Abstract
BACKGROUND:
Acute kidney injury (AKI) is a common and serious complication of cardiac surgery for which unrecognized heterogeneity may underpin poor success in identifying effective therapies. We aimed to identify phenotypically similar groups of patients as defined by their postoperative creatinine trajectories.
METHODS:
This was a retrospective, single center cohort study in an academic tertiary care center including patients undergoing coronary artery bypass graft procedures. AKI phenotypes were evaluated through latent class mixed modeling of serum creatinine patterns (trajectories) in post-cardiac surgery patients. To identify trajectory phenotypes modeling was performed using postoperative creatinine values from 50% of subjects (development cohort), and for comparison similarly conducted for the remaining sample (validation cohort). Subsequent assessments included comparisons of classes between development and validation cohorts for consistency and stability, and among classes for patient and procedural characteristics, complications and long-term survival.
RESULTS:
Twelve AKI trajectories were identified in both the development (n=2647) and validation cohorts (n=2647). Discrimination among classes was good (mean posterior class membership probability 66–88%), with differences in rate, timing and degree of serum creatinine rise/fall, and recovery. In matched class comparisons between cohorts, many other phenotypic similarities were present. Notably, four high-risk phenotypes had greater long-term mortality rates relative to lower risk classes.
CONCLUSIONS:
Latent class mixed modeling identified 12 reproducible post-cardiac surgery AKI classes (serum creatinine trajectory phenotypes) including four with higher risk of poor outcome. Such hidden structure offers a novel approach to grouping post-cardiac surgery patients for renoprotection investigations in addition to re-analysis of previously conducted trials.
Classifications: acute kidney injury, cardiac surgery, biomarkers, heterogeneity, latent class mixed modeling
Graphical Abstract
In settings where organ protection strategies have been disappointing, research that identifies novel disease subphenotypes has generated new hypotheses and recognized important heterogeneity in response to intervention. Such approaches to acute respiratory distress syndrome, and sepsis have advanced understanding, using novel analytic approaches to distinguish un-noticed, but important, subgroupings [1, 2]. Acute kidney injury (AKI) is a common serious complication of cardiac surgery, but the search for effective AKI interventions has been disappointing [3–5]. However, evident sources of perioperative renal insult (e.g., atheroembolism, systemic inflammation, toxins) suggest AKI heterogeneity may be important [4], and prevalent serum creatinine and/or urine output-based AKI definitions (i.e., thresholds [6, 7]) do not allow for differentiation based on temporal variation in biomarker levels.
Serum creatinine “trajectories” (longitudinal postoperative patterns) may contain useful information to expose obscure AKI subphenotypes. Unbiased statistical modeling tools such as latent class mixed modeling (LCMM) are useful to identify hidden sub-structure within heterogeneous data. The application of LCMM to look for clusters of patients with similar creatinine trajectories in patient cohorts relative to AKI is unexplored and represents a novel framework to re-explore this common perioperative complication. We used LCMM within a robust development-validation approach to test the hypothesis that unrecognized serum creatinine trajectory subgroups exist in a cohort of post-cardiac surgery patients.
PATIENTS AND METHODS
After Institutional Review Board approval (Duke Health IRB; Pro00086586), we conducted a retrospective analysis of prospectively collected data from the Duke electronic medical record. Subjects included all patients 18 years and older who underwent isolated, non-emergent coronary artery bypass graft surgery with cardiopulmonary bypass between 1/1/2000 and 12/31/2009 at Duke University Medical Center (Durham, NC). For subjects undergoing more than one procedure during the study period, only the earliest procedure was included for analysis. Preoperative exclusions included subjects with missing preoperative serum creatinine values, those receiving renal replacement therapy, and individuals with unstable cardiac disease reflected by ventricular assist device support, myocardial infarction within 24 hours prior to surgery, preoperative inotrope infusion requirement, cardiac arrest prior to surgery during the index hospitalization, or a diagnosis of cardiogenic shock. Additional exclusions related to postoperative factors included death prior to the fifth postoperative day (due to modeling of trajectories through day 4), and reoperation after the initial cardiac surgical procedure. Included subjects were subdivided randomly into equally sized development and validation cohorts for analysis.
