Large Language Model Shows Promise for Nephrology Prior Authorization Letter Drafting
Key Clinical Summary
- In 29 nephrology prior authorization scenarios, ChatGPT-5 generated letters with strong clinical reasoning in 89.7% of cases and valid references in 93.1%.
- Coding reliability was lower: 79.3% of letters used correct International Classification of Diseases, 10th Revision (ICD-10) codes, with most errors involving chronic kidney disease staging or internal diagnostic inconsistencies.
- Investigators concluded that artificial intelligence (AI)-assisted drafting may help reduce administrative burden, but clinician oversight remains necessary to verify coding, citations, and safety details.
Prior authorization remains a major administrative challenge in nephrology, contributing to treatment delays, clinician burden, and potential barriers to patient access. A new evaluation of ChatGPT-5 suggests that large language models may be able to support prior authorization documentation in nephrology, although coding errors and citation mismatches still require clinician review before submission.
Study Findings
Investigators developed 29 standardized nephrology scenarios involving medications that commonly require prior authorization and used a single prompt to generate letters with ChatGPT-5. Each letter was assessed for false statements, correctness of ICD-10 coding, validity of citations, and strength of clinical reasoning. Reference standards included US Food and Drug Administration drug labels, Kidney Disease: Improving Global Outcomes (KDIGO) guidelines, and related randomized controlled trials.
Overall performance was favorable. Only 1 of 29 letters, or 3.5%, contained a false statement. That error occurred in a velphoro scenario, in which the model cited the unrelated INNO2VATE vadadustat trial. ICD-10 coding was correct in 23 letters, or 79.3%, with most inaccuracies involving chronic kidney disease stage 3a vs stage 3b assignments. Reference validity was high, with 27 of 29 letters, or 93.1%, citing appropriate sources.
Clinical reasoning was the strongest domain. Twenty-six letters, or 89.7%, were rated as strong, while the remaining 3 were rated as adequate. The letters rated as adequate were not clinically wrong, but they omitted details such as gastrointestinal tolerability for tenapanor, stronger evidence support for nedosiran, or meningococcal vaccination and prophylaxis considerations for eculizumab.
The authors note that nearly 90% of letters offered patient-specific, guideline-aligned justification, suggesting that ChatGPT-5 can reproduce the structure and logic of payer-facing medical necessity documentation in many nephrology scenarios.
Clinical Implications
For managed care stakeholders, the findings are notable because prior authorization consumes substantial physician time and can delay therapy initiation. The study cites estimates that physicians spend 12 to 13 hours per week on prior authorization and submit about 39 requests weekly. If AI can draft clinically coherent letters, health systems and practices may be able to streamline nephrology workflows and reduce documentation burden.
However, the study also shows why unsupervised deployment would be risky. Even infrequent factual errors, incorrect trial references, or imprecise diagnostic codes could undermine payer confidence or lead to denials. In nephrology, where coverage often depends on disease stage, prior treatment history, and precise indication, coding accuracy is especially important.
From a payer and operations perspective, the most practical use case may be assisted drafting rather than autonomous submission. Embedding automated checks for diagnosis codes, citation verification, and boxed-warning or risk evaluation and mitigation strategy requirements could make AI-assisted prior authorization more reliable and scalable.
Conclusion
ChatGPT-5 performed well in drafting nephrology prior authorization letters, particularly for clinical reasoning and citation use, but important weaknesses remain in diagnostic coding and safety detail inclusion. For managed care and nephrology practices, the technology appears promising as a supervised workflow tool rather than a stand-alone solution.
Reference
Aiumtrakul N, Thongprayoon C, Kookanok C, Poochanasri M, Phichedwanichskul K, Cheungpasitporn W. Quality assessment of large language model-generated prior authorization letters in nephrology. Front Digit Health. 2026;3(8):1767648. doi: 10.3389/fdgth.2026.1767648
