A Simple Clinical Tool for Stratifying Risk of Clinically Significant CKD after Nephrectomy: Development and Multinational Validation

Significance Statement

Patients undergoing surgical management of kidney tumors are at increased risk of developing CKD. However, it is often difficult to identify patients at higher risk of clinically significant CKD before surgery, and there is a lack of validated tools to assist clinicians in this process. The authors developed and validated a simple scoring system that accurately and reproducibly stratifies risk of developing clinically significant CKD after nephrectomy on the basis of readily available parameters. This system provides an evidence-based quantitative tool for clinicians to balance the risk of CKD against other considerations when planning management of kidney tumors, and it will facilitate earlier identification of patients with a higher risk of developing clinically significant CKD, potentially leading to earlier intervention.

Abstract

Background

Clinically significant CKD following surgery for kidney cancer is associated with increased morbidity and mortality, but identifying patients at increased CKD risk remains difficult. Simple methods to stratify risk of clinically significant CKD after nephrectomy are needed.

Methods

To develop a tool for stratifying patients’ risk of CKD arising after surgery for kidney cancer, we tested models in a population-based cohort of 699 patients with kidney cancer in Queensland, Australia (2012–2013). We validated these models in a population-based cohort of 423 patients from Victoria, Australia, and in patient cohorts from single centers in Queensland, Scotland, and England. Eligible patients had two functioning kidneys and a preoperative eGFR ≥60 ml/min per 1.73 m². The main outcome was incident eGFR <45 ml/min per 1.73 m² at 12 months postnephrectomy. We used prespecified predictors—age ≥65 years old, diabetes mellitus, preoperative eGFR, and nephrectomy type (partial/radical)—to fit logistic regression models and grouped patients according to degree of risk of clinically significant CKD (negligible, low, moderate, or high risk).

Results

Absolute risks of stage 3b or higher CKD were <2%, 3% to 14%, 21% to 26%, and 46% to 69% across the four strata of negligible, low, moderate, and high risk, respectively. The negative predictive value of the negligible risk category was 98.9% for clinically significant CKD. The c statistic for this score ranged from 0.84 to 0.88 across derivation and validation cohorts.

Conclusions

Our simple scoring system can reproducibly stratify postnephrectomy CKD risk on the basis of readily available parameters. This clinical tool’s quantitative assessment of CKD risk may be weighed against other considerations when planning management of kidney tumors and help inform shared decision making between clinicians and patients.

Worldwide, the kidney is the 12th and 16th most common site of primary cancer for men and women, respectively.¹ Most patients with kidney cancer are managed with either radical or partial nephrectomy,² which carries attendant risks of CKD. Postnephrectomy CKD is associated with increased likelihood of all-cause^3,4 and cancer-specific mortality.⁵

Although many studies have evaluated kidney function following nephrectomy, only a few unvalidated clinical tools for predicting postoperative kidney function are available.^6–9 There is also a paucity of recommendations regarding postnephrectomy CKD risk in international guidelines such that follow-up planning is predominantly on the basis of oncologic risk, with minimal regard for kidney function.^2,10

There is a need for evidence-based methods to assess risk of clinically significant CKD following nephrectomy because such methods could guide preemptive/preventative management. The aim of this study was to develop and validate a simple clinical score using variables routinely available in preadmission/preoperative clinics, which could accurately and reproducibly stratify the risk of developing clinically significant CKD 12 months postnephrectomy.

Methods

Study Population

Patients undergoing radical or partial nephrectomy from five distinct cohorts were included: one derivation cohort and four validation cohorts. Reporting was consistent with the Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement. We excluded patients (Supplemental Table 1) if they had missing data on pre-/postoperative eGFR, had an anatomically abnormal contralateral kidney (determined on imaging and considered likely to affect differential kidney function; considered on a patient by patient basis), were not managed surgically, had a preoperative eGFR <60 ml/min per 1.73 m², required kidney replacement before surgery, or had a previous nephrectomy. Serum creatinine concentration was recorded within 1 month before surgery and on clinical follow-up as close to 12 postoperative months as possible (specified in Results). The Chronic Kidney Disease Epidemiology Collaboration equation was used to calculate eGFR from serum creatinine.