Basic demographic and laboratory parameters were obtained from the Duke electronic medical record. Additional patient and procedural characteristics (i.e., comorbidities, postoperative complications, European System for Cardiac Operative Risk Evaluation [EURO score][8], postoperative complications, intraoperative and postoperative transfusion) with prespecified registry definitions were collected as reported to the Society of Thoracic Surgeons Adult Cardiac Surgery Database (https://www.sts.org/registries-research-center/sts-national-database/adult-cardiac-surgery-database).
As part of standard clinical care, serum creatinine was determined in the hospital laboratory using a dry slide enzymatic reflectance technique (Vitros 950, Johnson and Johnson, New Brunswick, NJ) with a normal range of 0.5–1.4 mg/dl. Preoperative serum creatinine was defined as the pre-operative value closest value within 30 days prior to, but not on, the day of surgery. Per institutional protocol, serum creatinine was measured daily throughout the postoperative course. If multiple assessments occurred on one day, the earliest value was selected to avoid bias. AKI was defined using serum creatinine KDIGO criteria for the preoperative and first four postoperative days [6]. Mortality data was derived from the United States Social Security Death Index. Patient survival status was assessed for three years postoperatively at which point censoring occurred.
ANALYSIS PLAN
Summary statistics were generated for continuous (mean [standard deviation], or median [quartile 1 – quartile 3]), and categorical parameters (frequency [%]). While serum creatinine data for the primary analysis was complete, some patient and procedural characteristics for secondary analyses were incomplete (Table 1). Multiple imputation with chained equations was used to impute missing values for use in secondary analyses (mice package in R)[9]. Imputation models used predictive mean matching (continuous parameters) or multinomial logistic regression (categorical parameters) and included only pre- and intraoperative parameters. Estimates were pooled from multiply imputed datasets when comparing these parameters across classes and in Cox proportional-hazards regression using the rms[10] and miceadds[11] packages in R.
Table 1:
Patient, Procedural, and Postoperative Characteristics
| Development Cohort (n = 2,647) | Validation Cohort (n = 2,647) | Standardized Difference | |
|---|---|---|---|
| Patient Characteristics | |||
| Age, yr | 64.0 (±10.8) | 63.9 (±10.8) | −0.01 |
| Female, No. (%) | 763 (28.8%) | 764 (28.9%) | 0.00 |
| Race, No. (%) | |||
| white | 2,070 (78.2%) | 2,033 (76.8%) | −0.03 |
| black | 425 (16.1%) | 447 (16.9%) | 0.02 |
| other | 152 (5.7%) | 167 (6.3%) | 0.02 |
| Congestive Heart Failure, No. (%) | 466 (17.6%) | 464 (17.5%) | 0.00 |
| Chronic Obstructive Lung Disease, No. (%) | 372 (14.1%) | 362 (13.7%) | −0.01 |
| Cerebrovascular Disease, No. (%) | 300 (11.3%) | 312 (11.8%) | 0.01 |
| Diabetes, No. (%) | 929 (35.1%) | 926 (35.0%) | 0.00 |
| Hypertension, No. (%) | 1,985 (75.0%) | 2,029 (76.7%) | 0.04 |
| History of Myocardial Infarction, No. (%) | 1,127 (42.6%) | 1,093 (41.3%) | −0.03 |
| Peripheral Vascular Disease, No. (%) | 395 (14.9%) | 368 (13.9%) | −0.03 |
| Prior Cardiac Surgery, No. (%) | 137 (5.2%) | 125 (4.7%) | −0.02 |
| Smoking History, No. (%) | 1,255 (47.4%) | 1,251 (47.3%) | 0.00 |
| Logistic EuroSCORE | 4.6 (2.6 – 8.5) | 4.5 (2.5 – 8.3) | −0.02 |
| Preoperative Hematocrit | 39.28 (±4.92) | 39.50 (±4.90) | 0.05 |
| Ejection Fraction, No. (%) | |||
| < 45 | 679 (25.7%) | 649 (24.5%) | −0.03 |
| 45–55 | 746 (28.2%) | 710 (26.8%) | −0.04 |
| > 55 | 1,050 (39.7%) | 1,144 (43.2%) | 0.07 |
| Procedural Characteristics | |||
| Intraoperative Transfusion (any products), No. (%) | 844 (31.9%) | 809 (30.6%) | −0.03 |
| Bypass Time, minutes | 116.0 (95.0 – 141.0) | 115.0 (95.0 – 141.0) | −0.02 |
| Aortic Cross Clamp Time, minutes | 67.0 (52.0 – 85.0) | 67.0 (51.0 – 85.0) | −0.02 |
| Postoperative Complications | |||
| Major Cardiovascular Complication, No. (%) | 17 (0.6%) | 23 (0.9%) | 0.03 |
| Infectious Complication, No. (%) | 30 (1.1%) | 45 (1.7%) | 0.05 |
| Neurologic Complication, No. (%) | 44 (1.7%) | 49 (1.9%) | 0.01 |
| Pulmonary Complication, No. (%) | 77 (2.9%) | 105 (4.0%) | 0.06 |
| Postoperative Transfusion, No. (%) | 1,151 (43.5%) | 1,196 (45.2%) | 0.03 |
