Uncertainty-aware prediction of 48-month eGFR decline in type 2 diabetes mellitus: a secondary analysis of ACCORD

This is the authors' abstract. We don't add key points for this paper.

Long-horizon kidney trajectory models in type 2 diabetes mellitus (T2DM) commonly report point estimates or event risks, leaving uncertainty around predicted eGFR change less clear. We developed and internally evaluated a baseline clinical model with conformal intervals for annualized 48-month eGFR change. We analyzed 6,853 adults with T2DM and high cardiovascular risk from the multicenter Action to Control Cardiovascular Risk in Diabetes (ACCORD) trial with baseline and 48-month eGFR and at least three measurements. We compared penalized linear, random forest, histogram gradient boosting, and XGBoost models with mean-outcome and clinical linear benchmarks. The primary model included baseline eGFR, demographics, cardiometabolic measures, medications, and history. Five-fold training-set cross-validation selected the model, while separate calibration and test sets evaluated point accuracy and calibration, rapid-decline discrimination, and split and locally adaptive 90% interval coverage and width. In 1,371 test participants, random forest mean absolute error was 3.370 mL/min/1.73 m² per year (95% CI 3.221 to 3.527) and R² was 0.359 (95% CI 0.233 to 0.466). Locally adaptive intervals achieved coverage of 0.909 (95% CI 0.894 to 0.923) and mean width of 14.706 mL/min/1.73 m² per year. Rapid decline ( ≤ − 5 mL/min/1.73 m² per year) occurred in 467 (34.1%), with AUC of 0.732 (95% CI 0.708 to 0.758). Unadjusted Q5 had higher error and rapid-decline prevalence than Q1. After width quintiles were assigned within 20 baseline-eGFR strata, Q5 had higher HbA1c (8.6% vs. 8.2%), triglycerides (253.3 vs. 177.5 mg/dL), and female proportion (49.6% vs. 30.7%) than Q1 despite similar baseline eGFR. Conformal intervals achieved near-nominal marginal coverage but remained broad. Conditioned width differences across HbA1c, triglycerides, and sex revealed variation in modeled uncertainty beyond baseline eGFR. These findings make forecast uncertainty explicit and provide a foundation for external, decision-focused evaluation. Not applicable.