The derivation cohort included 699 patients managed surgically for renal cell carcinoma across 38 centers in Queensland (Australia) between January 2012 and December 2013. Validation cohort 1 included 423 patients with renal cell carcinoma from 79 centers in Victoria (Australia) within the same timeframe. Ascertainment was through whole of population sampling using state-based registries. A single value for serum creatinine was recorded on follow-up. Data collection was retrospective and described previously.¹¹

Validation cohort 2 was a prospectively recruited convenience sample of 179 patients managed for kidney tumors between June 2013 and January 2018 at a single tertiary hospital (Princess Alexandra Hospital) in Brisbane, Australia. Data collection was described previously.¹² Twenty-four patients presumably overlapping with the derivation cohort were excluded in sensitivity analyses. Serum creatinine values were recorded at 12 postoperative months. Because serum creatinine can be variable and affected by AKI, which may lead to inaccurate estimates of GFR, serum creatinine data were scrutinized against longitudinal trends and concurrent medical documentation for cases where multiple serum creatinine measurements were recorded at 12 months after nephrectomy, to identify values that accurately represented GFR. Serum creatinine values were considered unreliable if they increased from any known stable postoperative baseline value by ≥1.5 times or >26.5 μmol/L and subsequently returned to baseline or in the absence of a known baseline, if there was documentation indicating potential inaccuracy (e.g., intercurrent critical illness or oliguria), with subsequent reductions in creatinine observed at a later time point.

Validation cohorts 3 and 4 were retrospectively ascertained single-center convenience samples of patients who underwent nephrectomy for kidney tumors. Data were collected from chart review for clinical audit purposes. Validation cohort 3 included 205 patients from Western General Hospital, Edinburgh, Scotland (from January 2002 to December 2013). Clinicians recorded stable values for serum creatinine 12 months after surgery (exact dates were not recorded). Validation cohort 4 included 221 patients from Frimley Park Hospital, Frimley, England (from January 2002 to December 2013).

Population-based data for 1866 living kidney donors from Australia/New Zealand who donated a kidney between January 2004 and December 2017 were included as an exploratory analysis because these patients are also undergoing nephrectomy but are not affected by kidney tumors. Data were obtained from the Australia and New Zealand Dialysis and Transplant Registry.

Outcome

The primary outcome was postoperative eGFR <45 ml/min per 1.73 m² (incident stage ≥3b CKD) at 12 postoperative months. This outcome was considered clinically significant because, on the basis of large observational studies,³ it is associated with increased risk of mortality and further eGFR decline following nephrectomy and because reaching this threshold is an indication for changes to CKD management/surveillance protocols.^13,14 The chosen follow-up time was important because it is a common time point for postoperative review; therefore, patients potentially at risk could be targeted for earlier follow-up.²

Model Derivation

Multivariable logistic regression models were used to evaluate associations with CKD. The logistic coefficient was later used to develop clinical scores. Predictors were selected on the basis of background knowledge/clinical reasoning, focusing on parameters that would plausibly be available in outpatient settings. Because only 2% of patients in the derivation cohort were excluded due to missing data on predictors, we did not perform multiple imputation. Candidate predictor variables were age, sex, preoperative eGFR, diabetes mellitus, and nephrectomy type. If a diagnosis of diabetes prior to surgery was not documented, patients were assumed not to have diabetes. Both categorical and continuous measures were assessed in models where appropriate. Preoperative eGFR was grouped in 10-unit intervals from 60 to 90 ml/min per 1.73 m²; values higher than this were grouped because most laboratories do not specify values above this threshold. Living donors were scored as having undergone radical nephrectomy. We screened for interactions between selected predictors using logistic regression models, including two variables and a first-order multiplicative interaction term, evaluating all possible two-variable combinations of selected predictors; no interactions were observed. Visual tests for collinearity were performed; age was significantly correlated with preoperative eGFR, and it was instead included as a dichotomous predictor (<65 and ≥65 years old). Additional age categories using thresholds older than 65 years old were not considered because there were relatively small numbers of older patients undergoing nephrectomy, which would have adversely affected model precision. Stepwise deletion of nonsignificant variables with low effect size was undertaken, favoring parsimony to simplify models for clinical applications. Multivariable models run in the derivation cohort were compared objectively using Akaike and Bayesian information criteria, Harrell c statistic, and Brier score. Analysis was performed using Stata versions 13.0 and 14.2 (StataCorp, College Station, TX).

