Serum Uric Acid and Blood Urea Nitrogen as Primary Determinants of eGFR in 23,960 Chinese Adults: A Cross-Sectional Study with Machine Learning and Longitudinal Validation.

Abstract

Background: Chronic kidney disease (CKD) affects approximately 850 million individuals worldwide. Machine learning (ML) approaches offer potential advantages over traditional regression for identifying eGFR determinants, but prior studies frequently include serum creatinine or cystatin C in predictor sets, creating tautological prediction bias. We aimed to identify eGFR determinants using ML with methodological safeguards against data leakage in a large Chinese population.

Methods: We conducted a population-based cross-sectional study of 23,960 unique adult participants (first visit per patient) recruited across five annual waves (2011-2015) in Eastern China, with a longitudinal validation subset of 14,309 patients with two visits. eGFR was calculated using the CKD-EPI 2021 race-free equation. LightGBM and traditional regression were applied to 35 clinical and laboratory predictors, explicitly excluding serum creatinine and cystatin C to prevent tautological prediction. Patient-level data splitting (70/15/15) ensured no individual appeared in both training and test sets.

Results: CKD (eGFR <60 mL/min/1.73m²) was present in 2.3% (n=554) of participants; mean eGFR was 91.9±15.3 mL/min/1.73m². LightGBM achieved R²=0.461 (RMSE=11.63 mL/min/1.73m²) compared to linear regression R²=0.450. Feature importance analysis identified serum uric acid (SUA, 10.1%), blood urea nitrogen (BUN, 9.6%), and age (7.2%) as primary eGFR determinants. Multivariable regression confirmed age (β=-0.809 per year, 95% CI: -0.843 to -0.774) and SUA (β=-0.064 per µmol/L, 95% CI: -0.069 to -0.060) as independent predictors (both p<0.001). In the longitudinal subset, mean eGFR declined by 4.44±12.50 mL/min/1.73m² between visits. CKD prevalence increased from 2.5% (2011) to 3.6% (2015; p<0.001).

Conclusions: Using a methodologically rigorous approach that excludes GFR-derived metabolites from predictor sets, SUA and BUN are identified as primary eGFR determinants in a large Chinese population. The modest ML advantage over traditional regression (ΔR²≈0.011) reflects capture of non-linear interactions. These findings support targeted screening of individuals with elevated uric acid and BUN for CKD prevention.