Assessing the impact of race on glomerular filtration rate (GFR) prediction

Background The appropriate use and the implications of using variables that attempt to encode a patient's race in clinical predictive algorithms remain unclear. The clinical algorithm for estimating glomerular filtration rate (GFR) adjusts for race, but the observed difference between Black and non-Black participants lacks biologically substantiated evidence. We investigated the impact of using a race variable on GFR prediction by race-stratified error analysis. Methods We implemented three predictive algorithms with varied amount of input information from an electronic health record database to estimate GFR. We compared the prediction error of the estimated GFR with and without the variable race between Black patients and White patients. Results The prediction error for patients coded as Black was higher than that for patients coded as White across all three algorithms. Removing race from the prediction algorithm did not lower the prediction error for patients coded as Black, neither did it decrease the difference in error between the two groups. The algorithm that included the most information with thousands of variables but excluding race produced the most accurate estimate for both groups and minimized the difference in performance between the two groups. Conclusion The prediction error for patients coded as Black was higher compared to those coded as White, regardless of inclusion of race as a variable. Using a large amount of information represented in electronic health record variables achieved a more accurate prediction of GFR and the least difference in prediction error across racial groups.