Clinical Score and Risk Stratification

We developed a clinical score using the multivariable model determined to have the most clinical utility on the basis of objective measures and clinical judgement by assigning points proportional to the logistic coefficient rounded to the nearest integer.¹⁵ Scores were grouped into risk strata on the basis of predicted likelihood of the primary outcome, with thresholds selected by plotting predicted probabilities against the clinical score. We used log-binomial models to estimate relative risks of the primary outcome across risk strata in the derivation cohort. The score was also evaluated using univariable logistic regression, both by considering each risk stratum as a four-level categorical variable, and the clinical score as a continuous variable. Postoperative kidney function was compared descriptively among risk strata.

External Validation and Sensitivity Analyses

External validation of initial multivariable models was undertaken in validation cohorts 1 and 2. Sensitivity analyses were done in patients from the derivation cohort and validation cohort 1 who had stage T1 or T2–T4 tumors to assess the influence of disease severity. A sensitivity analysis in validation cohort 2 was also conducted, excluding patients managed before January 2014 (presumed overlap with derivation cohort). Following score development, data from two international nephrectomy cohorts were obtained for external validation. An exploratory analysis was also carried out using a cohort of living kidney donors.

To validate the clinical score, univariable logistic regression models were run, considering the score as a continuous variable. Discrimination was assessed using receiver operating characteristics curves. We assessed calibration visually, plotting predicted probability and observed absolute risk of the primary outcome against the clinical score for all cohorts. A calibration belt was used to facilitate formal hypothesis testing of model calibration,¹⁶ and Brier scores were calculated as an alternate measure of general model fit.

Ethical Considerations

For the derivation cohort, approval was obtained under the Queensland Public Health Act (RD004755) and by the human research ethics committees of Metro South Hospital and Health Service (HREC/13/QPAH/226) and QIMR Berghofer Medical Research Institute (P1529). For validation cohort 1, approval was granted by the human research ethics committee of Cancer Council Victoria (HREC-0911). For validation cohort 2, approval was granted by the human research ethics committee of Metro South Hospital and Health Service (HREC/05/QPAH/95; HREC/16/QPAH/353). Patients provided written informed consent. For validation cohorts 3 and 4, data were obtained for the purposes of service evaluation, and they were considered exempt from local institutional ethical review on the basis of guidance from the National Health Service (NHS) Health Research Authority; transfer of nonidentifiable data was deemed exempt from Caldicott Guardian approval on the basis of advice obtained from NHS Lothian and Frimley Health NHS Foundation Trust. Human research ethics approval for data analysis was obtained from the University of Queensland (2018002648). For living kidney donors, data access was designated low risk and approved by the University of Queensland (2017-SOMILRE-0193).

Results

Study cohort characteristics are presented in Supplemental Table 2. Postoperative kidney function was recorded at median (interquartile range [IQR]) follow-up times of 11.9 (IQR, 10.2–13.2), 11.7 (IQR, 7.1–13.7), 11.4 (IQR, 8.1–13.1), 17.4 (IQR, 12.0–26.2), and 14.0 (IQR, 12.0–22.0) months in the derivation cohort; validation cohorts 1, 2, and 4; and living kidney donors, respectively. Patients in validation cohort 3 were followed up as close to 12 postoperative months as possible; exact dates were not recorded in the dataset.

