Brief
Prior authorization (PA) requests create a significant administrative burden at Duke Health, contributing to staff and physician burnout, delays in care, and increased costs for the health system. To address this challenge, our team developed an AI-based solution that automates chart review to surface relevant patient data and prepare prior-authorization requests or denial appeals. The tool is currently being implemented and evaluated in real-time, with metrics including time to complete PA requests, rate of medication denials, appeal turnaround time, staff administrative time, and overall impact on patient care and outcomes.
Problem
Prior authorization (PA) is utilized by insurers to assess whether a proposed medication or medical service is medically necessary,1 requiring approval from a patient’s health plan before a patient can access certain healthcare goods and services, including medications, surgeries, radiology, and laboratory testing. If PA is not obtained, insurers can deny coverage of the medication or medical service. However, the current process is manual and inefficient as providers are often unaware of payer-specific criteria, multiple staff must compile and submit documentation, and relevant information is not consistently highlighted or complete. Additionally, high-cost medications often must be paid for up-front by the health system at Duke, with reimbursement contingent on prior authorization approval. When requests are denied, this can lead to significant financial losses and delays in patient care.
Solution
To address these challenges, Joanna Kipnes, Deborah Kaye, the PRMO team, and the Duke Institute for Health Innovation developed an AI platform to automate chart review and draft prior authorization submissions and denial appeals. Phase 1 of the project focuses on medications with high average monthly volume, denials, and treatment costs, and include Ocrevus, Botox for Migraine, Pluvicto, and Darzalex Faspro.
For each medication, we created a Medical Prior Authorization Assistant (MPAA) tool that retrieves patient data from Maestro Care (e.g., provider notes, flowsheets, labs, orders, medications, and ICD codes) in real-time and combines it with payer-specific prior authorization and medical necessity criteria. See Figure 1 for the MPAA clinical workflow for Ocrevus. On the solution’s user interface, users enter the patient’s MRN. The current date auto-populates to support authorization submissions for patients’ upcoming medication infusions. It can be adjusted to support denial appeal workflows. The user then selects “search” to pull in the relevant chart information and “Evaluate the patient for Ocrevus Prior Authorization”. The patient’s chart is then evaluated by a Large Language Model (LLM) to determine whether the patient meets criteria for the requested medication therapy. When all authorization criteria are met, the solution generates a documentation checklist for user to follow to select the relevant records in the release-of-information (ROI) workflow. It also drafts a cover letter that incorporates supporting clinical information for submission or appeal. If the criteria is not met, the solution tells the user to stop and lists the criteria that were met as well as those that were not met.
To validate the solution, we evaluated retrospective prior authorization cases for all four medications, including approvals, denials, and supporting data. From these cases, we created “ground truth” files that included the authorization status and relevant supporting data. These ground truth cases were then compared to outputs generated by the LLM. Each individual criterion was marked as a pass if the LLM output matched the ground truth. If a criterion generated by the LLM did not match ground truth, we reviewed the case, refined the prompt and augmented the raw source data as needed, and re-tested to confirm a subsequent match between LLM and ground truth.
Outcomes
The evaluations for Ocrevus, Botox for Migraine, and Pluvicto have all been completed, with scores of 100%, 99%, and 98% respectively (see Table 1). These scores reflect the LLM’s accuracy in matching the output of the ground truth files. Notably, the LLM correctly assessed the authorization status for all 83 cases across the three medications, with rare permutations from ground truth related to supporting documentation (7 errors total out of 970 criteria assessed). The Medical Prior Authorization Assistant tool for Ocrevus went live on January 12, 2026 with the Medical Necessity & Denials Team. The Medical Prior Authorization Assistant tool for Botox went live on February 23rd with the PSR Team and financial care counseling teams.
Next Steps (as of April 2026)
We plan to go live in May 2026 with the Outpatient Pre-Service Verification (PSV) teams for Ocrevus, Pluvicto, and Darzalex Faspro. In Phase 2, we are currently refining the prompt and evaluating the tool to authorize GLP 1 inhibitors, Pembrolizumab, and Romidepsin and go live with these medications in Summer 2026. The long-term goal is to continue testing, refining, and scaling the solution across Duke Health to support prior authorization for a broader range of services, including radiology, procedures, laboratory testing, and durable medical equipment.
Access the solution (requires DH authorization)
Reference:
[1] Gupta R, Fein J, Newhouse J P, Schwartz A L. Comparison of prior authorization across insurers: cross sectional evidence from Medicare Advantage BMJ 2024; 384 :e077797 doi:10.1136/bmj-2023-077797
