Panel Estimated Glomerular Filtration Rate: Statistical Considerations for Maximizing Accuracy in Diverse Clinical Populations

Background: Accuracy of eGFR is limited by non-GFR determinants that affect creatine or cystatin C. We hypothesized that a panel eGFR that incorporates multiple filtration markers with unrelated non-GFR determinants could further improve eGFR accuracy, but methods for robust estimation that can accommodate anomalous values of some filtration markers are required for optimal accuracy. Here, we evaluated methods for estimating GFR based on multiple markers in applications with higher rates of anomalous predictors. Methods: We detected anomalous predictors in the application data using 3 approaches: univariate outliers, multivariate outliers, and inconsistencies among predictor markers. After removing anomalous predictors, we also compared 3 approaches for robust estimation: weighted trimmed mean prediction, linear model with K-nearest neighbors (KNN) marginalized estimation, and linear regression with screened predictors estimation. Using statistical simulation to emulate applications with increased anomalous predictors, we compared these approaches using Root Mean Square Error (RMSE). As proof of concept, we evaluated these methods in an initial panel of 8 markers (5 metabolites and 3 low molecular weight proteins) in 3,554 participants from 9 diverse studies. Results: Figure displays RMSE for each strategy. For example, when 10% of pseudouridine values were anomalous using the univariate and multivariable outliers perspective and applying the three robust estimation approaches, RMSEs were 0.227 - 0.276 and 0.238-0.255, respectively. However, when considering inconsistent outliers with linear regression with screened predictors estimation, the RMSE was 0.196. Results using other markers are similar. Conclusion: We found that focusing on markers with consistent predictors yields robust GFR estimates in application populations with substantial levels of anomalous predictors. Our findings demonstrate that even in cases where a subset of markers appears to be anomalous for a particular patient, accurate and unbiased estimates remain feasible. Funding: NIDDK Support