Table 1 shows results from logistic regression analysis. All predictors had statistically significant associations with the primary outcome in univariable models, with the exception of sex. The first two multivariable models included all prespecified predictors; the only difference was handling of preoperative eGFR as a categorical or continuous variable. In the derivation cohort, the c statistic was 0.84 for both models. There were no substantial differences in Akaike or Bayesian information criteria. A third model was developed that excluded sex and considered eGFR as a categorical variable. The c statistic for this model was 0.84 in the derivation cohort. Because performance was not inferior to other models, this model was selected as the most clinically useful. Model performance was similar across validation cohorts (Supplemental Figure 1) and in sensitivity analyses. All models performed better for patients with stage T1 tumors than in patients with advanced disease (Table 1).

Table 1.

Risk prediction models for postoperative eGFR <45 ml/min per 1.73 m² after nephrectomy

	β (95% CI)
	Univariable	Model 1	Model 2	Model 3
Age, yr
<65	Reference	Reference	Reference	Reference
≥65	1.49 (1.01 to 1.97)	0.66 (0.12 to 1.20)	0.67 (0.13 to 1.20)	0.66 (0.12 to 1.20)
Sex
Women	Reference	Reference	Reference
Men	0.26 (−0.24 to 0.76)	0.09 (−0.46 to 1.34)	0.08 (−0.46 to 0.63)	—
Diabetes mellitus
No	Reference	Reference	Reference	Reference
Yes	0.59 (0.02 to 1.15)	0.65 (−0.04 to 1.32)	0.61 (−0.08 to 1.31)	0.66 (−0.04 to 1.35)
Preoperative eGFR, ml/min per 1.73 m2
60–69	3.43 (2.70 to 4.16)	3.15 (2.13 to 4.16)		3.16 (2.15 to 4.17)
70–79	2.73 (2.01 to 3.46)	2.41 (1.43 to 3.40)	—	2.41 (1.43 to 3.40)
80–89	1.80 (1.02 to 2.58)	1.75 (0.70 to 2.80)		1.76 (0.72 to 2.79)
≥90	Reference	Reference		Reference
Preoperative eGFR, ml/min per 1.73 m2
Per unit	−0.09 (−0.11 to −0.07)	—	−0.08 (−0.11 to −0.06)	—
Nephrectomy type
Partial	Reference	Reference	Reference	Reference
Radical	1.89 (0.07 to 2.81)	1.80 (0.80 to 2.80)	1.81 (0.82 to 2.80)	1.79 (0.80 to 2.78)
Intercept		−5.88 (−7.10 to −4.67)	2.82 (0.65 to 5.00)	−5.82 (−7.01 to −4.64)
Model performance measures, derivation cohorta
Akaike information criterion		407.48	404.10	405.58
Bayesian information criterion		443.88	431.40	437.43
Brier scoreb		0.08	0.09	0.09
P value		0.47	0.43	0.47
Harrell c statistic		0.84 (0.81 to 0.88)	0.84 (0.80 to 0.88)	0.84 (0.80 to 0.88)
Harrell c statistic, validation/sensitivity analyses
Validation cohort 1c		0.84	0.84	0.83
Validation cohort 2d		0.83	0.84	0.85
Stage T1 RCCe		0.86	0.86	0.85
Stage T2–T4 RCCf		0.81	0.81	0.81
Sensitivity analysisg		0.84	0.85	0.85

Data are presented as logistic coefficient (β) and 95% CI estimated using logistic regression. Supplemental Table 5 has the results presented as odds ratios. All variables are included in the models presented. Models 1 and 2 included all listed variables, differing only in handling preoperative eGFR as a continuous or categorical variable. Model 3 did not include sex and considered preoperative eGFR as a categorical variable. The 95% CI for Harrell c statistic was reported only for the derivation cohort. 95% CI, 95% confidence interval; —, variable was excluded from the respective model; RCC, renal cell carcinoma.

^aPatients with RCC from Queensland (Australia) managed surgically between January 2012 and December 2013 (n=699).

^bValues of 0.00 reflect perfect forecasting; a P value of 0.05 reflects good model fit.

^cPatients with RCC from Victoria (Australia) managed surgically between January 2012 and December 2013 (n=423).

