Predictors for Augmented Renal Clearance in a Heterogeneous ICU Population as Defined by Creatinine and Cystatin C

Abstract

Introduction:

The incidence of augmented renal clearance (ARC) in the ICU is highly variable, and the identification of these patients remains challenging.

Objective:

The objective of this study is to define the incidence of ARC in a cohort of critically ill adults, using serum creatinine and cystatin C, and to identify factors associated with its development.

Methods:

Retrospective cohort study of critically ill patients without stage 2 or stage 3 acute kidney injury with both serum creatinine and cystatin C available. The incidence of ARC was defined as a CKD-EPI_{Creatinine-Cystatin C} estimated glomerular filtration rate (eGFR) >130 mL/min. A multivariable logistic regression model using a penalized lasso method was fit to identify independent predictors of ARC.

Results:

Among the 368 patients included in the study, the indication for ICU admission was non-operative in 55% of patients, and 9% of patients were admitted for major trauma. The overall incidence of ARC was low at 4.1%. In a multivariable logistic regression model, Charlson Comorbidity Index, major trauma, intracerebral hemorrhage, age, and SOFA score were found to predict ARC.

Conclusion:

The incidence of ARC in this study was low, but prediction models identified several factors for early identification of patients with risk factors for or who develop ARC, particularly in a cohort with a low baseline risk of ARC. These factors could be used to help identify patients who may develop ARC.

Introduction

An accurate assessment of kidney function in critically ill patients is a challenging proposition but essential for the optimization of medication use. Although dose adjustment in the setting of acute kidney injury (AKI) has been the focus of research among these patients, the opposite end of the spectrum, augmented renal clearance (ARC), defined as an estimated glomerular filtration rate (eGFR) of >130 ml/min, is also a clinical challenge (1, 2). Glomerular hyperfiltration associated with ARC results in more rapid elimination of renally-eliminated medications (e.g., antimicrobials, anticoagulants), which decreases drug exposure and heightens the risk for clinical failure (1, 3, 4). Although it is becoming more widely recognized, the reported incidence of ARC is highly variable, and more evidence is necessary to identify risk factors associated with its development (3, 5, 6).

One possible explanation for the variable reported incidence of ARC is the different biomarkers and equations that have been utilized to estimate GFR. The diagnosis of ARC is typically made through the use of serum creatinine, yet the predictive capability of this marker for ARC is suboptimal in the intensive care unit (ICU). Creatinine is insensitive to asymptomatic acute changes in the filtration function, and there is a well documented 24–48 hour delay in creatinine response to the underlying functional change (7, 8). Cystatin C, an alternative endogenous GFR marker, has also been assessed for the identification of ARC. In a study that compared systemic biomarkers to urine creatinine clearance, cystatin C better predicted ARC than creatinine (9). The published literature on ARC has failed to consider the combination of creatinine and cystatin C, which has been shown to more accurately predict measured GFR in stable ambulatory patients than when either biomarker is used in isolation (10, 11).

Patient-specific factors that indicate a high risk for ARC in previous studies include younger age, male sex, lower severity of illness, and vasopressor therapy (2, 5, 12–14). Disease-specific risk factors have included sepsis, trauma, general surgery, neurosurgery, febrile neutropenia due to hematologic malignancy, burn injury, and cystic fibrosis (13). Collectively, however, these estimates have poor precision, and the relationships are often studied in specific patient populations at higher risk for ARC. There is a need to more broadly characterize factors for ARC based on contemporary biomarker availability in a mixed ICU patient population.

The purpose of this study, therefore, is to describe the incidence of ARC, defined by CKD-EPI_{Creatinine-Cystatin C} eGFR > 130 mL/min, and derive a prediction model for its development.

