Kidney disease progression in children and young adults with pediatric CKD: Epidemiologic perspectives and clinical applications

Abstract

Chronic kidney disease (CKD) progression is typically characterized as either time to a clinically meaningful event (like dialysis or transplant), or longitudinal changes in kidney function. This review describes pediatric kidney disease progression using these two distinct frameworks by reviewing and discussing data from the Chronic Kidney Disease in Children (CKiD) study. We first describe new equations to estimate glomerular filtration rate (GFR) for patients under 25 (U25) years old, and how the average of serum creatinine-based and cystatin C-based GFR equations yield valid estimates than either alone. Next, we present a lifecourse description of CKD onset to kidney replacement therapy, prediction models based on clinical measurements, and show the importance of diagnosis (broadly classified as nonglomerular and glomerular in origin), GFR level and proteinuria on progression. Literature on longitudinal GFR in children and young adults are reviewed and new data are presented to characterize non-linear changes in U25 eGFR. These models demonstrated accelerated progression associated with glomerular diagnosis, lower GFR level and higher proteinuria, which was congruent with time-to-event analyses. Descriptions of online tools for GFR estimation and risk stratification for clinical applications are presented and we offer key epidemiologic considerations for the analysis of longitudinal pediatric CKD studies.

Chronic kidney disease (CKD) progression is characterized by cumulative loss of kidney function over time, eventually leading to end stage kidney disease (ESKD)^{1, 2}. The most widely used measure of kidney function in children, clinically and in research, is glomerular filtration rate (GFR) which can be measured directly^3–5 or estimated using equations based on commonly measured biomarkers such as serum creatinine or cystatin C^6–8. CKD patients who experience ESKD require kidney replacement therapy (KRT) in the form of either a kidney transplant or dialysis. Understanding disease progression through clinical studies, such as longitudinal cohorts, provide information to help anticipate disease risk, identify risk factors for progression, and ultimately improve overall clinical care and inform patients, family and caregivers.

The purpose of this article is to summarize and contextualize key findings characterizing CKD progression in children, adolescents and young adults from the Chronic Kidney Disease in Children (CKiD) study. CKiD is a longitudinal observational study of children enrolled with mild-to-moderate CKD and followed prospectively through annual study visits^{9, 10}. At the time of the study’s inception in 2003, it was recognized that evidence-based knowledge regarding pediatric CKD and the characterization of the disease’s natural history were incomplete. The scientific objectives of the CKiD study are broadly classified within four domains: kidney disease progression, cardiovascular heath, growth and development, and neurodevelopment, cognitive abilities, and behavior¹⁰. Over the course of the study, CKID investigators have made valuable clinical and epidemiological contributions to the pediatric CKD knowledge base.

Here, we first briefly discuss methods used to measure kidney function because it is central to evaluating CKD progression. We summarize recently developed equations designed to estimate kidney function in people under 25 years of age with a history of pediatric kidney disease, and highlight its clinical application. Second, we define what we mean by progression and consider two distinct ways to conceptualize and quantify progression from mild CKD to ESKD and KRT. Third, we present an analysis to describe longitudinal changes in kidney function across the pediatric lifespan and into young adulthood. In this review, we summarize key findings related to CKD progression from the longitudinal data of the CKiD study from an epidemiologic perspective and discuss how it can inform clinical care.

Estimation of glomerular filtration rate as a marker of kidney function

Valid estimates of kidney function are foundational for conducting research about CKD progression and for guiding clinical care of CKD patients. Developed in 2009, the CKID “bedside” equation is widely use in pediatric nephrology, clinically and in research⁶. The equation (eGFR = 41.3 × [ht(m)/sCr(mg/dL)]) benefits from a simple construct: a single constant multiplied to the ratio of a patient’s height and their serum creatinine. A drawback of this equation is that it overestimates GFR in young children (<5 years old) and underestimates in young adults (>18 years old)¹¹. Recently, CKiD investigators published two new eGFR equations for clinical and research use in populations under 25 years old using serum creatinine (sCr) and cystatin C, respectively⁸. These equations were of the form K₁ × (height/sCr) and K₂ × (1/cystatin C), and the values for K₁ and K₂ are sex- and age- dependent. Internal validation of these equations demonstrated estimated GFR without bias in both males and females across the pediatric and young adult age spectrum (1 to 25 years), and this was an improvement over the widely used “bedside” equation. Additionally, as shown in Table 1, for situations where both sCr and cystatin C are available, the simple average of the two eGFR values provides an intuitive linear combination of the two markers with eGFR values that offers better calibration, are more accurate and more precise than either of the two single marker estimates alone. These equations are race blind and are congruent with recent recommendations from the National Kidney Foundation and American Society of Nephrology to exclude race-based modifiers in estimating equations^{12, 13}. This calculator is available as a web-based application for clinical use (https://ckid-gfrcalculator.shinyapps.io/eGFR/)¹⁴.