^dPatients with kidney tumors managed surgically at the Princess Alexandra Hospital (Queensland, Australia) between June 2013 and January 2018 (n=179).

^ePatients from the derivation cohort and validation cohort 1 with stage T1 tumors.

^fPatients from the derivation cohort and validation cohort 1 with tumors more advanced than stage T1.

^gAnalysis conducted in validation cohort 2 excluded 24 patients who were managed before January 2014 and presumed to overlap with the derivation cohort. Data were deidentified, and therefore, overlap was not able to be confirmed directly.

We derived a clinical score on the basis of the most parsimonious multivariable model. Four groups were defined on the basis of predicted likelihood of a postoperative eGFR <45 ml/min per 1.73 m² (Tables 2 and 3): negligible (zero to three points; 0.8% likelihood), low (four to six points; 6% likelihood), moderate (seven to eight points; 23% likelihood), and high (nine to ten points; 51% likelihood) (Supplemental Table 3).

Table 2.

A clinical score to stratify risk of postnephrectomy eGFR <45 ml/min per 1.73 m²

Variable	Coefficient	Points	Score	Risk Stratum	Predicted Probabilitya	Observed Absolute Riskb	Relative Risk (95% CI)c
Age, yr			0	Negligible	0.8% (0.2% to 1.6%)	<2%	0.05 (0.00 to 0.35)
<65	0.00	0	1
≥65	0.66	1	2
Diabetes mellitus			3
No	0.00	0
Yes	0.66	1	4	Low	6% (3% to 10%)	3%–14%	1.00
Preoperative eGFRd			5
60–69	3.16	5	6
70–79	2.41	4
80–89	1.76	3	7	Moderate	23% (17% to 29%)	21%–26%	2.71 (1.51 to 4.91)
≥90	0.00	0	8
Management plan
Partial nephrectomy	0.00	0	9	High	51% (43% to 59%)	46%–69%	5.93 (3.30 to 10.64)
Radical nephrectomy	1.79	3	10

Logistic coefficients are from model 3 (Table 1). To derive the points system, each coefficient was divided by the lowest value (0.66) and rounded to the nearest integer. 95% CI, 95% confidence interval.

^aMean (range) of predicted probability of the primary outcome for scores contained within respective strata.

^bRange of absolute risk of the outcome within each stratum across all four cohorts (excluding living kidney donor analysis).

^cRelative risk was calculated by comparing risk strata using a univariable log-binomial regression in the derivation cohort (with reference to the “low-risk” stratum).

^deGFR in ml/min per 1.73 m².

Table 3.

Performance measures for the clinical score model

Model Performancea	Derivation Cohort	Validation Cohort 1	Validation Cohort 2	Validation Cohort 3	Validation Cohort 4	Living Kidney Donors
Harrell c statistic	0.84	0.84	0.86	0.86	0.88	0.83
95% CI	0.79 to 0.89	0.78 to 0.89	0.79 to 0.93	0.81 to 0.92	0.81 to 0.96	0.79 to 0.87
Calibration testb	0.74	0.15	0.40	0.14	0.53	<0.001
Brier scorec	0.11	0.11	0.09	0.10	0.07	0.05
P valued	0.06	0.06	0.10	0.08	0.58	0.99

Logistic coefficients are from model 3 (Table 1). To derive the points system, each coefficient was divided by the lowest value (0.66) and rounded to the nearest integer. 95% CI, 95% confidence interval.

^aInformation represents univariable logistic regression models that considered the clinical score as a continuous variable; refer to Figure 1 for further evaluation of model performance and fit.

^bP value reported; values of 0.05 reflect a well calibrated model.

^cValues of 0.00 reflect perfect forecasting; values of 0.25 reflect a noninformative model.

^dP value from Spiegelhalter z statistic; values of 0.05 reflect a well fitted model.

