Prediction of Kidney Failure in Children with Chronic Kidney Disease and Obstructive Uropathy

Abstract

Background:

Obstructive Uropathy (OU) is a leading cause of pediatric kidney injury. Accurate prediction of kidney disease progression may improve clinical outcomes. We aimed to examine discrimination and accuracy of a validated kidney failure risk equation (KFRE), previously developed in adults, in children with OU.

Methods:

We identified 118 children with OU and an estimated glomerular filtration rate (eGFR)<60 ml/min/1.73m² in the Chronic Kidney Disease in Children study, a national, longitudinal, observational cohort. Each patient’s 5-year risk of kidney failure was estimated using baseline data and published parameters for the 4- and 8-variable KFREs. Discriminative ability of the KFRE was estimated using the C statistic for time-to-event analysis. Sensitivity and specificity were evaluated across varying risk thresholds.

Results:

Among the 118 children, 100 (85%) were boys, with median baseline age of 10 years (interquartile range: 6–14). Median eGFR was 42 mL/min/1.73m² (32–53), with median follow-up duration of 4.5 years (2.7–7.2); 23 patients (19.5%) developed kidney failure within 5 years. The 4-variable KFRE discriminated kidney failure risk with a C statistic of 0.75 (95% CI: 0.68–0.82). A 4-variable risk threshold of ≥30% yielded 82.6% sensitivity and 75.0% specificity. Results were similar using the 8-variable KFRE.

Discussion:

In children with OU, the KFRE discriminated the 5-year risk of kidney failure at C statistic values lower than previously published in adults but comparable to suboptimal values reported in the overall CKiD population. The 8-variable equation did not improve model discrimination or accuracy, suggesting the need for continued research into additional, disease specific markers.

Introduction

Chronic kidney disease (CKD) is characterized by gradual loss of kidney function over time. Although it is extremely rare in children, CKD is associated with significant health complications, impaired growth, and shorter life expectancy [1,2]. Owing to the progressive nature of CKD, there is a strong emphasis on research to improve prognostic estimation of kidney failure. The kidney failure risk equation (KFRE) was developed to help guide clinical care and decision-making for adults with CKD [3]. Since its development and external validation in adults, the KFRE has also shown high discriminative ability in estimating the 1- and 2-year kidney failure risk in a heterogeneous group of children with CKD resulting from both glomerular and non-glomerular etiologies. However, in this same population the 5-year risk discrimination was less optimal [4]. We recognize that the discrimination and accuracy of such an equation may differ based on disease etiology and that optimizing kidney failure risk prediction could benefit from disease specific evaluation.

Obstructive uropathy (OU) is a leading cause of non-glomerular CKD, accounting for 21% of all CKD cases and 13–17% of kidney failure cases among North American children [1]. OU encompasses a constellation of diagnoses affecting the urinary tract, such as posterior urethral valves (PUV), which is characterized by the impedance of urine flow at the level of the posterior urethra, resulting in kidney injury. Despite advances in antenatal and early-life interventions, OU patients continue to experience a significant risk of developing kidney failure before adulthood [5–7]. Accurate prediction of disease progression is especially important for CKD patients with OU to inform individualized surveillance, given that progression to kidney failure can occur at any point during childhood or even adulthood [8]. The objective of this study was to examine the discrimination and accuracy of the 4- and 8-variable KFRE to predict the risk of kidney failure in children with CKD due to OU and to assess the need for future work focusing on model improvement.