| Renal Characteristics | |||
| Baseline Serum Creatinine, mg/dl | 1.0 (0.9 – 1.2) | 1.0 (0.9 – 1.2) | −0.03 |
| Baseline eGFR, mL/min | 71.2 (±21.7) | 72.3 (±21.6) | 0.05 |
| Acute Kidney Injury (KDIGO), No. (%) | 1,036 (39.1%) | 1,036 (39.1%) | 0.00 |
| Mortality | |||
| 30 days, No. (%) | 28 (1.1%) | 31 (1.2%) | −0.02 |
| 3 years, No. (%) | 285 (10.8%) | 262 (9.9%) | −0.03 |
Data is presented as mean (SD), median (Q1 - Q3) or number (%) as appropriate. Major cardiovascular complication includes cardiac arrest, cardiac tamponade, pulmonary embolism, myocardial infarction and aortic dissection. Infectious complication includes sepsis, wound infection, mediastinitis, and pneumonia. Neurologic complication includes coma, cerebrovascular accident or transient ischemic attack. Pulmonary complication includes requirement for tracheal reintubation and extended time on the ventilator (> 24 hours). Intraoperative transfusion includes red blood cells, platelets, fresh frozen plasma, and cryoprecipitate. Abbreviations: EURO - European System for Cardiac Operative Risk Evaluation, eGFR – estimated glomerular filtration rate (CKD-EPI). Standardized differences (added per reviewer request) are rounded to two decimal places. These values represent the difference in means/proportions for each variable between the validation and development cohorts in units of the variable’s pooled standard deviation. A standardized difference of < 0.1 is considered to signal a negligible difference between groups.
Primary Analysis: Latent Class Mixed Modeling
Briefly, LCMM analysis is an extension of linear mixed modeling, with the additional assumption that several latent classes exist within a population, each with a distinct class-specific longitudinal trajectory [12]. In the current study, LCMM was used to explore for subgroups of patients with relatively distinct serum creatinine trajectories following isolated, non-emergent coronary artery bypass graft surgery. An a priori definition of a benchmark “stable creatinine” class (class 1) with less than 10% serum creatinine variation, to reflect minimal AKI, was introduced to facilitate comparison with other classes reflecting elevation or decline.
The lcmm package in R[12] was used to construct LCMMs from daily serum creatinine data for the first four postoperative days in the development cohort. Creatinine was modeled as a change relative to baseline ([serum creatinine – baseline] / baseline * 100%). Since variance at baseline is 0%, this time point was excluded from models to allow for appropriate consideration of intra-subject correlation, with postoperative day 1 serving as the model intercept. Given the non-Gaussian distribution of relative serum creatinine change from baseline, values were transformed using a basis of quadratic I-splines estimated simultaneously with the LCMM [12]. To account for intra-subject correlation of repeated measures, a first-order autoregressive correlation process was included in the model. Class-specific trajectories were modeled as cubic functions in time without adjustment for covariates. Models with increasing numbers of latent classes were generated with the optimal class number selected using the model with the lowest Bayesian Information Criterion. To increase the likelihood of appropriate convergence, each model was initialized with 100 sets of random starting values. Patients were a posteriori assigned to the class for which their probability of class membership was highest. Model discrimination was assessed using the mean posterior class membership probability (MPCMP) derived from a multinomial logistic class membership model. In a latent class mixed model, each patient receives a vector of probabilities for membership within each class and is subsequently assigned to the class with the highest probability. The MPCMP is a class-specific metric that represents the mean value of the probability for class membership for patients assigned to that class. For example, if there are four patients in class 1 and their respective probabilities of class 1 membership are 0.7, 0.8, 0.8, and 0.9, the MPCMP for class 1 will be 0.8 (the average of the four values). Thus, MPCMP can be viewed as a metric of uncertainty in patient class assignment and thus model discrimination, with higher values representing better discrimination and less uncertainty.
