Performance and Agreement of Large Language Models (LLMs) on Radiology Resident in Training Examinations

Purpose: As LLMs “learn” more daily, it is vital to understand their capabilities, especially in highly specialized domains such as radiology. To determine the LLM output accuracy, reliability and limitations, this study seeks to assess the overall inter-model performance variability and agreement between the ChatGPT-4, ChatGPT-3.5, Google Bard, and Claude LLMs.

Methods/Materials: The American College of Radiology Diagnostic In-Training Examination questions from 2020-2022 were utilized (n=172). Image questions were excluded due to the inability of LLMs to process images. Questions were run through the LLMs in Nov/Dec 2023 without any prompting and were queried three times. Chi-squared analysis was used to compare the performances. Post-hoc stratification was used to determine an accuracy “model success rate” (“Easy” = No Models incorrect, “Moderate” = 1-2 incorrect, “Hard” = 3 or more incorrect) . For the “hard” questions, an "agreement rate” was calculated(If 25% of models answered “A” and 75% of models answered “C”, then agreement was 75%).

Results: There was no significant difference in performances overall (p=.08). GPT-4 performed the best (76±5.4%), followed by Google Bard (72±4.1%), Claude (71±1.2%) and GPT-3.5 (63±7.3%). Based on inter-model accuracy stratification, 41% were “Easy”, 35% “Moderate”, and 24% “Hard”. “Moderate” question performance was: GPT-4 (80%), Google Bard (71%), Claude (68%), and GPT-3.5 (45%) without significant difference (p=.10). “Hard” question performance across the models was between 12-17%. Amongst the “Hard” questions, agreement between models on wrong answers ranged from 0.33 (all different answers) to 1.00 (all same answer). The average showed that models agreed on wrong answers 75% of the time.

Conclusions: The results demonstrate LLM radiology domain specific capabilities and their limitations with significant variability within and across models. Moving forward, standardized evaluation metrics on the performance, accuracy, and reliability of these models will need to be developed prior to integration in educational settings. Future studies will include prompt engineering and improved training approaches to improve response accuracy on radiology domain-specific questions.