Methods

Study population and design

The Chronic Kidney Disease in Children (CKiD) study is a prospective, observational cohort of children with mild-to-moderate CKD recruited from 48 North American Pediatric Nephrology Centers. Enrollment for the initial cohort (cohort 1) started in 2005 and a second cohort was added in 2011 (cohort 2). A total of 891 children were enrolled between cohorts 1 and 2, with baseline visits occurring through 2013 [9]. Details regarding the study design and methods have been previously published [10]. Briefly, all participants underwent baseline evaluation and yearly follow-up visits to measure variables related to kidney function, cardiovascular disease, growth, cognition, and behavior [10]. The original CKiD study protocol was reviewed and approved by the Institutional Review Boards (IRBs) at each participating center. For the present analysis, we initially selected CKiD participants who had a primary diagnosis of OU. After identifying children with OU, we further restricted the study population to children with an eGFR<60 mL/min/1.73 m² at baseline because the KFRE was developed and validated in adults with an eGFR below this level [3]. Baseline was defined as the first study visit or the first visit at which eGFR fell below 60 mL/min/1.73 m². The final study cohort further excluded children who did not have complete baseline data for at least the 4 variables used in the 4-variable KFRE (age, sex, eGFR, and albumin-to-creatinine-ratio [ACR]). This specific study was exempt from IRB review owing to its use of only de-identified data.

Variables

The study outcome of interest was progression to kidney failure within 5 years of baseline. Kidney failure was defined as receipt of long-term dialysis or kidney transplant (whichever came first). The primary predictive factor was the calculated risk (probability) of kidney failure within 5 years from baseline. The KFRE is calculated based on a Cox proportional hazards regression model that uses patient age, sex, and urinary and serum measurements. The regression model parameters from the published KFRE for the 5-year risk of kidney failure were used in this study [11]. The discriminative ability, calibration, sensitivity, specificity, and negative and positive predictive values of the 4-variable (age, sex, eGFR, ACR) and the 8-variable KFRE (4 variables plus serum calcium, phosphate, bicarbonate, and albumin levels) were examined separately. The variables used to calculate the KFRE were measured at the baseline visit or the visit when eGFR first decreased to <60 mL/min/1.73 m². Albuminuria was not measured at all baseline visits as this measurement was added after the CKiD study began. Therefore, for patients who were missing ACR at baseline, the ratio of protein-to-creatinine-levels (PCR) was transformed to an ACR using a published equation (girls: PCR/2.655; boys: PCR/1.7566) [12]. To examine the adequacy of this conversion, we estimated the correlation between the measured ACR and the ACR calculated from PCR (Pearson r=0.934; p<.0001). Race (white, black, other/multiracial) and Hispanic ethnicity (yes/no) were additional variables included in descriptive analyses to characterize the study cohort.

Statistical analysis

Frequencies and summary measures were used to describe patient characteristics and baseline values for the 8 variables used to calculate the KFRE. Each patient’s predicted risk of progressing to kidney failure within 5 years was calculated using the published proportional hazards regression model intercept and coefficient (slope) parameters for the 4-variable and 8-variable KFRE equations [11]. The calculated 5-year KFRE risk scores for each patient were then used to determine the KFRE model discrimination, calibration, sensitivity, specificity and predictive values as described below. Separate analyses were conducted using the 4-variable and 8-variable KFRE risk scores.

Discrimination:

Model discrimination, the ability to differentiate between patients who progressed to kidney failure and those who did not, was determined using concordance statistics for times-to-event analysis (C statistic) [13,14].

Calibration:

Calibration is a measure of how close a predicted risk estimate from a model is to the actual proportion of patients that experience an outcome in the study population. To determine calibration, the study cohort was divided into two groups based on the median KFRE risk score (above vs. below median). Kaplan-Meier curves for times-to-event were then fit to each group to compare the median KFRE risk score in each group to the observed Kaplan-Meier estimate of the proportion who progressed to kidney failure within 5 years. Miscalibration was assessed using the Greenwood-Nam-D’Agostino test [15].

