Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2541 - Alleviating the Authorization Burden: Physician Testing of a Generative AI-Augmented Appeal Tool
Presenter(s)
Jonathan Massachi, MD, MS - University of California, Los Angeles, Los Angeles, CA
J. Massachi1, F. Velazco1, and M. L. Steinberg2; 1Department of Radiation Oncology, University of California, Los Angeles, Los Angeles, CA, 2Department of Radiation Oncology, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA
Purpose/Objective(s):
Documentation and insurance authorization are a significant burden on healthcare providers. Time that could be spent on patient care is increasingly being consumed by the need for peer-to-peer review or appealing denials of medically necessary treatments. This is especially true in radiation oncology given the high rate of denied claims. Furthermore, with the rise in artificial intelligence (AI) technologies, several insurers have begun utilizing these tools to review and flag claims for further review or potential denial. Conversely, large language models (LLMs) present an opportunity to relieve some of the cognitive load and time burden on clinicians by leveraging their key strengths in handling formulaic documentation tasks. They are, however, limited by concerns for data security and risk of hallucination. Here, we present results of testing an in-house developed secure web platform augmented by generative AI to produce insurance appeal letters in simulated real-world conditions.Materials/Methods:
Briefly, the platform consists of a browser-based graphical user interface (GUI) that takes clinical details as input and utilizes generative AI to produce an appeal letter personalized to the specific patient’s case. Protected health information (PHI), which is not transmitted from the user’s computer, is segregated from general clinical information used to query an LLM to produce a letter template. Users can then edit the text and download the completed letter on institutional letterhead ready to be signed and submitted with a single click. Physicians were presented with a clinical vignette and treatment plan along with the insurer’s reason for denial and asked to write an appeal letter using their usual methods first. They were then asked to use the experimental tool to complete the same task. Time to complete each task, physician’s perception of quality, accuracy, and completeness of the AI-generated appeal letter, and ease of use of the tool were all recorded. A 5-point Likert scale was used for qualitative variables.Results:
In total, 7 physicians tested the tool. Mean time to complete the task was much shorter with the tool than conventional methods (7 minutes vs 23 minutes, p < 0.01). All physicians agreed that the quality of the AI-generated appeal letter was much better than the letter they had written on their own. All physicians rated the content of the appeal letter as both accurate and complete (at least 4/5 on the Likert scale).Conclusion:
Overall, this work demonstrates that LLMs can provide useful time savings and augment clinicians’ capabilities in routine clinical documentation when appropriate safeguards for data security and content review are in place. In aggregate for a health system, these time-saving support efforts may result in better quality appeal letters, more provider availability for patient-facing clinical care, and decreased clinician burnout.