Materials and Methods

This was a retrospective cohort study of adults admitted to the intensive care unit at Mayo Clinic, in Rochester, MN, from October 2008 through May 2015. Included individuals were patients with available serum creatinine and cystatin C concentrations checked during their ICU admission as part of three studies for which the detailed eligibility criteria have been previously published (15–17). Briefly, included patients were critically ill adults with sepsis, shock, or major trauma who had no evidence of acute kidney injury (KDIGO stage 2 or 3). Individuals in whom the creatinine and cystatin C were drawn greater than seven days after ICU admission were excluded to establish a temporality between baseline severity of illness and disease states and the outcome of interest. During the study timeframe, no specific cystatin C protocol was used at the study center, and the evaluation was at the discretion of the inter-professional care team. Creatinine clearance calculated from a 24-hour urinary collection is not routinely performed at the study center. The Mayo Clinic Institutional Review Board approved the protocol, and the requirement for informed consent was waived.

The primary endpoint of interest in this study was ARC, as defined by an eGFR > 130 mL/min (1, 3, 13). The CKD-Epi_Creatinine (2009), CKD-EPI_{Cystatin C} (2012), and CKD-EPI_{Creatinine-Cystatin C} (2012) equations, unadjusted for body surface area (BSA), were used to calculate eGFR; however, the CKD-EPI_{Creatinine-Cystatin C} equation (2012) was used for the primary analysis given the superior accuracy for prediction of measured GFR (10, 18, 19). It is recommended that eGFR measurements not be adjusted for body surface area for purposes of drug dosing, and as the intent of this model is to predict ARC for identification of patients with a potential need for medication adjustment or monitoring, the eGFR cutoff for ARC diagnosis was not adjusted for BSA(20–22). A secondary univariate analysis using eGFR adjusted for BSA was also performed.

Data extracted from the electronic health record included age, sex, race, height, admission weight, BSA, body mass index (BMI), and severity of illness [acute physiology and chronic health evaluation (APACHE III) score, sequential organ failure assessment (SOFA) score]. Chronic comorbidities, particularly factors which would affect creatinine concentrations (e.g., paraplegia, quadriplegia), were collected. Factors related to the acute illness and previously reported as predictive of ARC were collected, including the need for vasopressors, traumatic brain injury during the index hospitalization, cystic fibrosis, intracerebral hemorrhage, and febrile neutropenia. We also collected the cardiac output obtained by echocardiography within 48-hours of biomarker assessment when available as we hypothesized that hyperdynamic cardiac function (EF ≥ 70%) would be associated with ARC (23). Creatinine (mg/dL) measurement was assayed using the standardized (isotope dilution mass spectrometry traceable) enzymatic creatinine assay (Roche, Basel, Switzerland). Cystatin C (mg/L) was measured with a particle-enhanced turbidimetric assay (Gentian AS, Moss, Norway). This assay is traceable to the same international certified cystatin C reference material (ERM-DA471/IFCC) used to develop the cystatin C-based CKD-EPI equations. All data were collected in a password-protected Excel database designed expressly for the aims under study.

Summary statistics were reported, as appropriate for the variable type. Continuous data was described with the mean ± standard deviation (SD) or median and interquartile range (IQR). Categorical variables, including the incidence of ARC, were presented as frequencies and percentages. Logistic regression models were used to estimate associations with ARC status. Because of the small number of ARC patients relative to the number of predictors, to help prevent overfitting penalized lasso methods with 10-fold cross-validation were used to build a multivariable logistic regression model to predict ARC. Bootstrap resampling was used to obtain confidence intervals for the model parameters. Model discrimination was assessed using the area under the ROC curve. Model calibration was assessed by plotting the observed ARC risk versus the average predicted ARC risk for predicted ARC risk deciles. Summaries for ICU and hospital length of stay after ICU admission were done using the competing risk extension of the Kaplan-Meier method, where death was the competing risk for being discharged alive. Analyses used SAS version 9.4 statistical software (SAS Institute Inc., Cary, NC, USA) and the glmnet package in R (R: A language and environment for statistical computing, R Core Team, R Foundation for Statistical Computing, Vienna, Austria, URL https://www.R-project.org) (24). P < 0.05 was considered statistically significant.

