| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Letters to the Editor |
R1 Department of Clinical Chemistry and R2 Competence Centre, for Clinical Research, University Hospital, Lund, Sweden
aAddress correspondence to this author at: Department of Clinical Chemistry, University Hospital, S-22185 Lund, Sweden. Fax 46-46130064; e-mail anders.grubb{at}klinkem.lu.se.
To the Editor:
Although prediction equations for glomerular filtration rate (GFR), whether based on cystatin C or creatinine, provide superior diagnostic specificity and sensitivity (1)(2)(3)(4) in the diagnosis of renal disease, their general implementation in the clinical routine will not be without obstacles. Indeed, many of the previously discussed problems for creatinine-based prediction equations (5) also apply to the general implementation of cystatin C–based prediction equations. Twomey and Kroll raise important questions related to these implementation problems.
As they point out, the Modification of Diet in Renal Disease (MDRD) prediction equation applies only to persons above 17 years of age (1)(2)(6). Our logic in constructing Fig. 1, panels A–C, as shown in our article (2), was to demonstrate that the cystatin C–based GFR prediction equations can be used for both children and adults. The scales of the y axes were selected to accommodate all of the prediction errors of the cystatin C–based GFR estimations and most of the prediction errors of the MDRD-based GFR estimations. To avoid overextending the scales of the y axes, the most extreme prediction errors produced by the MDRD equation were given as numbers in Fig. 1C. Moreover, Tables 2 and 3 of our article report separate data for adults and children. Table 4 displays data for the combined group of adults and children, as the cystatin C–based prediction equation is useful for all ages.
It is true that the population we studied (2) is different from the one studied when the MDRD equation was produced (6). Whereas the population used to derive the MDRD prediction equation comprised only persons with chronic kidney disease, the population in our study comprised everyone for which an invasive determination of GFR was requested and thus included a high proportion of healthy persons. Previous investigations (7) have shown that the composition of the patient cohort influences the derived prediction equations, and further research must be performed to compare the diagnostic performances of cystatin C–based prediction equations and the MDRD prediction equation in identical populations. It might be observed, however, that the population used in our study is a very general one, with the only inclusion criterion being that a physician wanted to know the GFR of a patient.
Establishing a statistical prediction model always involves a trade-off between model complexity and model fit. The overall relationship between relative GFR and cystatin C in our data set, when both are measured on the log scale, is approximately linear. As Twomey and Kroll highlight, however, there might be some departure from the overall linearity among patients with high GFR (Fig. 3A in the article). Among 32 patients with GFR >110 mL . min–1. (1.73 m2)–1, underestimation occurred for 26, yielding a median percentage error of –16%. The accuracy of the predictions within this upper end of GFR is somewhat lower than that of the whole GFR range, with 80% of the GFR estimates within 30% of measured GFR. The diagnostic performance of a prediction equation for a specific GFR range is generally influenced by the relative number of data points in this area. In our adult population, a large number of the patients (189 of 451) had a GFR of 50 to 90 mL . min–1. (1.73 m2)–1, and it is in this clinically relevant area that the diagnostic performance of the cystatin C–based prediction equation is best, with 89% of the GFR estimates within 30% of measured GFR.
We agree completely with Twomey and Kroll that currently described GFR prediction equations cannot replace gold standard procedures for measuring GFR, but such equations might reduce the need to perform invasive GFR determinations and might permit a more precise selection of patients requiring such determinations.
We also agree with Twomey and Kroll that it is crucial to use the best models to generate GFR prediction equations with maximum diagnostic performance, and we welcome research in this important area. We would be happy to supply them with our data in such efforts.
We also share the view of Twomey and Kroll that the diagnostic performance of a derived prediction equation can be overestimated if it is evaluated only in the derivation set. In our study, we took this into account with an iterative cross-validation procedure, in which 
45 randomly selected patients were kept isolated from the derivation set and then used for validation in each iteration. The 100 generated test sets still showed overall high diagnostic performance of the cystatin C–based GFR prediction equation for all ages (median of 81% of the predicted GFR values were within 30% of measured GFR, and 96% were within 50%; as noted in the Cross-Validation section of our article).
References
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
