year mortality in the postacute kidney injury population. Risk factors in this population more closely resemble risk factors for mortality in the postdischarge population than risk factors for chronic kidney disease development. These findings may have implications for which patients are most likely to see a mortality benefit postdischarge follow-up.
Long-term risk of adverse outcomes after acute kidney injury: a systematic review and meta-analysis of cohort studies using consensus definitions of exposure.
The incidence, risk factors and in-hospital mortality of acute kidney injury in patients after surgery for acute type A aortic dissection: a single-center retrospective analysis of 335 patients.
Post-acute kidney injury proteinuria and subsequent kidney disease progression: the Assessment, Serial Evaluation, and Subsequent Sequelae in Acute Kidney Injury (ASSESS-AKI) study.
The relationship between total bilirubin levels and total mortality in older adults: the United States National Health and Nutrition Examination Survey (NHANES) 1999-2004.
Abnormal serum bilirubin/albumin concentrations in dementia patients with aβ deposition and the benefit of intravenous albumin infusion for Alzheimer's disease treatment.
Values are frequency and column percent unless otherwise specified.eGFR = estimated glomerular filtration rate.
00036-0&id=gr1.jpg){kind=link}