End-stage renal disease (ESRD) remains a major complication of diabetes, yet existing static risk scores may lose accuracy as patient profiles evolve and competing mortality risks change over time. We developed and externally validated a landmark-based ESRD Dynamic Risk Score (ESRD-DRS) that updates individual risk estimates at 1, 5, and 10 years after diabetes diagnosis using routinely collected EHR data. We assembled a retrospective cohort of 708,435 U.S. Veterans with newly diagnosed diabetes in the Veterans Health Administration (VHA). At 1, 5, and 10 years after diabetes diagnosis (LM1, LM5, and LM10), we fit penalized Fine-Gray subdistribution hazard models, drawing from more than 400 demographic, medication, comorbidity, and laboratory variables. Models were evaluated over 1-, 5-, and 10-year horizons for discrimination (area under the time-dependent receiver operating characteristic curve AUROC) and calibration (Brier score), and compared with the established RECODe and 4-variable KFRE risk equations. External validation was performed in an independent All of Us (AoU) cohort (n = 13,223). In the VHA cohort (median follow-up 7.6 years), 8,955 patients (1.26%) developed ESRD, and 136,666 (19.3%) died without ESRD. At LM1, ESRD-DRS achieved AUROCs of 0.93, 0.90, and 0.85 for 1-, 5-, and 10-year risk, respectively, and Brier scores ranging from 0.00098 to 0.0162. In the AoU cohort, corresponding AUROCs were 0.94, 0.91, and 0.86, with similar calibration performance. RECODe and KFRE yielded lower discrimination and poorer calibration. Top risk predictors, including estimated glomerular filtration rate, albuminuria, systolic blood pressure, and age, were consistent across landmarks and cohorts. ESRD-DRS, a scalable landmark approach that accounts for competing mortality and evolving patient profiles, outperformed existing static equations. Embedding ESRD-DRS into EHR workflows may support more timely, individualized ESRD risk assessment in patients with diabetes.
Jensen et al. (Mon,) studied this question.