Abstract / Summary
Abstract Background/Objective Albuminuria and estimated glomerular filtration rate (eGFR) provide complementary information about kidney disease risk and adverse outcomes, including mortality. Using pooled National Health and Nutrition Examination Survey (NHANES) 2007–2018 data, we evaluated the association between albuminuria and all-cause mortality among adults with preserved eGFR and compared conventional and machine-learning survival approaches for mortality prediction. Methods We analyzed pooled NHANES 2007–2018 data with linked mortality follow-up. The primary cohort was restricted to adults with eGFR ≥ 60 mL/min/1.73 m²; albuminuria was defined as urinary albumin-to-creatinine ratio (UACR) ≥ 30 mg/g, with UACR < 30 mg/g as the reference. Survey-weighted Cox regression estimated the association with all-cause mortality after multivariable adjustment. Standard Cox, least absolute shrinkage and selection operator (LASSO) Cox, Elastic Net Cox, and Random Survival Forest models were evaluated for mortality prediction in a held-out test set. Results Among 29,918 adults with preserved eGFR, 3,090 had albuminuria and 1,943 died during follow-up. Median potential follow-up was 86 months (95% CI, 84–87). Compared with UACR < 30 mg/g, albuminuria was associated with higher all-cause mortality in the fully adjusted survey-weighted model (HR, 2.10; 95% CI, 1.82–2.42). In the held-out test set, C-indices were 0.8400 for standard Cox, 0.8392 for LASSO Cox, 0.8393 for Elastic Net Cox, and 0.8407 for Random Survival Forest; paired bootstrap analysis showed no meaningful predictive advantage of Random Survival Forest over standard Cox. Conclusions Among adults with preserved eGFR, albuminuria was associated with higher all-cause mortality after multivariable survey-weighted adjustment. More complex machine-learning survival models did not show a meaningful improvement in mortality prediction beyond standard Cox regression.