The study protocol was registered in PROSPERO (CRD42019123434). We performed a systematic review following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines (18).
Literature search and study selection
In this systematic review, we included original observational studies that were performed in Latin American countries and compared both CKD-EPI and MDRD equations with mGFR (the gold standard, that used any exogenous filtration markers such as inulin, iohexol, iothalamate, 51Cr-EDTA, DTPA, among others) in adult populations (≥ 18 years). We did not exclude any study on the basis of language or any other criteria.
We carried out a literature search in PubMed and Scopus in January 2019, and in “Biblioteca Regional de Medicina” (BIREME) in February 2019. The search strategy for each database or virtual library is shown in Additional file 1.
Duplicated records were removed using the EndNote software. Later, two researchers (ABC and NBC) independently selected abstracts for full-text review and final inclusion, with any differences resolved by a third researcher (JHZT).
Also, we searched the lists of references of all included studies, and the lists of articles that cited each of the included studies (through Google Scholar) to identify other studies that fulfilled the inclusion criteria.
Data extraction
Two researchers (ABC and NBC) independently extracted data from each article that met the inclusion criteria using a standardized Microsoft Excel sheet, with any differences resolved by a third researcher (JHZT).
The following variables were extracted from each study: first author, year of publication, country, design (prospective or retrospective), population characteristics (inclusion and exclusion criteria, number of participants, sex, age, ethnic group, CKD diagnosis, and CKD etiology), intervention (type of MDRD and CKD-EPI equations), gold standard (exogenous filtration marker), mGFR, eGFR, and numerical results of diagnostic measures.
Main diagnostic measures were bias (defined as the mean of the difference between eGFR and mGFR), P30 (percentage of results of eGFR that did not deviate more than 30% from mGFR), and accuracy measurements (sensitivity, specificity, and area under the curve).
Other measures included: precision (defined as one standard deviation of bias, or as interquartile range), bias% (mean of the difference between eGFR and mGFR, in function of mGFR), P15, P10, combined root mean square error (CRMSE), Pearson coefficient, intraclass correlation coefficient, Kappa coefficient, and limits of agreement (defined as bias ± 2 standard deviations).
When there were doubts about some information reported in the studies, we sent an email to the authors in order to clarify the information.
Risk of bias and certainty of evidence
Two researchers (NBC and VEFR) assessed the four risk of bias domains of the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool (19): patient selection, index test, reference standard, and flow and timing. In case of disagreement, a consensus was achieved with a third researcher (JHZT).
We used the GRADE methodology (20) to report our certainty in the evidence of accuracy of diagnostic tests results. To show this certainty, we created tables of Summary of Findings (SoF) according to the GRADE specifications (21).
Statistical analyses
When possible, we performed meta-analyses of P30, bias, sensitivity, and specificity (when studies compared similar equations, showed their confidence intervals or standard deviations, or allowed to calculate these values).
For P30 and bias, we calculated mean differences (MD) and their 95% confidence intervals (95% CI). For sensitivity and specificity, we built a 2 × 2 table when possible. As there were less than four studies to meta-analyze, we could not perform a meta-analytical hierarchical regression for diagnostic accuracy, so we performed a meta-analysis of proportions using the exact binomial distribution. We assessed heterogeneity using an I² statistic and used random-effects models when I² was higher than 40%.
For bias and P30, we performed a subgroup analysis according to the presence of CKD (using the cut-off of 60 ml/min/1.73 m²) since a previous systematic review showed that the eGFR equations performance varies across these subgroups (22). We could not perform a subgroup analysis for comorbidities since there was no more than one study that assessed the same version of the equation in any of the comorbidities groups. The data were processed using the Stata v14.0 software.