The c statistic for the clinical score model in the derivation cohort was 0.84. There was a high degree of concordance between the observed cases and predicted probability across the validation cohorts (Figure 1). Formal hypothesis testing demonstrated adequate calibration across validation cohorts, with P values consistently >0.05 (Supplemental Figure 2). Model performance was not diminished in sensitivity analyses (Supplemental Figure 3). Calibration was diminished in the living kidney donor cohort when predicted probability was low (Supplemental Figure 4), but discrimination was not affected (c=0.83). Living kidney donor characteristics are presented in Supplemental Table 4.

Figure 1.

Discrimination and calibration of the proposed clinical score was reproducible across multiple cohorts. Receiver operating characteristics curves and graphs of predicted probability (95% confidence interval) and observed absolute risk of the primary outcome across the study cohorts. (A) Derivation cohort: c=0.84. (B) Validation cohort 1: c=0.84. (C) Validation cohort 2: c=0.86. (D) Validation cohort 3: c=0.86. (E) Validation cohort 4: c=0.88.

Postoperative kidney function compared by risk stratum is presented in Table 4. Across all cohorts, observed absolute risks of a postoperative eGFR <45 ml/min per 1.73 m² in the moderate- and high-risk groups were between 21% and 26% and between 46% and 69%, respectively. When considering all patients with kidney cancer, a score of zero to three (negligible risk) had a negative predictive value of 98.9% (95% confidence interval, 97.7% to 99.4%) for clinically significant CKD.

Table 4.

Postoperative kidney function according to assigned risk level

Risk Stratum (Score)	n (%)	eGFR Median [IQR]	eGFR=45–60, n (%)	eGFR=30–44, n (%)	eGFR<30, n (%)
Derivation cohort
Negligible (0–3)	278 (40)	76 [66–93]	42 (15)	1 (<1)	0 (0)
Low (4–6)	167 (24)	64 [55–75]	56 (34)	13 (8)	0 (0)
Moderate (7–8)	189 (27)	52 [46–59]	104 (55)	38 (20)	2 (1)
High (9–10)	65 (9)	45 [39–51]	31 (48)	29 (45)	1 (2)
Validation cohort 1
Negligible (0–3)	148 (35)	80 [65–93]	23 (16)	3 (2)	1 (<1)a
Low (4–6)	111 (26)	62 [54–74]	38 (43)	10 (9)	0 (0)
Moderate (7–8)	115 (27)	52 [45–60]	59 (51)	22 (19)	5 (4)
High (9–10)	49 (12)	42 [38–47]	16 (33)	27 (55)	3 (6)
Validation cohort 2
Negligible (0–3)	85 (47)	79 [64–90]	11 (13)	1 (1)	0 (0)
Low (4–6)	35 (20)	61 [55–73]	10 (29)	5 (14)	0 (0)
Moderate (7–8)	42 (23)	53 [45–63]	18 (43)	10 (24)	0 (0)
High (9–10)	17 (9)	48 [41–63]	3 (18)	7 (41)	1 (6)
Validation cohort 3
Negligible (0–3)	79 (39)	83 [70–102]	8 (10)	0 (0)	0 (0)
Low (4–6)	54 (26)	59 [53–69]	24 (44)	7 (13)	0 (0)
Moderate (7–8)	53 (26)	52 [45–57]	31 (58)	14 (26)	0 (0)
High (9–10)	19 (9)	41 [37–46]	6 (32)	12 (63)	1 (5)
Validation cohort 4
Negligible (0–3)	103 (47)	87 [77–92]	3 (3)	2 (2)	0 (0)
Low (4–6)	59 (27)	65 [58–79]	16 (27)	2 (3)	0 (0)
Moderate (7–8)	46 (21)	61 [50–71]	13 (28)	8 (17)	1 (2)
High (9–10)	13 (6)	42 [37–46]	2 (15)	9 (69)	0 (0)
Living kidney donors
Negligible (0–3)	998 (53)	68 [60–76]	230 (23)	11 (1)	1 (<1)
Low (4–6)	417 (22)	58 [53–65]	215 (52)	17 (4)	0 (0)
Moderate (7–8)	419 (22)	52 [47–57]	278 (66)	64 (15)	0 (0)
High (9–10)	32 (2)	44 [41–49]	8 (25)	21 (66)	0 (0)

eGFR is in milliliters per minute per 1.73 m².