Results

A total of 398 patients were available during the study time frame, of which 30 were excluded leaving a total of 368 patients available for analysis (Figure 1). Included patients had a mean age of 65±15 years, 93% of patients were Caucasian, 57% were male, and 45% and 24% required mechanical ventilation and vasopressors, respectively. The indication for ICU admission was non-operative in 55% of patients (Table 1), which included a respiratory (N=102), cardiovascular (N=36), gastrointestinal (N=15), genitourinary (N=15), or other (N=27) primary ICU diagnoses. On admission, 14 (4%) patients had an intracerebral hemorrhage, and 34 (9%) individuals were admitted with major trauma [12 of 34 (35%) with a diagnosis of traumatic brain injury]. One hundred twenty patients (33%) had a formal echocardiogram within 48-hours of the kidney biomarkers being drawn. Fifteen of these patients (13%) were noted to have hyperdynamic cardiac function (EF≥70%). Only two patients in the cohort had a history of quadriplegia. There were no patients who met inclusion criteria with febrile neutropenia diagnosis on hospital admission or a history of cystic fibrosis.

Figure 1.

Consort Diagram for Study Inclusion. A total of 368 patients were included in the study analysis, 15 with augmented renal clearance (ARC) and 353 without ARC.

Table 1.

Patient Demographics

Demographic	ARC (N = 15)	No ARC (N = 353)	N = 368	P-Value
Age (years, median±IQR)	37.4 (33.2–43.6)	67.2 (57.4–76.9)	66.8 (55.7–76.6)	<0.0011
Male sex	9 (60%)	199 (56.4%)	208 (56.5%)	0.7812
Caucasian	14 (93.3%)	328 (92.9%)	342 (92.9%)	0.9512
BMI (kg/m2; median±IQR)	27.6 (21.2–31.2)	27.2 (23.5–30.9)	27.2 (23.3–30.9)	0.7251
Non-operative Admission Diagnosis	6 (40%)	198 (56.1%)	204 (55.4%)	0.2202
TBI on Hospital Admit	3 (20%)	9 (2.5%)	12 (3.3%)	<0.0012
Major Trauma	5 (33.3%)	29 (8.2%)	34 (9.2%)	<0.0012
Intracerebral Hemorrhage	3 (20%)	11 (3.1%)	14 (3.8%)	<0.0012
Sepsis on ICU Admission	4 (26.7%)	142 (40.2%)	146 (39.7%)	0.2932
APACHE III score (median±IQR)‡	46 (36–50)	67 (54–81)	67.6±21.2	<0.0011
Charlson Comorbidity Index (median,IQR)	0 (0–2)	5 (3–7)	5 (3–7)	<0.0011
SOFA score ICU Day 1(median,IQR)‡	3.5 (1–6)	5 (3–8)	5 (3–8)	0.0721
Vasopressors on ICU Admission	3 (20%)	84 (23.8%)	87 (23.6%)	0.7352
Mechanical Ventilation ICU Day 1	8 (53.3%)	156 (44.2%)	164 (44.6%)	0.4852

Data presented as frequency (percentage) except where indicated otherwise.BMI (body mass index); TBI (traumatic brain injury); APACHE (Acute Physiology and Chronic Health Evaluation); SOFA (Sequential Organ Failure Assessment)

¹Kruskal Wallis

²Chi-Square

^‡9 missing values

Incidence of ARC

The median time from ICU admission to biomarker assessment was 1.5 days (IQR 0.2–2.4) and, in the patients with ARC, the biomarkers were drawn within one day of ICU admission in 83% of cases. Mean serum creatinine and cystatin C were 0.9±0.3 mg/dL and 1.4±0.6, respectively. The serum creatinine was ≤1 mg/dL in 276 (75%) patients. The incidence of ARC ranged from 3.3–7.9% depending on the eGFR equation used (Figure 2). The ranges of calculated eGFR were 15–29 ml/min= 12 (3.3%), 30–59 ml/min= 122 (33.2%), 60–89 ml/min= 121 (32.9%), 90–119 ml/min= 89 (24.2%), 120–149 ml/min= 21 (5.7%) ≥ 150 ml/min= 3 (0.8%). Median ICU length of stay (LOS) was four days (range 0, 108) versus three days (range 1, 17) in non-ARC versus ARC patients and hospital LOS after ICU admission was eleven days (range 1, 127) versus eight days (range 4, 26). In-hospital mortality occurred in 50 (14%) patients without ARC and 0 (0%) patients with ARC.

