Low agreement between modified-Schwartz and CKD-EPI eGFR in young adults: a retrospective longitudinal cohort study

Descriptive retrospective longitudinal observational cohort study.

Patients were recruited from the University of North Carolina STAR_x (Self-management and Transition to Adult-focused healthcare with R_x = treatment) Program (recruitment details noted in Ferris et al. 2015) [14‐16].

Patients from the STAR_x Program were eligible for inclusion in this study if they were prevalent or incident CKD cases followed annually at either the paediatric- or adult-focused nephrology clinics. Patients were excluded if they had spina bifida or muscular dystrophy, conditions that would drastically alter muscle mass. We consented patients and parents to access their electronic health record (system available since 1984) under a protocol and using forms approved by the University of North Carolina at Chapel Hill. The study adhered to the Declaration of Helsinki. The patients’ longitudinal medical and laboratory information since their first visit to the medical centre was obtained. For this study, we collected the patients’ basic demographic information (e.g., age, sex, and race), serum creatinine, anthropometrics (e.g., height and weight), and if applicable, dates of transplants and dialysis. Individuals were censored from the population after their first dialysis or after renal transplant due to potential inaccuracy in creatinine clearance estimation because of renal replacement therapy or graft function. Analyses (with the exception of the graphical population curves) required complete case data on age, race, sex, gender, and creatinine.

The primary outcome variable was their eGFR, which was calculated by the new Schwartz or new CKD-EPI estimating equations based upon creatinine, with all required covariates abstracted from the medical record. These equations were chosen because cystatin C measurements were rarely, if ever, ordered during the time period covered by the study. Schwartz eGFRs were calculated using the new Schwartz estimating equation: [7].

CKD-EPI eGFRs were estimated with the 2009 iterations of the CKD-EPI equation: [8]

Where C = 144 and S = 0.7 if the patient was female and C = 141 and S = 0.9 if the patient was male. We used k = − 0.329 if the patient was female with serum creatinine ≤0.7, k = − 1.209 if the patient was female with serum creatinine > 0.7, k = − 0.411 if the patient was male with serum creatinine ≤0.9 and k = − 1.209 if the patient was male with serum creatinine > 0.9. For use with SI units, S and serum creatinine bounds were replaced with 61.9 for females and 79.6 for males; while the other coefficients remained the same.

The UNC laboratory measured creatinine using the OrthoClinical Diagnostics Vitros Creatinine slide. They used Jaffe’s reaction and switched to an enzymatic IDMS-traceable assay in May 2008; prior measurements were corrected by multiplying them by a factor of 0.95. All heights were measured with a standing stadiometer at routine clinic visits. Heights were cleaned to remove implausible values (i.e. > 300 cm) or values that were entered with incorrect units (i.e. a patient going from 100 cm to 254 cm back to 100 cm), and if a particular creatinine did not have a height associated with the visit, heights were interpolated as a weighted average of the previous and next height measurements. Measurements for each patient within each half-year were averaged to obtain an overall estimate for that half-year of age.

Age was used as a continuous variable. Z-score of height was also used to divide patients into thirds relative to the z-score distribution of the overall cohort and allow assessment of concordance differences based upon relative stunted growth. Height z-scores were based on the Centers for Disease Control and Prevention’s growth charts [17].

No formal methods were used to correct for systematic biases. However, patients were enrolled consecutively during the study period with a response rate of 97% and lack of time was the sole reason given for non-enrolment. Thus, the patient sample is representative of the patients UNC serves in terms of race, sex and insurance. Given that more than half of the individuals studied had an eGFR> 60 at study entry, there remains the possibility that the results could be biased due to systematic underestimation of the eGFR in that group using the Schwartz equation. However, bias of the Schwartz equation in children is greater in children with low GFR, rather than GFR close to normal [18].

We used graphical and statistical methods to capture changes in the extent of disagreement between the new Schwartz and the CKDEPI eGFR estimations as patients in our cohort aged. Locally weighted polynomial regression (LOESS) curves were generated for each measure of eGFR over time using R packages “ggplot2” [19] and “ggExtra” [20]. These curves represent the average of the eGFR from a given equation at each age across the total population (with each individual representing at most one data point at a given age). The closer the two lines are to one another at a given age, the more similar the overall population estimates generated by the new Schwartz and CKDEPI are. Similar graphs were constructed to compare the new Schwartz and CKDEPI eGFR estimation in relevant subgroups of interest with stratification by sex, race, type of chronic kidney disease, and z-score of height at current age.

