The Application of Large Language Models for Radiologic Decision Making

Abstract

Background and purposeLarge language models (LLMs) have seen explosive growth, but their potential role in medical applications remains underexplored. Our study investigates the capability of LLMs to predict the most appropriate imaging study for specific clinical presentations in various subspecialty areas in radiology. Methods and materialsChat Generative Pretrained Transformer (ChatGPT), by OpenAI and Glass AI by Glass Health were tested on 1,075 clinical scenarios from 11 ACR expert panels to determine the most appropriate imaging study, benchmarked against the ACR Appropriateness Criteria. Two responses per clinical presentation were generated and averaged for the final clinical presentation score. Clinical presentation scores for each topic area were averaged as its final score. The average of the topic scores within a panel determined the final score of each panel. LLM responses were on a scale of 0 to 3. Partial scores were given for nonspecific answers. Pearson correlation coefficient (R-value) was calculated for each panel to determine a context-specific performance. ResultsGlass AI scored significantly higher than ChatGPT (2.32 ± 0.67 versus 2.08 ± 0.74, P = .002). Both LLMs performed the best in the Polytrauma, Breast, and Vascular panels, and performed the worst in the Neurologic, Musculoskeletal, and Cardiac panels. Glass AI outperformed ChatGPT in 10 of 11 panels, except Obstetrics and Gynecology. Maximum agreement was in the Pediatrics, Neurologic, and Thoracic panels, and the most disagreement occurred in the Vascular, Breast, and Urologic panels. ConclusionLLMs can be used to predict imaging studies, with Glass AI’s superior performance indicating the benefits of extra medical-text training. This supports the potential of LLMs in radiologic decision making.