Table 1.

Validation metrics of the CKiD under 25 years (“U25”) serum creatinine and cystatin C estimating GFR equations and their average from 308 children contributing 784 observations with age <18 years and 69 children with 107 observations with age ≥ 18 years.

	Arithmetic bias: eGFR – mGFR‡ (95%CI)		p10: eGFR within 10% of mGFR		p30: eGFR within 30% of mGFR		Test RMSE¥, ml/min|1.73m2
	<18 yrs	≥18 yrs	<18 yrs	≥18 yrs	<18 yrs	≥18 yrs	<18 yrs	≥18 yrs
Plasma/Serum creatinine (pCr or sCr)-based equation
CKiD U25 age- and sex-dependent	0.05 (−0.93, 1.03)	0.70 (−1.60, 3.00)	36.4%	44.9%	86.2%	90.7%	9.88	9.64
IFCC Cystatin C (cysC)-based equation
CKiD U25, age- and sex-dependent	−0.79 (−1.67, 0.09)	−0.89 (−3.05, 1.28)	39.3%	45.7%	86.6%	85.7%	10.48	10.81
Average of the creatinine and cystatin-based U25 GFR estimates	−0.33 (−1.12, 0.45)	−0.21 (−2.03, 1.61)	43.6%	52.4%	91.5%	87.6%	8.55	8.49

Abbreviations: CKID, Chronic Kidney Disease in Children Study; pCr, plasma creatinine; sCr, serum creatinine; U25, under 25 years;

^‡Calculated using the estimated fixed effect intercept of an intercept-only linear mixed effects model with a random intercept (participant) with the difference of eGFR-mGFR as the outcome where eGFR and mGFR are the estimated and measured (via iohexol) GFR values.

^¥Test root mean square error (RMSE) = $\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left(eGF{R}_i-mGF{R}_i\right)}^2}$

In the next sections, we expand on published findings on CKD progression from an epidemiologic perspective. Conceptually, disease progression is typically investigated by two different constructs: (1) time to an event of interest related to CKD, and (2) longitudinal patterns in markers of kidney function. We discuss each in turn starting with time to an event.

Chronic kidney disease progression and time to an event

It is common in epidemiological research to characterize disease progression using the time to some defined event. Such analyses benefit from a well-defined and meaningful event (i.e., outcome) as well as thoughtful consideration and construction of a time origin from which people are followed. For CKD progression, KRT initiation (dialysis or transplant) is frequently used because it is a milestone in clinical care and patient prognosis. Occurrence of an individual’s GFR dropping below a defined threshold (e.g., GFR < 15 ml/min|1.73m² which is the KDIGO-defined threshold for ESKD¹⁵) may also be used. A variant of this event is time to crossing a threshold that is dependent on and defined by a baseline marker level. Specifically, one may define a 50% (or 30%) decline in baseline GFR as an event that captures meaningful progression^16–18. This approach can be enhanced by interpolating the time at which the individual crossed the threshold using a linear assumption of decline from two GFR levels at known dates, rather than the time at first observed GFR less than the threshold.

For smaller studies with limited follow-up time, it is common to create a composite outcome combining any or all of these events and using the first occurrence of any as the time of the event. In such cases, defining the distribution of distinct event types and summarizing their cumulative incidence is critical for valid epidemiologic inference¹⁹. We note that events of accelerated kidney function decline (i.e., 50% decline in GFR) represent a different aspect of disease progression, both clinically and epidemiologically, compared to events like occurrence of ESKD (GFR < 15 ml/min|1.73m²) or initiation of KRT, which are “harder” outcomes.

Time origin and axis for time-to-event studies are defined independently of the event but must also be considered when designing a study and for inferences. Baseline in an observational cohort study usually refers to the time of study enrollment. While it is often the richest point in study time with respect to data completeness, it can be arbitrary in terms of the biological time scale of disease progression, especially when the eligibility criteria are broad. In CKiD, age at enrollment was between 2 to 16 years and with heterogeneous time since disease onset. We will present and discuss three papers investigating the impact of diagnosis²⁰, GFR and proteinuria¹⁷, and other risk factors for disease progression¹⁶ among children with CKD. Each paper has a unique study design including distinct time origins for targeted epidemiologic inference.