Considering that we lack access to true measured GFR data in this longitudinal data set (i.e. inulin clearance), statistical disagreement was captured using the concordance correlation coefficient. This method assesses the extent of agreement for two measures of the same base quantity more effectively than a t-test or normal correlation coefficient [21]. A variant of this method has been adapted for use through repeated measurements on individuals, which allows us to leverage the full breadth of our longitudinal data set rather than relying on cross-sectional cuts. This method also allows us to estimate the concordance correlation coefficient across different age periods, assessing changes in our estimate of the concordance correlation coefficient as patients aged. We estimated concordance correlation coefficients using R package “cccrm” [22] and creatinine measurements from age 10–15, 15–20, 20–25, and 25–30; these age groups were chosen to make sure early adulthood could be split into two phases, as well as give one final category to assess agreement once individuals were far from adolescence. We also created a Bland-Altman plot (created using R package “BlandAltmanLeh” [23]) for each age group to visualize information about disagreement; unfortunately, unlike the concordance correlation coefficient, these plots cannot integrate and adjust for information about repeated measurements. To assess whether this would dramatically alter results, we created a second set of supplementary Bland-Altman plots in which individuals only contributed one dot per age group based upon their mean values in that age range. To assess whether subgroup effects might differ, sub-analyses were conducted across relevant strata of interest, including sex, race (African-American vs. not African-American), type of kidney disease (glomerular vs. not glomerular), and Z-score of height based upon the CDC’s growth charts [17].

To investigate potential differences in CKD stage categorization between the equations at age 18 differentially across height groups, we constructed an additional graph. Data management and table generation were performed in both SAS 9.4 for Windows (SAS Institute Cary, North Carolina) and R using packages “dplyr” [24] and “xtable [25].”

We used novel methods to estimate the proportions of the variance in the ratio of the CKD-EPI value to the new Schwartz value explained by height Z-score, type of CKD, age, sex, and race. The classical R² for linear models measures the proportion of variance explained by a set of fixed predictors. Additionally, semi-partial R² statistics measure the relative explanatory power of one specific predictor in the set after accounting for the others. Several analogues of the R² are available for linear mixed models (LMMs), which are frequently used to characterize longitudinal data. Using the “lme4” [26] package, we fitted a LMM with the ratio of eGFRs modelled as a function of race, sex, age, and height Z-score. Using the “r2glmm” [27] package, measures of generalized explained variance were computed for the model and each predictor. Generalized explained variance may be interpreted as the proportion of multi-dimensional scatter that can be attributed to structural components (i.e. predictors) of a LMM [28].

We conducted a complete case analysis; that is, observations from a patient were only eligible for inclusion in our assessments if they were not missing data in any of the variables used to assess their eGFR. Once patients were lost to follow-up in creatinine measures, they no longer contributed to our eGFR comparisons.

We enrolled 292 AYAs from the STAR_x Program who had the diagnosis of CKD. We excluded 50 AYAs from the analysis since they did not receive care at the clinic prior to their transplant or initiation of dialysis, yielding no usable creatinine measurements. No AYAs were identified with either spina bifida or neuromuscular dystrophy.

A majority (59%) of these AYAs had glomerular disease, and slightly less than half (47%) of patients had CKD stages ≥3 based upon an estimate of their renal function using the new Schwartz and the KDIGO categories [29] at their first recorded height. Over half (55%) of patients were male, and over one-third (38%) were African-American, representative of the population we serve. The median age at entry into this cohort was 13.6 years (interquartile range 11.4; 16.3 years) and the median duration of follow-up was 4.0 years (interquartile range 1.4; 6.3 years); both were not normally distributed. In general, AYAs had greater than average BMI, though there was a wide variance and while 26 AYAs had missing heights at all their visits these patients represented very little potential follow-up time overall. Otherwise, missing data for key variables was negligible. For full details of the descriptive population statistics, see Table 1.

Figure 1 is the smoothed population eGFR trajectories for the patients included in the study population across time. Panel A is the trajectory for the overall population, while Panel B shows stratified trajectories of men and women. This figure depicts agreement between equations improving over time as the trend lines become closer together with much more improvement in women in men. Additional file 1 shows additional smoothed population eGFR trajectories in additional subpopulations.