Sensitivity, specificity, positive and negative predictive values:

to evaluate the clinical usefulness of the KFRE, several different thresholds for the estimated 5-year KFRE risk score were specified to classify patients as likely or unlikely to progress to kidney failure within 5 years. At each threshold, patients with a risk score above that value were classified as likely to progress to kidney failure within 5 years and the resulting classification was compared to the actual number of patients who did or did not progress to kidney failure within 5 years of study follow-up. This analysis excluded patients who were lost to follow-up before 5 years and whom did not reach kidney failure during their available follow-up. Sensitivity was defined as the proportion of true-positives among patients who progressed to kidney failure by 5 years, and specificity was defined as the proportion of true-negatives among patients who did not reach kidney failure by 5 years. Positive and negative predictive values (PPV and NPV) were also estimated as the proportion of patients who progressed to kidney failure among those classified as likely to progress to kidney failure, and the proportion of patients who did not progress to kidney failure among those classified as unlikely to progress to kidney failure, respectively. Exact confidence intervals were computed for each estimate of sensitivity, specificity, positive and negative predictive values.

Results

A total of 149 children in the CKiD study had a primary diagnosis of OU. Of the initial 149 patients, 25 patients were excluded because they did not have eGFR<60 mL/min/1.73 m² at either their first or any subsequent study visit, and an additional 6 patients were excluded because they did not have either ACR or PCR data, resulting in a final study cohort of 118 patients (Figure 1). Demographic characteristics and baseline physical examination and laboratory data are shown in Table 1. The median age at study entry was 10 years (interquartile range [IQR]: 6–14). Just under 85% of patients were boys; 59% were white, and 11% identified as having Hispanic ethnicity. The median baseline eGFR was 42 mL/min/1.73 m² (IQR: 32–53). During the follow-up period, 1 patient progressed to kidney failure by 2 years, 23 patients (19.5%) by 5 years, and 36 (30.5%) by the end of follow-up. The median follow-up time for the cohort was 4.5 years (2.7–7.2). Kaplan-Meier survival estimates for the risk of kidney failure over time are displayed and summarized in the supplemental Figure S1 and Table S1, respectively.

Figure 1.

Obstructive uropathy study cohort selection

CKiD, Chronic Kidney Disease in Children Study; CKD, Chronic Kidney Disease; eGFR, estimated glomerular filtration rate; Microalbumn/CR, ratio of urine microalbumin to creatinine

Table 1.

Baseline patient characteristics and outcomes in the CKiD obstructive uropathy cohort (n=118).

Demographic		N (%) or median (IQR)
Age, y		10 (6–14)
Male sex		100 (84.8)
Race	White	70 (59.3)
	Black	32 (27.1)
	Multiracial/other	16 (13.6)
Hispanic		13 (11.3)
Laboratory
eGFR, mL/min/1.73m2		42 (32–53)
Urine protein/creatinine ratio, mg/g		358 (200–844)
Calculated urine albumin/creatinine ratio, mg/g		139 (77–353)
Serum calcium (mg/dL)		9.7 (9.3–10)
Serum phosphate (mg/dL)		4.6 (4.2–5)
Serum albumin (g/dL)		4.3 (4.2–4.5)
Serum bicarbonate, mEq/L (n=96)		23 (21–24.0)
Outcomes
Follow-up time from baseline visit, y		4.5 (2.7–7.2)
Progression to kidney failure within 5 y from baseline visit		23 (19.5)
Total patients who progressed to kidney failure		36 (30.5)

CKiD, Chronic Kidney Disease in Children Study; IQR, interquartile range; SD, standard deviation; eGFR, estimated glomerular filtration rate; Baseline: first visit with the GFR<60; For 13 patients, baseline visit was after the first study visit

With regard to model discrimination, a C statistic value of 0.5 suggests random concordance, and a value of 1 suggests perfect concordance. In the OU cohort the KFRE provided a 5-year discrimination with similar C statistic values for the 4-variable (C: 0.75; 95% confidence interval [CI]: 0.68–0.82) and 8-variable equation (C: 0.78; 95% CI: 0.70–0.86; restricted to 94 patients with complete data for the 8 variables). The median estimated risk scores based on the 4- and 8-variable KFREs were 14% and 15%, respectively. Results of our analysis of model calibration are shown in Figure 2. The estimated proportion of participants expected to develop kidney failure was close to the observed proportion who progressed to kidney failure within 5 years, with some underestimation among patients with a KFRE score below the median, and overestimation among patients with a KFRE score above the median.