Figure 2.

Incidence of Augmented Renal Clearance (ARC). The incidence of ARC varied depending on the eGFR equation used; CKD-EPI_Creatinine 29 patients (7.9%) , CKD-EPI_{Cystatin C} 12 (3.3%), CKD-EPI_{Creatinine-Cystatin C} 15 (4.1%).

Prediction models for ARC

Lower Charlson Comorbidity Index, younger age, lower APACHE III score, traumatic brain injury (TBI), major trauma at ICU admission, intracerebral hemorrhage, and lower SOFA score were each associated with ARC in univariate models (Table 2). Predictors of ARC, when adjusted for BSA, were similar (Table S1). We were unable to assess the impact of quadriplegia or hyperdynamic cardiac function on ARC prediction as no patients with ARC had these diagnoses. In the multivariable penalized logistic regression model, Charlson Comorbidity Index, major trauma, intracerebral hemorrhage, age, and SOFA score were found to predict ARC (Table 3). APACHE III score was not included in the model to optimize the usability of the prediction model in daily clinical practice. The multivariable model was converted to a score to predict the probability of ARC within seven days of ICU admission (Table 3).

Table 2.

Univariate Predictors for ARC

Characteristic	OR (95% CI)	p-value
Charlson Comorbidity Index	0.35 (0.21–0.51)	<0.001
APACHE III	0.94(0.9–0.97)	<0.001
TBI	9.56 (1.94–37.1)	0.009
Major Trauma	5.59 (1.65–16.88)	0.008
Intracerebral Hemorrhage	7.77 (1.61–29.06)	0.014
Age	0.89 (0.86–0.93)	<0.001
SOFA Score (ICU Admit)	0.83 (0.66–1)	0.048
Operative Admission Diagnosis	1.9 (0.67–5.82)	0.221
BMI	0.97 (0.89–1.05)	0.51
Vasopressors (ICU Admit)	0.8 (0.18–2.59)	0.73
Mechanical Ventilation (ICU Admit)	1.44 (0.51–4.2)	0.487
Sepsis	0.54 (0.15–1.62)	0.28
Male Gender	1.16 (0.41–3.53)	0.781

Table 3.

Penalized Multivariable Regression Predictors for ARC

Characteristic	OR (95% CI)
Charlson Comorbidity Index	0.80 (0.16–1)
Major Trauma	1.75 (1–16.59)
Intracerebral Hemorrhage	2.82 (1–69.1)
Age	0.94 (0.83–1)
SOFA Score (ICU Admit)	0.99 (0.52–1)

ARC score = 1.07 – 0.22*Charlson Comorbidity Index + 0.56*Major trauma + 1.04*Intracerebral hemorrhage – 0.07*Age – 0.01*SOFA

ARC Probability = e^{ARC score}/(1 + e^{ARC score})

Figures 3 and S1 depict model discrimination and calibration, respectively. Overall the model showed excellent discrimination for prediction of ARC with an AUC of 0.95 (95% CI 0.92–1) in this mixed-ICU population. The median (IQR) predicted probability in the patient cohort using the ARC probability equation in Table 3 was 0.01 (0.004–0.03), while the minimum and maximum were 0.0003 and 0.76. Assuming sensitivity and specificity are equally important, then a calculated predicted probability of 0.087 might be a good cutoff value to classify patients (i.e., predict ARC if probability≥0.087) because it maximizes the sum of sensitivity and specificity (sensitivity=0.86, specificity=0.94).

Figure 3.

ROC Analysis. Multivariable model performance for prediction of augmented renal clearance (ARC) [AUC of 0.95 (95% CI 0.92–1)], in a population at low risk for ARC.

