Part 1 | Predicting Renal Recovery After Dialysis-Requiring Acute Kidney Injury Predicting Renal Recovery After Dialysis-Requiring Acute Kidney Injury

Predicting Renal Recovery After Dialysis-Requiring Acute Kidney Injury Predicting Renal Recovery After Dialysis-Requiring Acute Kidney Injury

Contact: emily.li@wecistanche.com

Benjamin J. Lee^1,2,3, Chi-yuan Hsu^1,4, Rishi Parikh⁴, Charles E. McCulloch⁵, Thida C. Tan⁴, Kathleen D. Liu^1,6, Raymond K. Hsu¹, Leonid Pravoverov⁷, Sijie Zheng^4,7 and Alan S. Go^1,4,5

¹Division of Nephrology, Department of Medicine, University of California, San Francisco, San Francisco, California, USA; ²Houston Kidney Consultants, Houston, Texas, USA; ³Houston Methodist Institute for Academic Medicine, Houston, Texas, USA; ⁴Division of Research, Kaiser Permanente Northern California, Oakland, California, USA; ⁵Department of Epidemiology and Biostatistics, University of California, San Francisco, San Francisco, California, USA; 6Division of Critical Care, Department of Anesthesia, University of California, San Francisco, San Francisco, California, USA; and ⁷Department of Nephrology, Kaiser Permanente Oakland Medical Center, Oakland, California, USA

Introduction:

After dialysis-requiring acute kidney injury (AKI-D), recovery of sufficient kidney function to discontinue dialysis is an important clinical and patient-oriented outcome. Predicting the probability of recovery in individual patients is a common dilemma.

Methods:

This cohort study examined all adult members of Kaiser Permanente Northern California who experienced AKI-D between January 2009 and September 2015 and had predicted inpatient mortality of <20%. Candidate predictors included demographic characteristics, comorbidities, laboratory values, and medication use. We used logistic regression and classification and regression tree (CART) approaches

to develop and cross-validate prediction models for recovery.

Results:

Among 2214 patients with AKI-D, the mean age was 67.1 years, 40.8% were women, and 54.0% were white; 40.9% of patients recovered. Patients who recovered were younger, had higher baseline estimated glomerular filtration rates (eGFR) and preadmission hemoglobin levels, and were less likely to have prior heart failure or chronic liver disease. Stepwise logistic regression applied to bootstrapped samples identified baseline eGFR, preadmission hemoglobin level, chronic liver disease, and age as the predictors most commonly associated with coming off dialysis within 90 days. Our final logistic regression model including these predictors had a correlation coefficient between observed and predicted probabilities of 0.97, with a c-index of 0.64. An alternate CART approach did not outperform the logistic regression model (c-index 0.61).

Conclusion:

We developed and cross-validated a parsimonious prediction model for recovery after AKI-D with excellent calibration using routinely available clinical data. However, the model’s modest discrimination limits its clinical utility. Further research is needed to develop better prediction tools.

Kidney Int Rep (2019) 4, 571–581; https://doi.org/10.1016/j.ekir.2019.01.015

KEYWORDS: dialysis, dialysis-requiring, kidney function, acute kidney injury, prediction model, renal recovery

Cistanche protects and improves kidney function

AKI-D is a serious acute medical condition that affects 3% to 13% of critically ill patients.^1–3 Although in-hospital mortality among patients with AKI-D has declined,^4–8 a sizeable fraction of survivors remain dialysis-dependent at the time of hospital discharge and beyond.^2,5,9–11 Renal recovery after AKI- D, defined as the return of sufficient kidney function to dis- continue dialysis, is an important clinical and patient-oriented outcome. Although most patients with normal baseline kidney function eventually recover if they survive the AKI-D hospitalization,12 many patients with AKI-D experience acute kidney injury (AKI) superimposed on chronic kidney disease (CKD) and do not recover.^13–15