Figure 2.

Observed versus predicted 5-year risk of kidney failure using the 4-variable (top) and 8-variable (bottom) KFRE in the CKiD obstructive uropathy cohort

Observed: Kaplan-Meier estimate of the proportion who developed kidney failure by 5 years; Predicted: Median estimated kidney failure risk equation scores in each risk group

Measures of the accuracy of the 4- and 8-variable KFRE risk scores for classifying patients as likely or unlikely to progress to kidney failure by 5 years are shown in Tables 2 and S2, respectively. The sensitivity, specificity, and predictive values associated with different thresholds for the KFRE risk scores were estimated. Using the 4-variable KFRE, a threshold of 30% was the only threshold that yielded sensitivity and specificity values that were both greater than 70%. Classifying patients as likely to progress to kidney failure within 5 years if they had a risk score of 30% or greater resulted in the correct identification of 19 out of 23 patients who progressed to kidney failure within 5 years, for 82.6% sensitivity (95% CI: 61.2–95.1). This threshold also correctly identified 39 of 52 patients who did not progress to kidney failure within 5 years, for 75.0% specificity (95% CI: 61.0–86.0) (Table 2). In general, PPVs were <60% and NPVs were >70% across most thresholds (Table 2). Similar to the results for the 4-variable KFRE, a 30% risk threshold based on the 8-variable equation yielded 84.2% sensitivity (95% CI: 60.4–96.6) and 68.4% specificity (95% CI: 51.3–82.5) (Table S2).

Table 2.

Sensitivity, specificity, and predictive values resulting from the application of different thresholds of 4-variable KFRE scores to the CKiD obstructive uropathy cohort.

Predicted risk thresholda		Progressed to kidney failure by 5 years (n=23)	Did not progress to kidney failure by 5 years (n=52)	Sensitivity, % (95% CI)	Specificity, % (95% CI)	PPV, % (95% CI)	NPV, % (95% CI)
9%	At/above	23	27	100 (85.0–100)	48.1 (34.0–62.4)	46.0 (31.8–60.7)	100 (86.2–100)
	Below	0	25
20%	At/above	19	19	82.6 (61.2–95.1)	63.5 (48.9–76.4)	50.0 (33.4–66.6)	89.2 (74.6–97.0)
	Below	4	33
30%	At/above	19	13	82.6 (61.2–95.1)	75.0 (61.0–86.0)	59.4 (40.6–76.3)	90.7 (77.9–97.4)
	Below	4	39
40%	At/above	15	12	65.2 (42.7–83.6)	76.9 (63.2–87.5)	55.6 (35.3–74.5)	83.3 (69.8–92.5)
	Below	8	40
50%	At/above	9	10	39.1 (19.7–61.5)	80.8 (67.5–90.4)	47.4 (24.4–71.1)	75.0 (61.6–85.6)
	Below	14	42
60%	At/above	6	7	26.1 (10.2–48.4)	86.5 (74.2–94.4)	46.1 (19.2–74.9)	72.6 (59.8–83.2)
	Below	17	45
70%	At/above	5	6	21.7 (7.5–43.7)	88.5 (76.6–94.8)	45.5 (16.8–76.6)	71.9 (59.2–82.4)
	Below	18	46
80%	At/above	4	4	17.4 (5.0–38.8)	92.3 (81.5–97.9)	50.0 (15.7–84.3)	71.6 (59.3–82.0)
	Below	19	48
90%	At/above	3	1	13.0 (2.8–33.6)	98.1 (89.7–99.9)	75.0 (19.4–99.4)	71.8 (59.9–81.9)
	Below	20	51