Discussion

The incidence of ARC in the ICU in the literature is highly variable depending on the patient population (3, 5, 6, 14). Many variables have been found to be associated with the ARC. However, most ARC studies fail to include a mixed ICU patient population, and ARC remains challenging to identify clinically (2, 5, 12, 13). This study included patients from a variety of ICU settings with less clear potential for ARC. In fact, many of the included patients had acute conditions that heighten the risk for not only ARC but also AKI (e.g., sepsis, shock), a dichotomous clinical scenario challenging for clinicians to navigate. We identified patients with ARC using the CKD-EPI_{Creatinine-Cystatin C} equation, and successfully developed a prediction score that included Charlson Comorbidity Index, major trauma, intracerebral hemorrhage, age, and SOFA score. Given the impact of ARC on renally-eliminated medications and risk for clinical failure due to decreased drug exposure, it is essential to appropriately identify these patients early in their ICU stay (1, 3, 4). Existing literature has failed to utilize the cystatin C and serum creatinine combination to identify patients with ARC, which may be a more accurate marker of kidney function compared to cystatin C alone. Additionally, the majority of previous studies have focused on specific patient populations known to be high risk for ARC and fail to study the clinically challenging patients that fall in the middle, with risk factors for both AKI and ARC. A clinical score to identify at-risk patients early in the ICU stay would allow clinicians to tailor antimicrobials, adjust other renally-eliminated medications, and provide enhanced monitoring earlier in the ICU stay.

Historically other studies have defined ARC according to calculated creatinine clearance via serum creatinine or measured urine creatinine clearance (1–4, 6, 12, 14, 25, 26). It is known that creatinine is a suboptimal measure for acute changes in kidney function in critically ill individuals and lags behind changes in GFR (7, 8). Urine creatinine clearance overestimates true GFR depending on the patient’s end-organ function due to tubular secretion, and accuracy may be jeopardized by incomplete urine collection (10, 27–29). Also, creatinine lacks specificity for GFR and is affected by a number of non-renal factors including age, weight, sex, ethnicity, the presence of acute or chronic inflammation, underlying skeletal muscle mass, dietary intake, and magnitude of fluid resuscitation (10, 30–38). Cystatin C is an endogenous biomarker utilized to calculate eGFR, which is less affected by some of the above non-renal determinants and has been shown to have better predictive performance than serum creatinine-based equations in critically ill patients. In the few studies of ARC that have attempted to evaluate the performance of cystatin C, it has also been shown to better predict ARC than creatinine [cystatin C AUC= 0.92 (95% CI 0.86–0.98); creatinine AUC = 0.86 (95% 0.79–0.94)] (9, 39). While the best cystatin-C based equation has not been fully elucidated in critically ill patients the combination creatinine-cystatin C based equation, CKD-EPI _{Creatinine-Cystatin C} eGFR equation was shown to most accurately reflect measured GFR in a heterogeneous group of 6471 stable outpatients across the spectrum of kidney function (7, 8, 19). Additionally, a recent study found that using the CKD-EPI _{Creatinine-Cystatin C} eGFR equation to determine vancomycin dosing in critically ill patients achieved goal vancomycin trough concentrations better compared to standard dosing strategies using serum creatinine (16, 17). A recent systematic review supports the notion that estimates of GFR inclusive of cystatin C may be better at predicting drug clearance compared to creatinine(40). Given this information, the current study utilized the CKD-EPI _{Creatinine-Cystatin C} eGFR equation for defining ARC. As expected, the incidence of ARC was variable depending on the equation used.