The aim of a paper published by Ng et al.²⁰ was to describe the lifecourse of disease from onset to kidney replacement therapy by diagnosis. We broadly grouped CKD etiology into three categories: nonglomerular disease, hemolytic uremic syndrome (HUS), and other glomerular, non-HUS disease. Nonglomerular diagnoses (N=650, 71% of 915 participants in the study population) comprised mostly congenital anomalies of the kidney or urinary track (CAKUT) diagnoses, including: obstructive uropathies (25%), dysplastic kidneys (24%), reflux nephropathy (18%), bilateral hydronephrosis, agenesis of abnominal musculature and VACTERL, and other non-CAKUT diagnoses, including autosomal recessive polycystic kidney disease, renal infarct and cystinosis. Hemolytic uremeic syndrome (N=49, 5% of study population), as a glomerular diagnosis, was treated as a separate category because of its unique disease trajectories noted in previously published research²¹. Among Glomerular, non-HUS diseases (N=216, 24% of study population), specific diagnoses included focal segmental glomerulosclerosis (35%), systemic Lupus erythematosus (17%), and other diagnoses like chronic glomerulonephritis and familial nephritis. While we recognize that there are likely diagnosis-specific trajectories, broadly aggregating diagnoses with common physiologic processes was necessary because of limited representation due to the rare nature of pediatric CKD.

In Ng et al.¹⁸, the time origin (t=0) was disease onset and the time scale was years from disease onset to KRT or censoring (e.g., loss to follow up or administrative censoring at date of analysis). Late entry methodology was required to deal with study enrollment at heterogeneous durations of disease^{20, 22}. The analysis used parametric survival regression and fit generalized gamma distributions to characterize the diagnosis-specific cumulative incidence of KRT and used the model to predict years after disease onset by when nearly all children with CKD were likely to experience KRT^{23, 24}.

Figure 1 presents the cumulative incidence of KRT by diagnosis, with corresponding coefficients for the generalized gamma distributions denoted by GG(β, σ, κ). The study found that half of children with nonglomerular CKD diagnoses will experience KRT by 18.9 years of age (corresponding to disease duration, since diagnoses in this group were almost entirely congenital). Furthermore, it was estimated that nearly all of these children will experience KRT by 42.5 years of age. Obstructive vs. non-obstructive uropathy diagnoses did not differ in times from diagnosis to KRT. For those with non-HUS glomerular diagnoses, the median time to KRT was significantly shorter (7.6 years after disease onset) with nearly all expected to experience KRT by 25.4 years after disease onset. Since glomerular diagnoses are generally not congenital (average age of onset in CKiD was older than 10 years), these times estimates do not equate directly with age. To estimate ages, the analysis was extended to provide an average age of KRT among those with glomerular, non-HUS diagnoses using a simulation approach. The age at which 50% of children with glomerular, non-HUS diagnoses experience KRT was estimated at 15.2 years and at 37.9 years for KRT initiation in nearly all (i.e., 99%) of this population. Interestingly, the 99^th percentile for age were remarkably similar between those with nonglomerular and glomerular non-HUS diagnoses (42.5 vs. 37.9 years). We note that these upper limits were predictions based on parametric (model-based) extrapolation. Continued follow-up of this cohort will confirm or modify these estimates. We also note that these results are best interpreted from a broad, epidemiologic and public health perspective. The estimates are not specific to a particular patient profile, but may help clinicians convey the natural history of CKD to patients and families. Indeed, clinicians may find it useful to explain that nearly all children who have similar profile to those who enrolled in the CKiD study will be expected to receive KRT by about 40 years of age. We will discuss several modifiable factors to delay progression to KRT, and this epidemiologic perspective of expected time course of CKD progression may help motivate effective disease management and therapy adherence for patients and families through knowledge and realistic expectations²⁵.

Figure 1.

Incidence of kidney replacement therapy (RRT) after kidney disease onset among participants with nonglomerular (blue; n= 650), hemolytic uremic syndrome (HUS; green; n= 49), glomerular non-HUS (red; n= 216) diagnoses. Continuous step functions represent non-parametric estimates of the cumulative incidence of RRT. Dashed lines represent group-specific parametric survival models based on the generalized gamma (GG) family with parameters listed as GG(β, σ, κ). Median and 99th percentile times to RRT in years after kidney disease onset are presented with 95% confidence intervals for the 99th percentile. Reprinted with permission²⁰.