LCMM-generated Classes: Characteristics, Postoperative Complications and Survival Analysis
Comparisons among classes identified in the primary analysis for patient and procedural characteristics and postoperative complications used Kruskal-Wallis (continuous parameters) and Chi-square (categorical parameters) tests. Similarly, survival curves (from postoperative day 5 onward) were developed using the Kaplan-Meier method and compared using the log-rank test. Following visual inspection, unadjusted Kaplan-Meier curves were empirically clustered into survival risk groups with similar profiles. Survival differences between these groups was assessed using Cox proportional-hazards regression adjusted for perioperative comorbidities, baseline creatinine, and maximum relative change in postoperative creatinine.
Validation Cohort Secondary Analyses
To assess generalizability of primary model findings from the development cohort, two analytic approaches were conducted in the remaining subjects (validation cohort). First, the abovementioned LCMM process was repeated (repeat validation approach). Second, class definitions from the primary analysis were directly applied to the validation cohort, with subjects a posteriori assigned to the class for which they had highest posterior probability of membership (applied validation approach). Results from each approach were compared to the primary analysis for the number of classes and qualitative class similarities. All analyses were completed in R (version 3.3.2) [13].
RESULTS
From the study period, 5294 subjects undergoing coronary artery bypass graft surgery procedures met inclusion and exclusion criteria (Figure 1). Patient and procedural characteristics and outcomes for the development (n=2647) and validation (n=2647) cohorts are shown in Table 1. In the entire cohort there were 547 patient deaths during 14937 person-years of follow-up data.
Figure 1:
Flow chart of study population. Patients were excluded for meeting 1 or more of the exclusion criteria. Abbreviation: CABG – coronary artery bypass graft surgery.
The primary LCMM analysis identified 12 distinct creatinine trajectory phenotypes (i.e., latent classes) in the development cohort (Figure 2A, Figure 3, Supplemental Table 1). Discrimination amongst classes was good with MPCMP ranging from 66 to 88% (Supplemental Table 2). Serum creatinine trajectory classes were heterogeneous by direction, degree and rate of day-by-day serum creatinine rise (or decline), timing of peak (or nadir), and presence or absence of recovery. Class descriptions were assigned qualitatively:
Class 1: stable creatinine
Class 2: minor nadir with sustained decline
Class 3: modest nadir with sustained decline
Class 4: mild rise, peak on postoperative day 1, with recovery
Class 5: mild rise, peak on postoperative day 2, with recovery
Class 6: mild rise, peak on postoperative day 2, without recovery
Class 7: mild rise, peak on postoperative day 1–2, without recovery
Class 8: mild rise, peak on postoperative day 3 or 4
Class 9: modest rise, peak on postoperative day 2, with recovery
Class 10: modest rise, peak on postoperative day 2, without recovery
Class 11: severe rise, peak on postoperative day 3 or 4, with late rise
Class 12: severe rise, peak on postoperative day 3 or 4, with early rise
Figure 2:
Serum creatinine trajectories identified in the 12-class model in the (A) development cohort and (B) validation cohort (repeat validation approach). Colored lines represent the estimated class-specific creatinine trajectories. Colored shapes represent the observed class-specific mean creatinine values. Dashed lines outline a linear transition from baseline to the initial point on each trajectory (postoperative day 1).
Figure 3:
Class specific and individual serum creatinine trajectories stratified by class in the development cohort. Colored lines represent the estimated mean class-specific creatinine trajectories. Black lines (spaghetti plots) represent individual patient creatinine trajectories.
Both validation cohort analyses strongly supported these primary findings. Repeat validation LCMM analysis identified 12 patient clusters, also with good discrimination (MPCMP 68–95%; Supplementary Figure 1, Supplementary Table 3), and serum creatinine trajectories that highly resembled those observed in the development cohort (Figure 2B). The applied validation LCMM analysis also demonstrated good discrimination among patient clusters (MPCMP 68–92%; Supplementary Figure 2, Supplementary Table 4).