^aThis outlying value is most likely the consequence of the patient in question being predisposed to developing kidney insufficiency for a reason that would have warranted exclusion from this study but was not recorded in the dataset (e.g., kidney transplant recipient or anatomically/functionally solitary kidney).

Discussion

The aim of this study was to develop and validate a simple clinical score that could reliably stratify risk of CKD in patients with normal kidney function who were undergoing nephrectomy. We developed a ten-point clinical score using three patient factors (age ≥65 years old, diabetes mellitus, and preoperative eGFR) and the proposed management approach (radical/partial nephrectomy). These variables have previously been shown to be strongly associated with risk of CKD after nephrectomy.^11,17 The score reliably categorized risk of clinically significant CKD.

External validation of the clinical score in geographically and temporally distinct cohorts demonstrated good discriminatory performance and calibration; c statistics were consistently ≥0.84, with high concordance between predicted risk and observed cases across all validation cohorts. The model also had good discrimination and calibration in the cohort of living donors, although risk was overestimated at the lower bound of the score (because donors could not score less than three points). This exploratory analysis supports the robustness of the proposed model.

We propose three potential applications for this score. The first is in terms of primary prevention of postoperative CKD. Because the proposed management strategy is included in the model, there is potential to assess the CKD risk under different management approaches. For example, the same patient could have a score of six (low risk) or nine (high risk) depending on whether they were undergoing partial or radical nephrectomy, a distinction that is potentially associated with ≥30% difference in absolute risk of significant CKD. Depending on technical feasibility and urologic/oncologic risk, partial rather than radical nephrectomy may be considered favorable in this scenario. Although bias, particularly confounding by indication, may have affected our estimate of absolute risk reduction associated with partial nephrectomy, this is unlikely to be a major consideration in practice because the strength of the causal association between nephrectomy type and postoperative kidney function has been clearly demonstrated by a randomized, controlled trial.¹⁸

International guidelines recommend partial nephrectomy as standard of care when technically feasible^2,10; thus, it could be argued that partial nephrectomy would most likely be performed regardless of preoperatively determined CKD risk. Partial nephrectomy is widely used at higher-volume centers,^19,20 but this may not be the case at lower-volume/regional centers, where patients have a higher likelihood of radical nephrectomy, even for small renal masses.^11,21,22 This preoperative scoring system could facilitate appropriate referral to tertiary hospitals to minimize risk of postoperative CKD. Conversely, the score could also assist in selecting patients for whom management with radical nephrectomy at a smaller center would have low probability of leading to significant CKD. This is particularly relevant given the high negative predictive value of the negligible risk category.

The clinical score did not perform as well for patients with advanced disease compared with stage T1 tumors. This likely reflects the heterogeneous clinical picture associated with advanced kidney cancer. In particular, patients with larger tumors were substantially more likely to undergo radical nephrectomy, more likely to have preoperative CKD,²³ and less likely to experience the same preservation of kidney function compared with patients who had smaller tumors (if undergoing partial nephrectomy).¹¹ There may be value in augmenting the scoring system by estimating the possible volume of parenchymal preservation with partial nephrectomy.²⁴

Another application of this is for guiding postoperative follow-up. Current guidelines for postnephrectomy follow-up are on the basis of oncologic risk^2,25; therefore, patients at higher risk of significant CKD but with low oncologic risk may be more likely to be lost to follow-up compared with patients with high oncologic risk.

Finally, the tool could be useful for counseling patients about postnephrectomy CKD risk. Providing accurate information about surgical risks is important for informed consent, and this score may facilitate discussions with patients. Figure 2 shows a simplified version of this score for clinical applications. A caveat is that this score does not predict lifetime risk of stage ≥3b CKD, only likelihood at 12 postoperative months. Longitudinal exploration of the utility of the clinical score for this purpose would be desirable.

Figure 2.

