3382 - Hype or Reality: Utility of Large Language Models in Radiation Oncology

B. S. Jang¹, S. R. Alcorn², T. R. McNutt³, and U. Ehsan⁴; ¹University of Minnesota, Minneapolis, MN, ²University of Minnesota: Department of Radiation Oncology, Minneapolis, MN, ³Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, ⁴Georgia Institute of Technology, Atlanta, GA

Purpose/Objective(s): Generative artificial intelligence (AI)--especially large language models (LLMs)--promises to make a profound impact on providers’ productivity and wellness. Algorithmic benchmarks make impressive claims like LLMs having superior bedside manner and diagnostic abilities than providers. But how useful are LLMs for typical tasks in radiation oncology? Do LLMs save time or cognitive burden? How does it impact provider frustration? Materials/

Methods: Using a scenario-based design with a multidisciplinary team (radiation oncologists (RO), AI researcher, and medical physicist) we created four ecologically-situated tasks that would be helpful to an RO– medical decision making (MDM), letter of medical necessity (LOMN), summary of a consult note, summary of a radiology note). Using a diary study method with standardized prompts, we prompted three publicly available LLMs (Bard, BingChat, ChatGPT) for all use cases. The prompts contained deidentified text from patient notes, case details, and radiology reports. All LLMs were prompted on the same date and multiple points across multiple dates to avoid model update confounds. Using an iterative inductive qualitative coding process, we assessed each use case for perceived usefulness according to themes including accuracy, breadth of response, consistency, efficiency, cognitive burden, and frustration.
Results: Across use cases, while LLMs usually produced outputs, the quality of the output, especially for MDM, added more complexity and frustration. Time spent checking for hallucinations increased the overall burden. While LLMs performed well for summarization tasks and produced moderately useful LOMN, efficiency was offset by time required to review output. MDM output lacked breadth of treatment options but also failed to provide patient-specific recommendations. Table 1 summarizes findings.
Conclusion: In four ecology-situated scenarios investigated, utility of LLMs fell below expectations. While summarization tasks were well-performed, cases where the assistance is most useful–such as medical decision-making– the LLMs performed the worst by including overly generalized guidance and hallucinations. This added to the RO’s cognitive burden and frustration. Gains in efficiency where LLMs successfully summarized technical text may be obviated by liability concerns.

Abstract 3382 – Table 1

	Bard	Bing	ChatGPT 3.5
MDM	Unsatisfactory†	Mostly authentic citations but included non-primary sources Provided dose constraints for incorrect technique	Hallucinates citations Responses vary significantly over time
LOMN	Hallucinates citations	Mostly authentic citations Arguments for specific use case are too general	Hallucinates citations
Radiology Summary	Satisfactory	Unsatisfactory*	Satisfactory
Consult Summary	Unsatisfactory†	Unsatisfactory*	Satisfactory