Prediction of recovery after the onset of AKI-D is a common dilemma confronted by patients, their families, and physicians across multiple specialties, from nephrologists to intensivists, hospitalists, and primary care physicians. Baseline eGFR, proteinuria, age, diabetes mellitus, and comorbidity burden have been shown to influence the probability of recovery.9,10,13–17 The only published prediction model was constructed by Srisawat et al.,¹⁸ who found that Charlson comorbidity index and APACHE II score were predictors. However, their study was small (n ¼ 76) and included only highly selected participants enrolled in a clinical trial that excluded patients with preexisting stage 4 or 5 CKD, so generalizability was limited for multiple reasons.^19,20 Overall, data on the natural history of AKI-D are variable, and it is difficult to know whether an individual patient with AKI-D will recover.^9,21

The ability to predict recovery more accurately could potentially guide counseling and decision-making in both the inpatient and outpatient settings. Many hospitalized patients with AKI ask about their chances of recovery even before initiating acute dialysis, and some may decline to start dialysis altogether if they understand that the chances of recovery are very low and they will likely be on dialysis for the rest of their lives. Accurate prediction of recovery would inform dialysis access decisions for patients with AKI-D: both the choice of temporary versus tunneled catheters in the short term and the timing of fistula or graft placement in the medium term. Improved prognostic abilities would also influence the timing of outpatient dialysis chair placement (i.e., establishing a time and location for outpatient dialysis), which could potentially affect hospital length of stay. In the outpatient setting, when patients consider procedures that may prolong AKI-D (e.g., iodinated contrast administration), the ability to predict recovery would help patients and their providers appropriately weigh risks and benefits. From a research perspective, improved prognostic abilities would allow for targeted enrollment of patients with AKI-D who have a reasonable chance of recovery into trials testing potential treatments.

There are currently no validated AKI-D recovery prediction models, and expert panels have identified this knowledge gap as a key barrier to improving outcomes in this vulnerable population.^9,22 Using a diverse, community-based cohort, our objective was to develop a prediction model for recovery after AKI-D that would be applicable to routine clinical practice.

This study was approved by the institutional review boards at KPNC and the University of California, San Francisco, with the waiver of informed consent obtained because of the nature of the study.

Cistanche protects and improves kidney function

METHODS

Source Population

The source population was based within Kaiser Per- Permanente Northern California (KPNC), a large, integrated health care delivery system that provides comprehensive care for >4.4 million members. These patients were treated in 21 Kaiser Permanente–owned hospitals (Supplementary Appendix S1). The KPNC membership is highly representative of the surrounding local and statewide populations.23 Nearly all aspects of care are captured through KPNC’s electronic medical record system, which is integrated across inpatient, emergency department, and outpatient care settings.

This study was approved by the institutional review boards at KPNC and the University of California, San Francisco, with the waiver of informed consent obtained because of the nature of the study.

Study Sample

We conducted a retrospective cohort study of all adult (age ≥18 years) KPNC members who developed AKI-D between January 1, 2009, and September 30, 2015, and who had ≥12 consecutive months of health plan membership and pharmacy benefits before the index hospitalization to ensure adequate capture of relevant comorbidities, laboratory tests, and prescription medication use. For this analysis, we classified patients as having AKI-D if they underwent renal replacement therapy (RRT; acute intermittent hemodialysis and/or continuous RRT) during hospitalization in the absence of any preadmission chronic RRT and had peak inpatient serum creatinine concentration ≥50% of preadmission baseline (defined as the most recent non–emergency department outpatient measurement between 7 and 365 days before admission). Chronic RRT before admission was ascertained through a comprehensive KPNC End-Stage Renal Disease (ESRD) Treatment Registry that tracks initiation and cessation of RRT treatments and date(s) of renal transplantation.13,15,24,25 We excluded patients who had baseline eGFR values <15 ml/min per 1.73 m2 (because it is difficult in this eGFR range to distinguish true AKI-D from the progression of severe CKD) or predicted probability of inpatient mortality ≥20% using a KPNC-validated risk score26 (because the issue of renal recovery is clinically relevant only among those patients with AKI-D who are likely to survive the acute hospitalization and also to reduce analytic issues introduced when death can be interpreted as a state of “nonrecovery” after AKI-D). We also conducted 2 sensitivity analyses: one that did not exclude patients with a predicted probability of inpatient mortality ≥ 20% and one that used serum creatinine instead of eGFR.

