Master's Papers
Permanent URI for this collectionhttps://hdl.handle.net/10161/34322
Browse
Recent Submissions
Item type: Item , Access status: Open Access , Implementation and Evaluation of an AI-Based Prior Authorization Assistant in a Large Academic Health System(2026) Nikkie DuttaAbstract: Implementation and Evaluation of an AI-Based Prior Authorization Assistant in a Large Academic Health System Introduction: Prior Authorization (PA) refers to the process used by payers to ensure that proposed care, including treatments, procedures, and prescriptions, is medically necessary. While PA is well intended, the process is administratively burdensome, has high costs, contributes to provider burden, and prevents timely access to treatment. As a result, health systems such as the Duke University Health System (DUHS) are actively exploring solutions driven by AI to reduce PA-related bottlenecks. Objective: To address these challenges, Joanna Kipnes, Medical Director of Utilization Management, Patient Revenue Management Office (PRMO), Deborah Kaye, Assistant Professor, Department of Urology, and the PRMO team, with support from Duke University Health System and in collaboration with Duke Institute for Health Innovation, developed an AI tool to automate chart review and generate complete PA submissions and appeals. Methods: A Medical Prior Authorization Assistant tool, accessed through a web-based interface, was developed for each medication in order to automate the evaluation and preparation of PA submissions. The tool extracts real-time patient data including clinical notes, flowsheets, laboratory results, medication orders, and active medications, from Maestro Care (Epic), which is Duke Health’s unified electronic medical record (EMR) and clinical care application. Results: Currently, DIHI has gone live with 2 of the Phase 1 medications, Ocrevus and Botox for Migraine. The evaluation for Ocrevus has been completed, with the LLM achieving a perfect score of 100% across 30 retrospective PA cases, each assessed on 12 criteria. Evaluation for Botox achieved a score of 99%, where 31 cases were evaluated across 14 criteria. Discussion: Teams at DUHS are accustomed to processing PAs using established workflows, and transitioning to a new approach involved learning how the tool could complement their existing processes. Continued engagement with the teams, a collaborative approach to feedback, and a willingness to refine the tool based on user input have helped build trust and improve adoption of the tool. Conclusion: This project developed and evaluated an AI-based Medical Prior Authorization Assistant designed to automate chart review and generate documentation for PA submissions within DUHS. Initial evaluations demonstrated high accuracy (100% for Ocrevus and 99% for Botox for Migraine), suggesting that AI tools have strong potential to reduce administrative burden and improve the efficiency of the PA process.