Effect of removing race correction factor in glomerular filtration rate estimation on predicting acute kidney injury after percutaneous coronary intervention

BackgroundRace is a social and not a biological construct. Major societies have recommended removing a race correction factor when estimating glomerular filtration rate (eGFR). eGFR is a strong predictor for acute kidney injury (AKI) after percutaneous coronary intervention (PCI). We assessed the effect of removal of the race correction factor in eGFR calculation on the predictive ability of contemporary models for AKI after PCI. MethodsWe used data from the American College of Cardiologys National Cardiovascular Data Registry (NCDR) CathPCI registry to assess the effect of removing a race correction in eGFR calculation on two previously published AKI models developed using NCDR data - a logistic regression-based model and a gradient boosting machine learning (ML) model. We first assessed the calibration and discrimination of these models with and without race correction and subsequently included race as an independent predictor in a model without race-corrected eGFR. We assessed model performance overall, and stratified by Black and non-Black subgroups. The models were trained on the same cohort used to develop the NCDR models and validated with a contemporaneous validation data set. ResultsWe included 947,091 PCI procedures in 915,223 patients (mean age 64.8 years; 7.9% were Black and 32.8% women) with an AKI rate of 7.4%. In the NCDR model, inclusion of race correction in eGFR significantly underestimated AKI risk among Black patients (predicted 7.6% vs observed 10.2%) while slightly overestimating risk among non-Black patients (predicted 7.4% vs observed 7.1%). Removing the race correction partially corrected the underestimation among Black patients (predicted 8.2%). Including race as an independent predictor and introducing interaction terms further reduced the underestimation (predicted 10.1%). The receiver-operating-characteristic curve (AUC) was similar among these models, but was consistently lower in Black patients. Compared with the logistic model, the ML model had better calibration in Black patients (with race correction predicted 8.6% and without 8.7%) but still underestimated AKI risk. With race included as a predictor, the ML model achieved similarly good calibration in Black patients (predicted 10.1%). The AUCs of ML models were better than the logistic models but did not differ based on the inclusion of the race correction. ConclusionRemoving the race correction in eGFR calculation has a positive effect of reducing the underestimation of the risk of AKI following PCI for Black patients. However, despite the reduction in underestimation, Black patients remain at elevated risk for AKI compared to non-Black patients. There is a need to better capture the determinants of this higher AKI risk through richer data and advanced modeling techniques.