Ng et al.¹⁸ also explored timing of dialysis and transplant separately, by CKD diagnosis and the results of the cumulative incidence of dialysis and survival transplant-free are presented in Figure 2. The unique feature of this analysis was that first dialysis and first transplant were treated as competing risk events: that is, once a participant initiated dialysis as their first KRT event, they were no longer considered at risk of transplant as their first KRT event. While this logically makes sense, it is methodologically important to make this distinction to avoid treating an outcome as being censored for the other (i.e., still at risk for the other at a later point in time). Those with glomerular non-HUS diagnoses were significantly more likely to experience dialysis as a first KRT event than those with nonglomerular diagnoses. Children with nonglomerular diagnoses were about equally likely to initiate dialysis as to receive a transplant (49% vs. 51%); but those with glomerular non-HUS disease had a significantly higher probability (84%) of starting KRT via dialysis initiation. We hypothesized that higher risk of dialysis was due to accelerated CKD progression associated with glomerular diagnoses relative to nonglomerular diagnoses, and the greater difficulties of planning efforts for transplant under faster progression and potentially shortened timelines. It is important to note that Atkinson et al.²⁶ demonstrated that among children who experienced dialysis as the first modality of KRT, about 50% received a kidney transplant within 12 months of initiating dialysis. Nonetheless, the evidence from CKiD has been consistent and clear that diagnosis strongly influences progression and disease course, and we further demonstrate that this can inform our clinical understanding for prediction and risk stratification.

Figure 2.

Incidence of first transplant or dialysis as competing events among (A) nonglomerular (n= 650; blue) and (B) glomerular non-hemolytic uremic syndrome (n= 216; red) diagnoses. Continuous step functions represent non-parametric competing risk estimates of the cumulative incidence of first dialysis (bottom, pale color) or first transplant (top, dark color). Dashed lines represent group-specific parametric mixture models with corresponding mixture estimate (%) describing the proportion expected to receive dialysis or transplant. The full parametric survival models estimated by maximum likelihood methods can be described by mixtures of Generalized Gamma (GG(β, σ, κ)) or Weibull (WE(β, σ)) distributions. The coefficients for nonglomerular diagnoses are 49% ~ WE(3.047, 0.399) for dialysis and 51% ~ WE(3.106, 0.367) for transplantation; for glomerular diagnoses are 84% ~ GG(3.160, 0.208, 9.740) for dialysis and 16% ~ WE(2.500, 0.488) for transplantation. Source: Reprinted with permission²⁰.

Prediction model with longitudinal data for risk stratification

Warady et al.¹⁶ published an analysis identifying predictors of rapid progression, stratified by glomerular and nonglomerular diagnoses. This analysis used parametric survival models to estimate times to a composite outcome of KRT or a 50% decline in GFR from baseline. The unique feature of this analysis was that the time origin was the second study visit observed, or approximately one year into the study. This approach allowed the authors to relate one-year longitudinal changes in risk factors as well as current clinical status as predictors of progression. Longitudinal factors were classified as persistent, incident, or resolved conditions. Persistent conditions were those present at two clinical presentations about one year apart; incident conditions were those that were absent and then present about a year apart, and resolved conditions were the reverse. The reference group comprised those free of the condition of interest in both visits. Using a prediction modeling approach and internally validated using crossvalidation, several risk factors were identified, each with varying strength of association on the outcome.

For those with nonglomerular diagnoses, age and sex were modifiers, and important current clinical factors included: GFR (categorized as >= 45, 30 to 45, and < 30), hypoalbuminemia, BP ≥ 90th percentile, and dyslipidemia. Longitudinal factors associated with accelerated progression for nonglomerular diagnoses were nephrotic range proteinuria (urine protein to creatinine ratio > 2 mg/mg), and anemia (low hemoglobin for age and sex standards) and ACEi/ARB use, classified as persistent use, incident use and discontinuation. Interestingly, ACEi/ARB was associated with shorter time to the composite outcome, but this finding should be interpreted from a purely predictive perspective. Specifically, therapy use is an indicator for more severe disease, and this association is very likely confounding by indication (in which those with more advanced disease are likely to be treated). Indeed, ACEi/ARB have been consistently associated with effective CKD management and delayed progression²⁷. In this setting, the association of ACEi/ARB use and progression must be interpreted within the construct of prediction, not as a measure of medication effectiveness.