Among the 12 latent classes identified within the development cohort, patient and procedural factors, postoperative complications, and long-term outcome were heterogeneous, with notable common features within some classes, but few distinguishing features in others (Supplementary Table 5, Supplementary Figures 5–7). Signature characteristics by class included: a) Class 3, over-representation of females and elevated baseline serum creatinine, b) Class 4, the youngest, c) Class 11, elevated baseline serum creatinine and logistic EURO scores, low preoperative hematocrit values, and elevated rates of transfusion (intra- and postoperative), and pulmonary complications, and d) Class 12, baseline elevated rate of diabetes and low hematocrit levels, longer cardiopulmonary bypass and aortic cross-clamp durations, higher transfusion rates (intra- and postoperative), and pulmonary complications. Similar findings were evident in validation cohort analyses (Supplementary Tables 6–7, Supplementary Figures 8–13) further supporting the possible generalizability of phenotype clustering by serum creatinine trajectory.
Postoperative survival was notably heterogeneous across classes. By simple empirical inspection of survival curves, a subset of four classes (3, 10, 11, and 12) demonstrated much poorer survival relative to the remaining classes (2, 4, 5, 6, 7, 8, and 9) and the “stable creatinine” class (class 1) (Figures 3A, B). Notably, of the four classes with higher risk for mortality, three (classes 10, 11 and 12) were characterized by relatively substantial elevation of postoperative creatinine while one (class 3) was characterized by a moderate and sustained decline in creatinine. High-risk group membership in the development sample was associated with higher mortality rates using Cox proportional-hazard regression (HR 3.10 [95% CI, 1.89–5.07]), and this association persisted after multivariable adjustment (baseline comorbidities and creatinine, intraoperative factors, and peak postoperative serum creatinine rise) (aHR 2.38 [95% CI, 1.35–4.20]). Similar additional risk for these classes was evident in the validation cohort, including the repeat (Figure 4C, Supplementary Figure 3; HR 4.07 [95% CI, 2.29–7.22]; aHR 2.24 [95% CI, 1.19–4.22]) and applied validation analyses (Figure 4D, Supplementary Figure 4; HR 4.07 [95% CI, 2.25–6.93]; aHR 2.44 [95% CI, 1.33–4.48]).
Figure 4:
Kaplan-Meir survival curves for the 12 trajectory classes identified in the development cohort primary analysis, displayed separately (A) and clustered into 3 groups (B): stable, low risk and high risk AKI trajectory risk groups. Secondary findings from the validation cohort, depict similarly clustered survival curves, from the repeat (C), and the applied (D) validation analyses (see text).
COMMENT
In a large cohort of post-cardiac surgery patients, we identified 12 novel AKI trajectory phenotypes with distinct postoperative serum creatinine trajectories. Trajectory phenotypes varied not only by timing and character of serum creatinine rise (or fall) but also several other features including patient and procedural characteristics, postoperative complications, and long-term outcome. Importantly, four classes associated with markedly poorer survival were identified. Two validation analyses probing an equivalently large patient cohort identified 12 similar trajectory phenotypes supporting generalizability of these primary findings. Use of an unbiased methodology to identify previously hidden AKI phenotypes among post-cardiac surgery patients offers an alternate approach to classifying (defining) and investigating perioperative renal injury.
Although no reports have previously explored longitudinal postoperative serum creatinine trajectories, two renal recovery studies have characterized patterns of “post-peak decline” in serum creatinine. Bhatraju and colleagues described a “non-resolving AKI” subgroup among critically ill patients that was associated with increased mortality [14]. These authors developed their criteria from several a priori defined candidates in a development cohort and used the best performing of those to successfully predict mortality risk in a validation cohort. Similarly, Kellum and colleagues used an a priori approach to define five serum creatinine-based renal recovery subtypes, all of which were significantly associated with increased 1-year mortality [15]. As with the current study, both publications aimed to subgroup AKI episodes for better understanding. However, the current analysis strategy fundamentally differs in that its agnostic approach is a purely model-driven strategy for identifying distinct trajectory classes. This unbiased approach eliminates any potential for misleading assumptions regarding creatinine patterns that may even undermine the search for clusters among patients with distinct AKI subtypes.
