A study presented at the American Society of Nephrology Kidney Week 2026 in Denver, Colorado, found that identifying high-risk patients with chronic kidney disease in electronic medical records and providing quarterly feedback to clinicians led to increased adoption of guideline-directed medical therapy and albuminuria testing. The research, funded by Klinrisk and AstraZeneca, utilized the Klinrisk model to predict disease progression in a large US nephrology practice with over 100 clinicians and 50,000 patients.
Risk Scores and GDMT Uptake
Patients with a 5-year risk of progressive CKD—a 40% decline in estimated glomerular filtration rate or kidney failure—of at least 25% were flagged as high risk. Clinicians received quarterly reports benchmarking their urine albumin-creatinine ratio (UACR) testing and GDMT use against practice-wide rates. The analysis included 51,097 patients at baseline and 54,257 at follow-up. Among high-risk patients, use of every GDMT class increased after implementation. SGLT2 inhibitor use rose from 30.7% to 37.9%, glucagon-like peptide-1 receptor agonist use from 14.5% to 18.2%, mineralocorticoid receptor antagonist use from 13.5% to 15.9%, and renin-angiotensin-aldosterone system inhibitor use from 61.2% to 63.5%. UACR testing across all patients increased from 43.0% to 50.7%.
The biggest prescription increases for SGLT2 inhibitors, mineralocorticoid blockers, and RAAS inhibitors occurred among those deemed highest risk. Researchers attributed the gains to the risk-scoring system and clinician feedback, noting stronger adoption among patients flagged by the Klinrisk tool. The results imply that prioritizing treatment based on risk could improve outcomes for those who stand to benefit most, offering a practical way to deliver more effective kidney care.
While the Klinrisk model showed clear results in a single practice, its effectiveness may vary widely across different health systems. A retrospective cohort study from the University of Connecticut examined a different approach to managing advanced kidney disease. Researchers looked at the first two years of the Comprehensive Kidney Care Contracting option under the CMS Kidney Care Choices Model.
Payment Models and Outcomes
Under the CKCC option, kidney contracting entities that include nephrologists and transplant providers take on varying levels of financial risk for total cost of care. The study compared Medicare fee-for-service beneficiaries in CKCC practices with geographically matched beneficiaries outside the model over eight quarters in 2022 and 2023 across nine outcomes in patients with stage 4 to 5 CKD and those with dialysis-dependent end-stage renal disease. In difference-in-differences analysis, CKCC practices showed a statistically significant differential quarterly improvement in permanent vascular access at hemodialysis initiation. Preemptive kidney transplant rates increased significantly over time among CKCC-aligned beneficiaries with CKD but not among comparators.
CKCC-aligned beneficiaries with stage 4 to 5 CKD had lower costs, with a mean difference of $3,876 per year. The author, Ohm S. Tripathi, wrote that the findings support extending voluntary, risk-bearing kidney care payment models and identify permanent vascular access planning as a priority for broader dissemination. The analysis differs from CMS’ independent evaluation of the full Kidney Care Choices Model, which found no statistically significant effect on total Medicare Parts A and B payments in 2023 and an increase of $304.8 million in net Medicare spending.
AI Tools for Prior Authorization
Mayo Clinic researchers examined two large language models for drafting prior authorization letters in nephrology. The team used OpenEvidence and ChatGPT-5 across 28 clinical scenarios. Writing these letters takes time and delays treatment, adding to clinician burden. The study found false statements in 7.1 percent of OpenEvidence letters and 3.6 percent of ChatGPT-5 letters. The difference was not statistically significant. ICD-10 coding was correct in 67.9 percent of OpenEvidence letters and 78.6 percent of ChatGPT-5 letters.
Both tools produced valid references in 92.9 percent of the drafts. Clinical reasoning was rated strong in 78.6 percent of OpenEvidence letters and 89.3 percent of ChatGPT-5 letters. OpenEvidence generated longer letters that often exceeded the 350-word target. The authors linked this length to more chances for inaccuracies. ChatGPT-5 produced concise drafts with fewer high-impact mistakes. Both tools generated factual errors, known as hallucinations. The authors concluded that LLMs could generate clinically usable letters, but the error rates were similar across models.