The model for those with glomerular diagnoses had fewer significant predictors, most likely due to the smaller sample size. Despite this potential limitation, many clinical variables were identified as strong predictors of faster progression and shorter times to KRT. Interestingly, these predictive variables were a subset of those identified for children with nonglomerular diagnoses and included: lower current GFR, hypoalbuminemia and BP > 90th percentile. Longitudinal factors included proteinuria and anemia classified as none, persistent, incident or resolved. For both diagnosis groups, the importance of proteinuria, blood pressure and other associated markers of CKD severity (anemia and albuminemia) was consistent with many other studies^28–32.

These predictive models identified risk factors for progression and operate as a useful clinical tool to estimate times to KRT or a 50% decline in GFR. A web-based application designed to facilitate clinical use of these progression prediction models is available on the National Kidney Foundation website (https://www.kidney.org/professionals/kdoqi/gfr_calculatorPedRiskCalc)³³. Using these models, one can estimate the time in years by which 50% of children with a specific clinical profile will experience KRT or 50% decline in GFR. To provide estimates of variability, we also estimate the time by which 10% and 25% of children with the specified clinical profile will experience the outcome. For example, the calculator provides a set of times quantifying aggregated incidence of the outcome for a population of patients with a given clinical profile (e.g., among patients with this profile, 10% of patients will have the outcome by 5 months, 25% by 1 year and 2 months years, and 50% by 3 years and 5 months).

A limitation of this calculator is that full data must be entered for the calculator to estimate times to the outcome. If data are missing (for example, dyslipidemia), an option may be to input two profiles, one with the condition present, and the other with the condition absent, to estimate a range of times. We note that this approach will only work for one missing data element. In addition, the calculator will not work if the clinician does not have patient data from approximately one year ago to calculate longitudinal changes. If more than one variable is missing, or the clinician is seeing the patient for the first time (or without access to medical charts or lab data from one year ago), this calculator cannot be used. Fortunately, a different calculator has been developed using only three widely available clinical markers of CKD - diagnosis, GFR level and proteinuria level¹⁷. It too offers valid and clinically meaningful risk stratification.

Simplified clinical prediction model for risk stratification

KDIGO (Kidney Disease: Improving Global Outcomes) disease progression risk models for adults define staging and progression risk by GFR and albuminuria (or proteinuria)¹⁵. Historically, CKD staging and disease progression have been more robustly researched and quantified in adults with CKD than in children. On this point, 2012 KDIGO guidelines acknowledged a lack of data and research dedicated to improving understanding and quantifying disease progression in children with CKD. To address this deficit, Furth et al.¹⁷ used data from 1169 children enrolled in two large multi-center longitudinal studies, the CKiD Study (n=857) and the Effect of Strict Blood Pressure Control and ACE Inhibition on Chronic Renal Failure Progression in Pediatric Patients (ESCAPE) clinical trial (n=312), to quantify risk of CKD progression as a function of three common clinical indicators: CKD diagnosis, GFR level, and proteinuria. Disease diagnosis was classified as glomerular or nonglomerular. Baseline estimated GFR was categorized using the KDIGO classification system: G1 (≥90), G2 (60–89), G3a (45–59), G3b (30–44) and G4 (15–29 ml/min|1.73m²). Proteinuria was measured as first morning urine total protein-to-creatinine ratio and was classified as <0.5, 0.5–2.0 and >2.0 mg/mgCr. These cut points were chosen for adequate data in each level rather than based on a priori clinically meaningful cutoffs. The outcome event was a composite of initiation of dialysis or a transplant, eGFR dropping below 15 mL/min/1.73m², or a longitudinal decline in eGFR from baseline of 50% or more.

Among 1169 children, the median baseline age was 12 years and median eGFR was 47 mL/min/1.73m². Children with glomerular CKD comprised 25% (n=297) of the study population. Total follow-up was 4698 person-years with median individual follow-up time of 3.8 years (4.1 years in nonglomerular CKD children; 2.3 years in glomerular CKD children). A total of 412 events were observed: n=65 initiated RRT, n=137 had eGFR drop below 15 mL/min/1.73m², and n=210 experienced a 50% decline in eGFR over follow-up. Using parametric survival models, GFR-proteinuria categories with similar risk of the outcome were combined until no further aggregation of similar risk groups was possible. This resulted in six distinct risk groups A through F of increasing risk and shorter times to the composite event. The combinations of GFR and proteinuria that define each risk group are described in Figure 3, and it is important to note that some risk groups are not contiguous. Diagnosis was included as a covariate in the model that modified the incidence of the composite event, but not the overall shape of the incidence function.

