Accounting for the Competing Risk of Death to Predict Kidney Failure in Adults With Stage 4 Chronic Kidney Disease

Abstract

Importance: Kidney failure risk prediction has implications for disease management, including advance care planning in adults with severe (ie, estimated glomerular filtration rate [eGFR] category 4, [G4]) chronic kidney disease (G4-CKD). Existing prediction tools do not account for the competing risk of death.

Objective: To compare predictions of kidney failure (defined as estimated glomerular filtration rate [eGFR] <10 mL/min/1.73 m2 or initiation of kidney replacement therapy) from models that do and do not account for the competing risk of death in adults with G4-CKD.

Design, setting, and participants: This prognostic study linked population-based laboratory and administrative data (2002-2017) from 2 Canadian provinces (Alberta and Manitoba) to compare 3 kidney risk models: the standard Cox regression, cause-specific Cox regression, and Fine-Gray subdistribution hazard model. Participants were adults with incident G4-CKD (eGFR 15-29 mL/min/1.73 m2). Data analysis occurred between July and December 2020.

Main outcomes and measures: The performance of kidney risk models at prespecified times and across categories of baseline characteristics, using calibration, reclassification, and discrimination (for competing risks). Predictive characteristics were age, sex, albuminuria, eGFR, diabetes, and cardiovascular disease.

Results: The development and validation cohorts included 14 619 (7070 [48.4%] men; mean [SD] age, 74.1 [12.8] years) and 2295 (1152 [50.2] men; mean [SD] age, 71.9 [14.0] years) adults, respectively. The 3 models had comparable calibration up to 2 years from entry. Beyond 2 years, the standard Cox regression overestimated the risk of kidney failure. At 4 years, for example, risks predicted from standard Cox were 40% for people whose observed risks were less than 30%. At 2 years (risk cutoffs 10%-20%) and 5 years (risk cutoffs 15%-30%), 788 (5.4%) and 2162 (14.8%) people in the development cohort were correctly reclassified into lower- or higher-risk categories by the Fine-Gray model and incorrectly reclassified by standard Cox regression (the opposite was observed in 272 patients [1.9%] and 0 patients, respectively). In the validation cohort, 115 (5.0%) individuals and 389 (16.9%) individuals at 2 and 5 years, respectively, were correctly reclassified into lower- or higher-risk categories by the Fine-Gray model and incorrectly reclassified by the standard Cox regression; the opposite was observed in 98 (4.3%) individuals and 0 individuals, respectively. Differences in discrimination emerged at 4 to 5 years in the development cohort and at 1 to 2 years in the validation cohort (0.85 vs 0.86 and 0.78 vs 0.8, respectively). Performance differences were minimal during the entire follow-up in people at lower risk of death (ie, aged ≤65 years or without cardiovascular disease or diabetes) and greater in those with a higher risk of death. At 5 years, for example, in people aged 65 years or older, predicted risks from standard Cox were 50% where observed risks were less than 30%. Similar miscalibration was observed at 5 years in people with albuminuria greater than 30 mg/mmol, diabetes, or cardiovascular disease.

Conclusions and relevance: In this study, predictions about the risk of kidney failure were minimally affected by consideration of competing risks during the first 2 years after developing G4-CKD. However, traditional methods increasingly overestimated the risk of kidney failure with longer follow-up time, especially among older patients and those with more comorbidity.

Figure 1.. Cumulative Incidence Functions vs Kaplan-Meier Failure Functions

Dashed lines indicate Kaplan-Meier estimator functions; solid lines, cumulative incidence functions; G4, severe CKD. All curves are stacked. The sum of the 2 Kaplan-Meier failure functions corresponding to the 2 competing events (kidney failure and death) is more than 1 before year 8 in the development cohort (A) and year 6 in the validation cohort (B).

Figure 2.. Calibration Plots at 2, 4, 6, and 8 Years in the Validation Cohort

Calibration plots summarize the graphical agreement between observed and predicted risks at years 1 through 8. In an ideal model, pairs of the observed and predicted risks lie on a 45-degree angle line. Curves falling under the 45-degree angle line indicate that predicted risks overestimate (are higher than) observed risks. Corresponding plots at years 1 through 8 for the development cohort and for years 1, 3, 5, and 7 for the validation cohort are provided in the eFigures 2-4 in the Supplement.

Figure 3.. Predicted vs Observed Risk of Kidney Failure at Years 2 and 5 in the Development and Validation Cohorts

Risk (%) was predicted for each member of the development (A, C) and validation (B, D) cohort according to the Fine and Gray subdistribution hazard model and the standard Cox model at years 2 and 5 from study entry. People were then assigned to each cell of a 3 × 3 table corresponding to the combination of the model predictions. Each cell of the 3 × 3 table includes the number of people (top) and their actual observed risk (crude cumulative incidence function) at 2 or 5 years (bottom, bold). FG-/SC+, total No. (%) of people incorrectly classified by Fine-Gray model and correctly classified by standard Cox regression with respect to the actual observed risk; FG+/SC-, total No. (%) of people correctly classified by the Fine-Gray model and incorrectly classified by standard Cox regression with respect to the actual observed risk; NA indicates not available.

Figure 4.. C Statistics and Brier Score in the Development and Validation Cohort

The C statistic in the development cohort (A) and validation cohort (B) assesses the ability of separating people with kidney failure from those without kidney failure, ranging from 0.5 (no prediction ability beyond chance) to 1 (perfect discrimination). Smaller Brier scores in the development cohort (C) and validation cohort (D) indicate better performance in terms of both discrimination and calibration.