Renal Recovery After AKI-D

The primary outcome was the recovery of native kidney function after AKI-D, defined as RRT independence within 90 days after RRT initiation and survival for ≥4 weeks after RRT discontinuation. Patients who stopped RRT within 4 weeks of the 90-day cutoff were observed past 90 days to confirm that they remained alive for the minimum 4-week period. We used status at 90 days because patients are conventionally considered to have ESRD if they remain dialysis-dependent for ≥90 days.9 We required that patients be alive and off dialysis for ≥4 weeks to reduce potential misclassification of people who discontinued dialysis due to withdrawal of care. Recovery could occur during the initial AKI-D hospitalization or in the outpatient setting after hospital discharge. We anchored our analysis based on the date of RRT initiation (rather than hospital discharge or some other date) to link it more closely to the natural history of the AKI episode rather than other extraneous factors that may influence length of hospitalization.

Covariates

Demographic characteristics (e.g., age, gender, self-reported race and ethnicity) were obtained from health plan databases.^27–29 Relevant comorbidities were defined by diagnostic or procedural International Classification of Diseases, Ninth Revision codes and supplemented with laboratory test results, outpatient vital signs, and prescribed medications using electronic health record-based data that were cleaned and linked at the individual-patient level into the Kaiser Permanente Virtual Data Warehouse as previously described and validated.^25,30–38 Patient vital status was determined using comprehensive information from health plan administrative and clinical databases, member proxy reporting, Social Security Administration vital status files, and California state death certificate information.^39,40 Demographic characteristics and inpatient laboratory values were measured on the date of RRT initiation for AKI-D, and baseline outpatient laboratory values and vital signs were measured 7 to 365 days before admission. For variables that had missing data, a category for missingness was created for each of those variables. Variables with >20% of values missing were not included in the modeling process.

Statistical Approach

Analyses were conducted using SAS, version 9.3 (SAS Inc., Cary, NC) and Salford Predictive Modeler, version 8.2 (Salford Systems, San Diego, CA). Baseline characteristics were compared across recovery groups using analysis of variance for continuous variables and x² tests for categorical variables.

We initially conducted multivariable logistic regression analysis for prediction of recovery after AKI-D, with the following candidate predictors: age, gender, self-reported race and ethnicity, smoking status, preadmission medication use, preexisting comorbidities (heart failure, coronary heart disease, prior ischemic stroke, peripheral artery disease, atrial fibrillation, mitral or aortic valvular disease, venous thromboembolism, hypertension, diabetes mellitus, dyslipidemia, prior hospitalized gastrointestinal bleed, thyroid disease, chronic liver disease, chronic lung disease, dementia, depression), and inpatient mortality risk score.²⁶ Additional candidate predictors included the following preadmission variables: body mass index, systolic blood pressure, preadmission high-density, and low-density lipoprotein levels, eGFR (using the Chronic Kidney Disease Epidemiology Collaboration creatinine equation⁴¹), dipstick proteinuria level, hemoglobin level, and platelet count. Body mass index, preadmission systolic blood pressure, and all laboratory-based variables were treated as ordinal categorical variables (partitions between categories shown in Tables 1 and 2; similar results were obtained when these covariates were treated as continuous variables). To identify important predictors, we first generated 1000 random samples of the analytic cohort through bootstrap resampling with replacement and then conducted automated stepwise logistic regression on each sample. Predictors that were selected by stepwise regression in ≥75% of the bootstrapped samples were included in the final model. We subsequently used 10-fold cross-validation to generate predicted probabilities of recovery for each patient, which were used to calculate a c-index and generate calibration statistics. Finally, model parameter estimates and odds ratios for the final set of predictors were generated through a logistic regression model using the full analytic cohort.

We also planned a priori to perform a CART analysis for recovery because it was not known whether this method would yield more clinically useful results than the logistic regression approach.⁴² Candidate predictors were the same as those used in the logistic regression analysis. CART treated all laboratory values as continuous variables and optimally selected cut-points to minimize information loss. No limits were set on a minimum node or terminal size. Trees were pruned and optimized using built-in 10-fold cross-validation to minimize the relative misclassifi-cation of cases while protecting against overfitting. C- indices, a confusion matrix, and a receiver operating characteristic curve was generated to evaluate the performance of the final decision tree.

For Part 2, please click here.