CKiD, Chronic Kidney Disease in Children study; ; KFRE, kidney failure risk equation; TPR, true positive rate; TNR, true negative rate; PPV, positive predictive value; NPV, negative predictive value

^aKFRE score based on the baseline values of 4 variables: Age, sex, eGFR, urine albumin/creatinine ratio 43 of the total 118 patients were excluded because they were lost to follow up before 5 years

Discussion

In this analysis of children with CKD and OU enrolled in the CKiD study, we identified similar discriminative ability of the 4- and 8-varible KFRE to predict the 5-year risk of progression to kidney failure compared to the overall CKiD population. Although our C statistic point estimate of discrimination in the 4–variable model was slightly lower compared to the previously reported overall population (0.75 vs. 0.81 respectively), confidence intervals were overlapping signifying no measurable difference [4]. Both the C statistic values for the 5-year risk discrimination noted in our study and reported for the entire cohort were lower than those reported in adult populations, signifying potential room for improvement in its use in children [11,3]. In addition, classifying patients in our study as likely to develop kidney failure if their age-, sex-, eGFR-, and microalbuminuria-based risk was ≥30% provided 82.6% sensitivity and 75.0% specificity to identify patients who progressed to kidney failure within 5 years. Results were similar when the 8-variable risk equation was applied to a subset of the cohort with complete data. The similar results between the 4- and 8-variable KFRE equations for discrimination and accuracy suggest that a search for alternative additional markers may help further improve risk prediction for OU patients.

Since initial development of the KFRE in adult patients in Canada [3], the equation has been internationally validated across several cohorts of adult patients [11]. Beyond the KFRE, several other risk equations have been reported for adults [16], and a risk classification system analogous to the Kidney Disease: Improving Global Outcomes classification system has been recently developed for children [17]. Although the importance of risk equation models to predict kidney failure is not in doubt, the ability of current models to accurately predict 5-year risk in children is. In 2018, Winnicki et al. examined the performance of the KFRE in the overall CKiD cohort of children with GFR<60 ml/min/1.73m², which was comprised of children with glomerular and non-glomerular disease [4]. In the overall CKiD cohort, the 4-variable KFRE discriminated 1-year risk of kidney failure with C statistics of 0.90 (95% CI: 0.86–0.93), however, this decreased to 0.81 (95% CI: 0.77–0.83) for 5-year risk. Discrimination was similar in the 8-variable KFRE model [4]. In the present study of CKiD patients with a primary diagnosis of OU, the discriminative ability of the KFRE could only be evaluated for the 5-year risk of kidney failure because only one patient experienced the outcome by 2 years. The observed C statistic of 0.75 (95% CI: 0.68–0.82) suggests that the 4-variable KFRE is expected to correctly differentiate between children with OU who will and will not progress to kidney failure within 5 years (i.e., assign a higher risk to the former versus the latter) 75% (95% CI: 68%-82%) of the time. While the discriminative ability of the KFRE in the OU cohort was comparable to the overall CKiD cohort, the observed C statistic suggests that the KFRE did not perform as well in the study cohort as it did in adult CKD patients (overall C statistic, 0.88 [95% CI: 0.86–0.90] for 5-year kidney failure risk across 30 countries) [11]. The value of the C statistic depends on the model being used and distribution of the model predictors in the sample upon which it is applied [18]. It is possible that the lower discriminative ability of the KFRE was caused by differences in the distribution and influence of KFRE variables on the risk of kidney failure for OU patients compared to adult CKD patients. For example, females comprised 33% of adult patients across the international validation cohorts but just over 15% of the CKiD OU cohort [11]. Lower heterogeneity in KFRE variables such as sex in the CKiD OU cohort may have contributed to a lower C statistic compared to adult CKD patients [18].