Figure 2 is a panelled Bland-Altman plot for each of the four age categories. Additional file 2 contains a panelled Bland-Altman plot removing the potential for repeat observations. Both show improvement in agreement across the age groups. When the analysis with repeated measures was performed, concordance is initially low at 0.74 (95% C.I. 0.67, 0.79) at the lowest age range of 10–15 but trends upwards at 0.78 (95% C.I. 0.71, 0.84) at age 15–20, 0.80 (0.70, 0.87) at ages 20–25, and 0.82 (95% C.I. 0.70, 0.90) at age 25–30. This trend across age holds within every stratum. There is a markedly higher concordance in every age range comparing females to males. Females have a value of 0.82 (95% C.I. 0.73, 0.88) at age 10–15 compared to males’ 0.69 (95% C.I. 0.59, 0.77) at age 10–15, 0.88 (95% C.I. 0.80, 0.93) compared to 0.73 (95% C.I. 0.62, 0.82) at age 15–20, 0.89 (95% C.I. 0.79, 0.94) compared to 0.76 (95% C.I. 0.59, 0.86) at age 20–25, and 0.92 (95% C.I. 0.83, 0.96) compared to 0.76 (95% C.I. 0.52, 0.89) at age 25–30, but none of the other stratification categories showed a large difference that was sustained across all age categories. For a full list of stratum-specific concordance correlation coefficient estimates including gender, race, glomerular versus non-glomerular disease and height z-score, see Additional file 3. Figure 3 presents a chart assessing the ratio between the CKD-EPI and Schwartz equations among patients at age 18 with patients divided into categories by Z-score. There is a trend towards a larger ratio (greater disparity) in stunted patients.

Table 2 summarizes the semi-partial R² estimates for specific predictors from the overall model with R² = 0.9 with approximate 95% confidence limits. Age explained most of the variance, followed by sex, then race, then height Z-score, and finally followed by type of CKD.

We explored the extent of agreement between the new Schwartz and the CKD-EPI equations for eGFR estimation in a longitudinal, racially diverse cohort of AYAs with CKD, with the intent to provide clinical recommendations for the use of these equations in patients with CKD transitioning into adulthood. We found disagreement between the Schwartz and CKD-EPI equations at all ages in this CKD cohort. The disagreement was largest in patients aged 10–15 years old, but there was better agreement as patients aged. CKD-EPI consistently estimated higher GFRs compared to the Schwartz measurements across all age groups and observed strata. In the absence of a gold standard GFR measurement, we cannot state which formula is better. Disagreement between the CKD-EPI [8] and Schwartz [7] equations was lower among female than male patients at every age, perhaps due to the lower muscle mass or less pronounced stunted growth in women. Our analyses also exhibited that age and sex are both important components of the overall variance between the two renal function estimating equations, reinforcing that both attributes play important roles in the relative performance of the two equations. The use of a sex-specific adjustment in CKD-EPI may have contributed to the fact that the equation performed more comparably to Schwartz in women than in men, though this bears further investigation. Most importantly, the switch from Schwartz to CKD-EPI eGFR estimation introduced discontinuity across our entire analysis due to discordance between both formulae.

We do not have a gold-standard measured GFR and can thus only assess agreement rather than which formula is most accurate. However, CKD-EPI and the new Schwartz [7] eGFR formulae have been well validated in their respective age groups. Due to the data’s abstraction from electronic medical health records, there may be concerns about the accuracy or reliability of these lab measures. There are also concerns with missing data on height preventing estimation of GFRs using the Schwartz equations for some of the patients at some time points, but we overcame this obstacle by imputing height values between measurements. While this is a limitation, growth is pretty linear, and the error due to height imputation can be considered negligible. Another limitation is due to the higher number of measurements in the younger age groups (see Table 1).

Overall, our study demonstrates that among AYAs with CKD, the CKD-EPI and Schwartz equations are discordant during the transition-age period and early adulthood. Selistre et al. [12] examined the difference in accuracy of the two equations across a broad spectrum of age categories using cross-sectional broad categories of age (2–12, 13–17, 18–40, 41–64, > = 65 years-old) but did not focus on the paediatric-adult transition. In Selistre’s study, a gold standard GFR measurement was performed, and he concluded that the Schwartz formula might be more accurate for young adults. Other studies assessing potential age cut-offs for kidney function equation [30] switching used measured GFRs for comparison or formulae incorporating both cystatin C and serum creatinine [31]. These studies identified similar trends in agreement as patients aged. A very recent study by Ng et al. compared creatinine and creatinine and cystatin C-based eGFR formulae in young adults with the diagnosis of paediatric CKD aged 18–26 years of age and concluded “clinicians should be aware that individually the paediatric and adult serum creatinine-based estimates of GFR had large discrepancies among emerging adults with paediatric CKD.” The group recommends taking the average of paediatric and adult serum creatinine-based formulae (i.e. the new Schwartz bedside and the CKD-EPI creatinine only formulae) as a valid tool for clinical use [32]. Of course, this approach needs further validation, but the results by Ng et al. are fully aligned with our findings. The current contribution augments the findings and suggests that this approach may be needed until age 30.

The results of this study apply to other nephrology centres and possibly other diseases where eGFR is an important biomarker in similar settings within North America and other populations with a mix of Caucasian, African American and Hispanic patients. Groups with different ethnic makeup may have differing agreement between the two equations, although our study did not show this.