Figure 3.

Combinations of GFR (y-axis) and urine protein:creatinine ratio (x-axis) to define CKD progression risk groups A (lowest risk) through F (highest risk) based on data from the Chronic Kidney Disease in Children study and the Effect of Strict Blood Pressure Control and ACE Inhibition on Chronic Renal Failure Progression in Pediatric Patients (ESCAPE) clinical trial. Adapted from Furth et al., Am J Kidney Dis 2018.¹⁷

The six progression risk groups offer strong discrimination for risk of disease progression as defined by the composite outcome in both nonglomerular [area under the receiver-operator characteristic (AUROC)=0.83] and glomerular (AUROC=0.80) CKD children. Use of parametric survival models allowed for the estimation of quantile-specific survival times, a clinically useful statistic that providers can use to communicated expected patient progression time lines to family and caregivers and plan for future clinical interventions. The ability to provide quantile specific times to progression allows the provider to communicate a richer picture of the anticipated progression timeline including the heterogeneity that exists within diagnosis-specific risk groups. The authors opted to reported the expected times by which 10%, 25% and 50% of a given diagnosis-specific risk group would experience significant disease progression. Crossvalidation methods and comparison to non-parametric Kaplan-Meier survival curves demonstrated strong calibration and expected times-to-event in each of the risk groups.

Filler and McIntyre suggested that these estimates may represent the “best case scenario” for children in clinical care since study populations often times over-represent patients with better care and disease management³⁴ and this was an example of the Hawthorne effect. The data for this risk calculator used the original CKiD “bedside” equation⁶ which has subsequently been shown to overestimate measured GFR in young children and underestimate it in young adults¹¹. Such misclassification of true GFR may attenuate the true discrimination of this risk model if a more precise estimation of GFR can be used. Recently published estimating equations by CKiD researchers described above have overcome the age-related biases of the “bedside” equation and their use in defining a child’s risk group may provide even stronger discrimination⁸. Another limitation of these risk models is the use of a composite outcome that includes both clinically hard (initiation of RRT and G5 CKD stage) and soft (50% decline in GFR) endpoints. In fact, 51% of the observed events were those who experienced 50% GFR decline. While such declines are indicative of significant disease progression, especially among patients with mild-to-moderate CKD, they represent a less clearly defined endpoint from the perspective of actionable clinical care. Nonetheless, this simple calculator is particularly useful for clinical care, prognosis and transplant planning. A web-based application allows for easy calculation based on diagnosis, GFR and proteinuria www.ckdprognosis.com. All pediatric calculators described in this review may be accessed through the CKiD website (www.ckidstudy.org)³⁵.

Disease progression and longitudinal changes in GFR

Longitudinal change in GFR is another useful metric to characterize progression. Similar to investigations of time to KRT, glomerular and nonglomerular classifications of diagnosis is helpful for structuring epidemiological summaries and clinical management practice. Using longitudinal data from 529 children (n=109 G; n=420 NG), Pierce et al.³⁶ characterized longitudinal changes in GFR as individual-specific annualized ratios. This metric quantifies the expected change in a child’s GFR over one year with values less than one capturing a loss of kidney function. An annualized ratio value of 0.90 captures an expected one-year decline in GFR of 10% while a ratio of 1 represented no longitudinal decline. Median follow-up time was three years and 89% of children had 3 or more GFR measurements.

Traditional regression methods based on Gaussian distributions are designed to detect and quantify a rigid shift in an outcome like longitudinal GFR rate of change associated with an exposure. These methods may fail to capture the nuanced differences in the distributions of GFR change between diagnosis groups. Pierce et al.⁸ used more flexible generalized gamma regression methods that allowed for percentile-specific heterogeneity of differences in GFR change between glomerular and nonglomerular CKD children.

While median changes in GFR were different between nonglomerular and glomerular diagnoses, −3.9% (−11.4 to 2.2) and −10.5% (−23.4 to −1.2) per year, respectively, the largest difference between the two diagnosis groups was in the lower tail of the distributions with a significantly larger proportion of those with glomerular diagnoses demonstrating rapid disease progression (> 30 % decline per year) compared to children in those with nonglomerular diagnoses. At the same time, there was little to no difference in the distributions’ upper tails indicating that both diagnosis groups contained an equal presence of slow or non-progressors. Generalized gamma regression models were able to accurately capture the asymmetric distributions of GFR change. A limitation of this analysis was that it assumed a constant, linear decline and also did not investigate modifiers of longitudinal GFR change.