A simplified version of the clinical score for predicting clinically significant CKD after nephrectomy that is suitable for clinical applications. Points are cumulative to a maximum of 10, and associated with a level of risk of clinically significant CKD one year postnephrectomy, expressed as a percentage.

This clinical score was limited by the fact that it was developed/validated using observational data, which introduces the risk of selection bias affecting the magnitude of estimated effect sizes. For example, older patients were more likely to undergo partial nephrectomy in Victoria compared with Queensland,²⁶ which might reasonably be expected to attenuate the size of the protective effect of partial nephrectomy, and this may explain some differences in calibration between Queensland and Victoria. Furthermore, additional information on postoperative conditions that could have contributed to deterioration in kidney function may further inform clinical practice. Accordingly, exploration of this score with a pragmatic clinical trial is warranted.

Albuminuria was not included in our model because these data are not routinely recorded in this clinical context. Although some studies reported an association between albuminuria and postoperative CKD,²⁷ the size/aggressiveness of the tumor also affects albuminuria,²⁸ which may reduce its sensitivity for glomerular damage. Because albuminuria may predict CKD risk,^8,29 it would be reasonable to evaluate albuminuria as an adjunct to this clinical score. Similarly, hypertension was not included because diagnoses were not reliably documented/standardized, which would have introduced higher risk of differential misclassification. Care should be taken when interpreting this clinical score in patients with poorly controlled hypertension.

Patients with an eGFR<60 ml/min per 1.73 m² were omitted from this study because the intention was to focus on patients with “normal” preoperative kidney function. Patients with CKD before surgery are at high risk of worsening kidney function and mortality³; therefore, they should always be considered at heightened risk of significant postoperative CKD and associated sequelae. We developed and validated a clinical tool, which could reproducibly stratify postnephrectomy CKD risk on the basis of readily available parameters. This model may help guide management decisions and follow-up periods on the basis of risk of developing clinically significant CKD after nephrectomy.

Disclosures

Dr. Donaldson reports personal fees from the European Association of Urology, outside the submitted work. Prof. Francis reports funding from Amgen and Novartis, outside the submitted work. Prof. Stewart reports grants and personal fees from Pfizer, grants from AstraZeneca, personal fees from CMR Surgical, personal fees from Merck, and personal fees from EUSA Pharma, outside the submitted work. Dr. Giles reports grants from National Health & Medical Research Council, during the conduct of the study. Dr. Hawley reports grants from Shire, grants from Fresenius, grants from Baxter Healthcare, other from Janssen, other from GlaxoSmithKline, personal fees from Otsuka, grants from Otsuka, outside the submitted work. All remaining authors have nothing to disclose.

Funding

Prof. Davis was supported by a National Health and Medical Research Council fellowship APP1102604. Dr. Ellis was supported by an Australian Government Research training stipend. Prof. Jordan was supported by a National Health and Medical Research Council fellowship APP1061341. Prof. Neale was supported by a National Health and Medical Research Council fellowship APP1043029. Australian population–based data collection was funded through Victorian Cancer Agency translational research grant EOI09_E36 and a Cancer Council Queensland project grant.

Supplemental Material

This article contains the following supplemental material online at http://jasn.asnjournals.org/lookup/suppl/doi:10.1681/ASN.2019121328/-/DCSupplemental.

Supplemental Figure 1. Receiver operating characteristics curves for models 1–3 (Table 1).

Supplemental Figure 2. Calibration belt plots for the clinical score model in the derivation and validation cohorts showing calibration at various confidence levels.

Supplemental Figure 3. Receiver operating characteristics curves for the clinical score model in sensitivity analyses.

Supplemental Figure 4. Discrimination and calibration of the clinical score model in living kidney donors.

Supplemental Table 1. Numbers of participants excluded according to prespecified criteria.

Supplemental Table 2. Patient and tumor characteristics.

Supplemental Table 3. Logistic regression analysis considering risk strata as the independent variable.

Supplemental Table 4. Characteristics of living kidney donors.

Supplemental Table 5. Risk prediction models in the derivation cohort with odds ratios reported.