An ARC score was previously described by Udy et al. in a population of septic and trauma ICU patients using measured urine creatinine clearance (12). Risk factors for ARC included age ≤50 years (6 points), trauma (3 points), and SOFA score of ≤ 4 (1 point). Patients with scores of 0–3 were considered low risk, 4–6 points intermediate-risk, and 7–10 points high risk. Akers et al. later validated this ARC score in a modified version (low score ≤6 versus high score ≥7) using piperacillin/tazobactam pharmacokinetic clearance data in critically ill surgical and trauma patients (41). The score was found to be 100% sensitive and 71.4% specific for detecting increased piperacillin/tazobactam clearance and volume of distribution and decreased AUC with a ROC score of 0.86. This study, however, included only 13 patients and was validated in a specific subset of ICU patients. More recently a study in a mixed ICU patient population (203 medical ICU patients) of 446 patients found an ARC prevalence of 25% and found that trauma, young age, and male sex were independent risk factors for ARC.(14) We found similar factors for predicting ARC including age, major trauma, and SOFA score, but added additional factors including the Charlson Comorbidity Index and intracerebral hemorrhage. Other unique preditors collected in this study included hyperdynamic cardiac function, quadriplegia, and cystic fibrosis. These were sufficiently infrequent, especially in the patients with ARC, such that they were excluded from multivariable models. While hyperdynamic cardiac function is thought to be a mechanism for ARC, no patients with hyperdynamic cardiac function detected by echocardiography in our study had ARC. The pathophysiology of ARC is complex, and the preferred measure to detect hyperdynamic circulation in ARC remains poorly defined. Given the small number of patients with available echocardiograms we are unable to draw firm conclusions regarding this association, or lack thereof. Risk factors such as the need for vasopressors, mechanical ventilation, sepsis, and male sex have intermittently been shown to be associated with ARC but were not found to be significant in our cohort of patients. Additionally, a wide variety of ICU patients were included in this study with only ~9% of patients admitted with trauma and more than half of the sample was non-operative admission indications which expands upon the previously published works.

The incidence of ARC in our study is much lower than previously reported, possibly related to the population included in our study (42). Patients included in the study had no evidence of acute kidney injury KDIGO Stage 2 or 3 based on inclusion criteria; however, patients with lower GFRs were eligible for inclusion so as not to bias the sample. In addition, the study population had an increased medical/surgical mix of patients, perhaps including patients with a lower likelihood of developing ARC. According to mean APACHE III, Charlson Comorbidity Index score, and SOFA score, it seems the included patients had a moderate degree of severity of illness and baseline comorbidities, which may have also impacted the incidence of ARC compared to previous studies. Despite limitations in sample size, this study still found similar risk factors to previous studies, which supports the validity of its findings.

Limitations of this study include the small sample size over the included timeframe, which may have contributed to selection bias, the small number of patients with ARC which limits the ability to identify risk factors, and the retrospective nature of the study prohibiting the ability to perform measured urinary clearance of filtrate markers, which have not been compared directly to cystatin C. Importantly, however, there is no gold standard for measurement of kidney function in the ICU and inulin, iohexol, I-iothalamate filtrate makers are expensive and complex for utilization in critically ill and may not be valid in this population. Building a model with a low event rate has inherent limitations. However, a penalized lasso method with 10-fold cross-validation was used to limit the influence of overfitting and enhance generalizability. Notably, the model is more appropriate for predicting ARC in populations with a lower incidence of the syndrome or unpredictable kidney function (e.g., mixed ICU population at risk for both ARC and acute kidney injury). It may be less suitable for populations with a higher risk for ARC, but more predictable kidney function (e.g., younger sample, limited comorbid conditions, surgical patient population). Also, certain information was not available retrospectively. For example, only 120 patients had echocardiograms temporally near biomarker collection. This limitation of available data did not allow us to evaluate certain risk factors previously shown to be associated with ARC (e.g., hyperdynamic cardiac function). Due to the lack of routine daily cystatin C trending in this cohort, we were unable to provide insight into the duration of ARC in the population. Lastly, the ARC score in this study needs to be independently validated in a separate cohort.

Conclusion

The incidence of ARC, as defined by CKD-EPI _{Creatinine-Cystatin C} > 130 mL/min, in a mixed cohort of critically ill patients, was lower than previously reported. A risk prediction score for an ARC within seven days of ICU admission was developed to help clinicians identify patients at risk for ARC. This study builds on previous literature by including a heterogeneous group of critically ill patients and utilizing a novel method of calculating ARC using serum creatinine and cystatin C-based equation. Knowing factors associated with ARC is an additional tool for clinicians to use when considering medication dose optimization and monitoring. Further validation is needed to determine the external validity of the ARC score before routine implementation in clinical practice.