The description of 12 reproducible AKI trajectory phenotypes within a cohort of patients following cardiac surgery at a single center is a novel finding of the current study. Such observations of heterogeneity among AKI cases may advance understanding and the search for targeted therapies for this vexing condition. While prevalent AKI diagnostic tools employ mostly single parameter thresholds, hidden phenotypes as described in this study are missed by such traditional criteria. Notably, phenotype refers to observable clinical characteristics while endotype refers to underlying pathophysiologic process or mechanism. For example, heart failure phenotypes include those with preserved and reduced left ventricular ejection fraction; whereas, chronic hypertension and amyloidosis are both endotypes of the heart failure phenotype with preserved left ventricular ejection. While clustering by phenotype can improve classification and in some cases treatment, in settings where this re-classification strategy has advanced care, endotype discovery has offered the most promise for targeting therapeutic interventions. Pertinent to the current study, beyond phenotype reproducibility in other cohorts, further investigations are needed to seek endotypes from among the 12 AKI phenotypes described. Such an approach has advanced knowledge for other important disorders with otherwise disappointing results from major research initiatives; among these, acute respiratory distress syndrome. Latent class analysis of acute respiratory distress syndrome has identified two broadly reproducible phenotypes [16] and re-grouping subjects from previously negative randomized therapeutic trials for re-analysis has yielded potentially important phenotype-related differential responses to therapy [1, 17]. Similar approaches are also advancing understanding for sepsis and asthma [2, 18].
The current study is a valuable first step in exploring the relevance of phenotypes to AKI, but subsequent investigations are needed; for example, to understand how such findings relate to already-studied potential prophylactic and/or therapeutic renoprotection interventions. Furthermore, elucidation of pathophysiologic mechanisms that underlie the clustering of patients into trajectory phenotypes may expose AKI endotypes among cardiac surgery patients. While the more simplistic high- vs. low-risk trajectory paradigm outlined in this study provides sufficient clinical prognostic value, the identification of more specific and nuanced trajectory classes is invaluable with respect to future identification of clinically and therapeutically heterogenous endotypes that may advance the treatment and prevention of AKI. Additionally, the transferability of AKI phenotypes from cardiac surgery to other high-risk populations requires investigation (e.g., critical illness, major trauma, cardiac valve surgery, etc.).
The current study is limited by its retrospective quality and single-center nature. Notably, most demographic, procedural, postoperative complication and outcome data was sourced from standardized prospectively-collected information that contributes to an audited tool for national quality measurement and improvement activities (i.e., Society of Thoracic Surgeons Adult Cardiac Surgery Database; https://www.sts.org/national-database). Furthermore, the reproducibility of the study findings and the development-validation approach using data from a large clinical cohort supports the stability and generalizability of the study findings. Nonetheless, while such findings are novel and interesting, equivalent findings in other samples are required. Notably, the present analysis describes creatinine trajectories only for the first four postoperative days. Since routine discharge home for such procedures often occurs at the reference institution on postoperative day 5, LCMM analysis beyond this point, while robust to missing data within the modelling period, would have been severely limited, not only by the burden of missing serum creatinine data but also by sample bias. Furthermore, the typical timing of renal insult related to cardiac surgery mostly relates to procedure-related filtration impairment AKI, which manifests through serum creatinine rise within the first 24–48 hours postoperatively, and should be sufficiently reflected in the current analysis. Finally, the analysis presented does not examine the association between AKI trajectory phenotype and persistence of renal dysfunction (i.e., new chronic kidney disease or the requirement for renal replacement therapy). Data describing these outcomes are not readily available. Nonetheless, this study is the first to describe agnostically identified subgroups among post-cardiac surgery serum creatinine trajectories, which are associated with heterogeneity in patient, procedural and recovery characteristics, as well as important survival differences.
