Evaluating the accuracy of the CKD-EPI equations in estimating the glomerular filtration rate among adult Africans in Malawi, Uganda, and South Africa

Abstract

Background Chronic kidney disease (CKD) is a significant global health concern with a growing burden, affecting millions of individuals worldwide. CKD was ranked 18th amongst the highest non- communicable diseases cause of death worldwide. CKD is a global health issue, which reduces kidney function and increases cardiovascular risk. Accurate glomerular filtration rate (GFR) estimation is crucial for CKD management. The commonly used CKD-EPI (Chronic Kidney Disease-Epidemiology Collaboration) equations were primarily developed for Caucasian and African American populations. This study aimed to evaluate existing equations and develop new ones specific to the African population using data from Malawi, Uganda, and South Africa. Methods This study was a secondary analysis of data collected in three African countries collectively referred to as the ARK (African Research on Kidney Disease) Consortium. The study used data from 2433 participants, with plasma iohexol clearance as an indication of GFR (mGFR). Estimating equations were developed using serum cystatin C and or in combination with serum creatinine, mirroring the CKD-EPI equations by adopting a non-linear modelling approach. The study developed a predictive model using supervised machine learning techniques. Bland-Altman plots were used to assess linearity and agreement of the eGFR methods. Accuracy within 10% and 30% of mGFR, bias, and precision were assessed overall and by CKD stage. Results iv Analysis of 2433 participants from the three African countries revealed significant differences in mean measured glomerular filtration rate (mGFR) by country and sex. New serum cystatin C and creatinine-cystatin C-based equations for estimating GFR were developed, showing high accuracy ranging between (94-95%) and (93-95%), respectively for GFR ≥90 ml/min/1.73m2. The equations however had lower precision (2.07 – 2.12) compared to existing ARKM (African Research on Kidney Disease Model) equations (2.36 – 2.37). Six machine learning (ML) classification models were evaluated, with Random Forest emerging as the top performer, followed by Logistic Regression. ML approaches demonstrated higher F1 score measures (89%-100%) than eGFR equations, accuracies ranging between 75% and 95% and less bias. Overall, it was concluded for this study that ML techniques provide better performing models in comparison to the existing and developed eGFR models. This was evident as AUC measures for all ML models were higher (93% - 100%) than the accuracy measures of the eGFR equations (75% - 95%). Conclusion The results highlight the value of cystatin C as a biomarker for improving GFR estimation and underscore the importance of population-specific GFR estimation tools for African populations. While no single method is perfect across all GFR levels, the findings demonstrate the potential of both refined eGFR equations and machine learning models in enhancing GFR estimation accuracy. However, confirmation in broader populations is needed, and regular monitoring and adaptation of ML models will be required to maintain predictive performance over time. These findings can inform efforts to improve GFR estimation and CKD evaluation in Africa.