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

Researchers developed a model to predict kidney function decline (eGFR) over 48 months in patients with type 2 diabetes mellitus using data from the ACCORD study with 6,853 participants. The model used variables such as age, sex, blood sugar level (HbA1c), triglycerides, and blood pressure, but excluded creatinine and urine biomarkers. The best model was random forest with accuracy R2=0.382. Confidence intervals achieved 91.7 percent coverage. Wider intervals were associated with larger prediction errors and faster kidney function decline. Wider intervals occurred in younger patients, women, and individuals with higher HbA1c, triglycerides, and blood pressure. The results show that the model can reliably predict long-term kidney function trajectory and interval width serves as an indicator of prediction reliability.