Characterizing CKD progression by GFR decline using new estimating equations and enriched models

In this section, we present new data to address limitations in Pierce et al.⁸. Specifically, we stratified by CKD diagnosis (glomerular vs nonglomerular CKD) and characterized non-linear longitudinal GFR decline by level of initial proteinuria level in each CKD diagnosis group using linear mixed models with a polynomial (quadratic) time parameter. GFR was estimated using the U25 eGFR equations based on both serum creatinine and cystatin C (the simple average of the two single marker estimates) when available. Analysis was restricted to GFR estimates from study visits prior to KRT initiation.

Specifically, we constructed models stratified by CKD diagnosis (glomerular and nonglomerular) with U25 GFR as the dependent variable. Time (t_ij) and time-squared $\left(t_{ij}^2\right)$ from baseline (study entry) were the independent variables; initial proteinuria level (<0.5, 0.5 to 2.0, > 2.0 mg/mgCr) modified the relationship between initial GFR level (the intercept; t₀ = study entry ) and both parameters for time (i indexes participant and j indexes the serial longitudinal visit within participant). Random effects for intercept (a_i) and the two time parameters (b_i and c_i, respectively ) accounted for within-person correlation of repeated measures. The model was of the form:

$$\text{Level}\text{at}t=0:GF{R}_{ij}={\alpha }_0+{\alpha }_1\ast medProteinuri{a}_i+{\alpha }_2\ast highProteinuri{a}_i+a_i$$

$$\text{Linear:}+\left({\beta }_0+{\beta }_1\ast medProteinuri{a}_i+{\beta }_2\ast highProteinuri{a}_i+b_i\right)\times t_{ij}$$

$$\text{Quadratic:}+\left({\gamma }_0+{\gamma }_1\ast medProteinuri{a}_i+{\gamma }_2\ast highProteinuri{a}_i+c_i\right)\times t_{ij}^2$$

$$\text{Residual:}+e_{ij}$$

Models assuming linear decline were also fit (i.e., no quadratic term, or γ = 0) to compare and contrast with non-linear assumptions. Bayesian information criteria (BIC) were calculated to determine goodness-of-fit of quadratic models versus linear models.

Table 1 provides a description of demographic, clinical characteristics and longitudinal data for the two diagnosis groups. A total of 271 children with glomerular diagnoses contributed 1261 observations. Their average age at study entry was 14.2 years, about half (53%) were boys, and 30% were of Black race. Their age at disease onset was, on average, 8.5 years. Average eGFR at entry for these participants was 58 ml/min|1.73m², and 57% had proteinuria in the elevated or nephrotic range (≥ 0.5 or > 2.0, respectively). Most glomerular CKD participants had at least 4 eGFR measurements (61%) and 50% had at least 3.4 years of follow-up.

A total of 778 children with nonglomerular diagnoses contributed 4182 eGFR observations over 3677 total years of follow-up time. These children entered the study at an average age of 8.4 years, two-thirds (67%) were boys, 19% were Black race and 14% were Hispanic ethnicity. Disease onset was nearly all at birth (89%). These participants entered with an average eGFR of 47 ml/min|1.73m² and 63% had proteinuria < 0.5 mg/mgCr. The average amount of study follow-up time was 4.2 years, and 64% contributed at least 4 eGFR observations over follow-up.

For both diagnosis groups, mixed effects models including a quadratic term for time fit significantly better than models with only a linear term for time. Specifically, the BIC was lower for the quadratic model compared to the linear model for both the nonglomerular (31188.2 vs. 31227.4) and glomerular (9983.2 vs. 9990.3) CKD groups.

Figure 4 presents a visual depiction the diagnosis-specific results of the longitudinal models for each proteinuria category including the average eGFR at entry and how eGFR changes over time. This figure demonstrates the accelerated decline associated with glomerular diagnoses relative to nonglomerular diagnoses, a result consistent with earlier times to KRT in a survival setting^{16, 17} and with decline modeled with linear assumptions³⁶. In addition, higher proteinuria levels were associated with lower GFR at study entry, and faster GFR decline in both diagnosis groups. The curves show the heterogeneous GFR trajectories, both in terms of initial level and changes over time, by diagnosis and proteinuria level. We note that GFR decline from these models is not a single value (i.e., a constant change per year), but is rather a function of time and this may be easily derived as the first derivative of the equations presented in the figure. The fastest progression observed were among those with glomerular diagnoses and high proteinuria. Within each diagnosis group, those with highest proteinuria entered the study with the lowest eGFR level indicating advanced disease progression prior to study entry. This graph clearly depicts the importance of proteinuria as a modifiable clinical risk factor as well as indicator for future disease progression.