In summary, using an agnostic statistical approach and a development-validation methodology to assess postoperative serum creatinine trajectories, we identified 12 reproducible AKI trajectory phenotypes with good discrimination in a large cohort of patients undergoing non-emergent coronary artery bypass graft procedures. These trajectory phenotypes also demonstrated many similarities among members including patient and procedural characteristics, recovery profiles and long-term outcomes. Most notably, four high-risk classes had markedly poorer long-term survival. Such a strategy for clustering cases among post-cardiac surgery patients offers an intriguing approach to grouping patients for categorizing and investigating perioperative AKI. More research is required to confirm generalizability of these findings among and beyond the cardiac surgery setting and to seek pathophysiologic endotypes and the promise of better renoprotection strategies.
Supplementary Material
Conflicts of Interest and Source of Funding:
Funding:
Research reported in this publication was supported by the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Numbers TL1TR001116 (to BYA), T32GM008600 (to ADC and JRP) and P30AG028716 (to CFP and JFP). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Footnotes
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Meeting Presentation: Presented in part at the Annual Meeting of the Society of Cardiovascular Anesthesiologists (Phoenix, AZ; 2018)
Conflicts of Interest: none
REFERENCES
- 1.Calfee CS, et al. , Acute respiratory distress syndrome subphenotypes and differential response to simvastatin: secondary analysis of a randomised controlled trial. Lancet Respir Med, 2018. 6(9): p. 691–698. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Gardlund B, et al. , Six subphenotypes in septic shock: Latent class analysis of the PROWESS Shock study. J Crit Care, 2018. 47: p. 70–79. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Bellomo R, Kellum JA, and Ronco C, Acute kidney injury. Lancet, 2012. 380(9843): p. 756–66. [DOI] [PubMed] [Google Scholar]
- 4.Wang Y and Bellomo R, Cardiac surgery-associated acute kidney injury: risk factors, pathophysiology and treatment. Nat Rev Nephrol, 2017. 13(11): p. 697–711. [DOI] [PubMed] [Google Scholar]
- 5.Kellum JA and Prowle JR, Paradigms of acute kidney injury in the intensive care setting. Nat Rev Nephrol, 2018. 14(4): p. 217–230. [DOI] [PubMed] [Google Scholar]
- 6.Khwaja A, KDIGO clinical practice guidelines for acute kidney injury. Nephron Clin Pract, 2012. 120(4): p. c179–84. [DOI] [PubMed] [Google Scholar]
- 7.Bellomo R, et al. , Acute renal failure - definition, outcome measures, animal models, fluid therapy and information technology needs: the Second International Consensus Conference of the Acute Dialysis Quality Initiative (ADQI) Group. Crit Care, 2004. 8(4): p. R204–12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Nashef SA, et al. , EuroSCORE II. Eur J Cardiothorac Surg, 2012. 41(4): p. 734–44; discussion 744–5. [DOI] [PubMed] [Google Scholar]
- 9.van Buuren S and Groothuis-Oudshoorn K, mice: Multivariate Imputation by Chained Equations in R. Journal of Statistical Software, 2011. 45(3): p. 1–67. [Google Scholar]
- 10.Harrell F, rms: Regression Modeling Strategies. 2018.
- 11.Robitzsch A, Grund S, and Henke T, miceadds: Some additional multiple imputation functions, especially for mice. 2018.
- 12.Proust-Lima C, Philipps V, and Liquet B, Estimation of Extended Mixed Models Using Latent Classes and Latent Processes: The R Package lcmm. Journal of Statistical Software, 2017. 78(2): p. 1–56. [Google Scholar]
- 13.R Core Team, R: A Language and Environment for Statistical Computing, in R Foundation for Statistical Computing. 2016: Vienna, Austria. [Google Scholar]
- 14.Bhatraju PK, et al. , Acute kidney injury subphenotypes based on creatinine trajectory identifies patients at increased risk of death. Crit Care, 2016. 20(1): p. 372. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Kellum JA, et al. , Recovery after Acute Kidney Injury. Am J Respir Crit Care Med, 2017. 195(6): p. 784–791. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Calfee CS, et al. , Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials. Lancet Respir Med, 2014. 2(8): p. 611–20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Famous KR, et al. , Acute Respiratory Distress Syndrome Subphenotypes Respond Differently to Randomized Fluid Management Strategy. Am J Respir Crit Care Med, 2017. 195(3): p. 331–338. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Bochenek G, et al. , Certain subphenotypes of aspirin-exacerbated respiratory disease distinguished by latent class analysis. J Allergy Clin Immunol, 2014. 133(1): p. 98–103 e1–6. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.