The current study contributes to the kidney failure risk prediction literature by providing the first evaluation of classification thresholds to inform the usefulness of the KFRE risk scores in predicting kidney failure in children with CKD and OU. The use of a 30% risk threshold to classify OU patients with CKD as high risk for progression within 5 years based on 4 variables provided 82.6% sensitivity (95% CI: 61.2–95.1) and 75.0% specificity (95% CI: 61.0–86.0). Sensitivity values were <70% at higher thresholds, and specificity values were <70% at lower thresholds. PPVs were less than 60% for most risk thresholds and NPVs were above 70% across all thresholds of the calculated risk in the OU cohort. From a patient perspective, the results for PPV and NPV suggest that patients with CKD due to OU who are classified as unlikely to progress to kidney failure within 5 years can be more certain that their predicted outcome is correct than those who are classified as likely to progress. The PPV of a screening test decreases and the NPV increases as the prevalence of an outcome decreases in a population [19]. Therefore, the low PPV and high NPV values identified in the OU cohort are not surprising given that children with non-glomerular CKD experience slower disease progression and lower incidence of kidney failure compared to those with glomerular causes [17,20]. Interestingly, in the CKiD OU cohort, the 8-variable equation did not appear to significantly improve any predictive measures we studied. This was similar to findings reported by the KFRE study using the entire CKiD population [4]. This is in contrast however, to the initial study of adult Canadians where the KFRE was first validated. In this population the 8-variable equation outperformed the 4-variable with integrated discrimination improvement of 3.2% (95% CI, 2.4%-4.2%) [3]. This suggests, that in children with CKD, a search for alternative markers, possibly those more disease specific may help improve the current 4-variable KFRE model.

This study has limitations that should be considered. Because the specific condition leading to the diagnosis of OU is not available in the CKiD database, we are unable to evaluate differences in the KFRE performance for PUV patients compared to those with obstructive kidney injury from other sources. This is salient, as there are significant differences in kidney injury mechanism in PUV compared to upper tract obstruction where bladder level dysfunction is not typical. This lack of diagnosis granularity and patient specific clinical parameters like bladder dysfunction, however, unlikely undermines our findings of poor KFRE performance in OU. In fact, this supports our conclusions that diagnosis specific risk equations are necessary to optimize these equations since the addition of general markers of kidney injury (calcium, phosphate, bicarbonate, and albumin) did not improve model performance.

The current study was restricted to patients with GFR<60 ml/min/1.73m² because the KFRE was developed in adult patients with GFR below this level [3]. However, by design, patients with GFR<30 ml/min/1.73m² at baseline were not eligible for enrollment in cohort 1 of the CKiD study, and the minimum GFR at entry was further restricted to 45 ml/min/1.73m² for cohort 2 to allow longer follow-up before progression to kidney failure [10]. Results from the current study may therefore not be generalizable to OU patients with CKD outside of this enrollment criterion. Also, because children younger than 1 are not enrolled in the CKiD [9], we are unable to comment on the accuracy of the KFRE for the small subset of OU patients who have been shown to progress rapidly to kidney failure within the first year of life [7].

Despite these limitations, our study adds to the scarce literature related to the study of CKD progression in children with OU, a predominant cause of non-glomerular CKD. Findings suggest that while the KFRE provides similar discrimination for the 5-year kidney failure risk in children with OU as those in the overall CKiD cohort, there continues to be significant room for improvement, that the 8-variable model failed to provide. Such opportunities may include updating the KFRE model to better take into account the baseline risk and risk factor distribution in children with OU, such as through a calibration factor for OU [11], and/or expanding the model with emerging markers not included in the original KFRE equation [21,22]. Further study is planned to evaluate approaches to improve such predictions.

Conclusions

In children with OU enrolled in the CKiD study, the 4-variable KFRE discriminated the 5-year risk of kidney failure with C statistic values that were lower than previously published in adults but comparable to suboptimal values reported in the overall CKiD population. Use of the 8-variable equation did not significantly improve model discrimination or accuracy, suggesting the need for continued research into additional, disease specific markers of pediatric CKD progression in OU.