Figure 4.

Nonlinear estimated GFR trajectories over time, by initial proteinuria levels and stratified by underlying kidney disease diagnosis (nonglomerular and glomerular).

We note that these models show average trajectories of the population, and not individual-specific ones. Furthermore, specifying a quadratic term is more flexible than a strict linear assumption, but these models still impose parametric assumptions of a quadratic trajectory. Lastly, these are simplified models in the sense that they only include proteinuria level as an exposure and there are several previously identified modifiers which were not included such as blood pressure^{16, 18, 32}, fibroblast growth factor-23 (FGF23)³⁷, anemia³⁸, race^{39, 40}, and therapy use^{27, 28}. These GFR trajectory models correspond with the key variables in the AJKD risk staging based on time-to-event analysis¹⁷ and underscore the importance of diagnosis, proteinuria and initial level of GFR as a predictors of disease severity, and also demonstrate the construct consistency between longitudinal changes in GFR and time to KRT or a 50% decline in GFR.

Conclusions

Characterizing disease progression in populations with pediatric CKD is not a simple matter, but conceptualizing progression by either time to a particular event, or GFR decline will yield unique and distinct insights. We underscore the importance of specifying a targeted research question which will inform the analytic structure and inference. We demonstrated the utility of these approaches for etiologic purposes, identifying important modifiable clinical factors (proteinuria and blood pressure), and prediction for risk stratification. We also note that our presentation and summary of findings were restricted to the latest literature on the epidemiology of disease progression and not other complex systems that modify CKD severity and progression. Indeed, this paper characterized the natural and treated history of CKD progression in children but these are not deterministic models and substantial heterogeneity in trajectories are present. Modifiable conditions, including management of blood pressure^{18, 32}, metabolic acidosis⁴¹, anemia²⁹, uric acid⁴² and growth failure⁴³, and improved medication adherence⁴⁴, have all demonstrated important clinical roles for delaying CKD progression in this observational setting. The scientific findings reviewed in this paper will help clinicians understand the concepts behind analytic designs and how unique models will yield different interpretations and aspects of disease progression.

We presented updated GFR equations that offer valid estimates of kidney function for patients under 25 years old, the utility of taking the simple average of creatinine- and cystatin-based equations, and useful clinical web-based prediction models for risk stratification. Lastly, we described what is currently known about the timing of clinically meaningful events such as KRT, and longitudinal GFR, as constructs of CKD progression and how a few key clinical variables (including diagnosis, GFR and proteinuria level) modify the course. Our hope is that these findings can be used for improved clinical decision making, sharing patient-centered information with family, and offer useful considerations in the design of epidemiologic studies of pediatric kidney diseases.

Table 2.

Demographic and clinical characteristics of the CKiD cohort and a description of longitudinal data available to characterize GFR changes over time, stratified by glomerular and nonglomerular diagnoses. Median [IQR] or % (n).

Variable	Glomerular diagnoses (n= 271)	Nonglomerular diagnoses (n= 778)
Demographic characteristics
Age at study entry, years	14.2 [10.8, 15.9]	8.4 [4.6, 12.9]
Male sex	53% (143)	67% (521)
Black race	30% (82)	19% (147)
Hispanic	15% (42)	14% (111)
Kidney disease characteristics
Age at disease onset, years	8.5 [2.5, 12.5]	89% (693) at birth
U25 eGFR, ml/min|1.73m2	58 [42, 75]	47 [35, 62]
Proteinuria category
uPrCr < 0.5 mg/mgCr	44% (118)	63% (489)
uPrCr 0.5 to 2 mg/mgCr	32% (86)	28% (220)
uPrCr > 2.0 mg/mgCr	25% (67)	9% (69)
Longitudinal data
Total follow-up time, years	1082	3677
Follow-up time per participant, years	3.4 [1.8, 6]	4.2 [1.5, 7.2]
Total eGFR observations	1261	4182
n eGFR observations per participant
1 to 3	39% (106)	36% (283)
4 to 6	38% (103)	27% (212)
7 to 9	19% (50)	23% (181)
≥10	4% (12)	13% (102)