Background: As artificial intelligence (AI) tools become widely accessible, more patients and medical professionals will turn to them for medical information. Large language models (LLMs), a subset of AI, excel in natural language processing tasks and hold considerable promise for clinical use. Fields such as oncology, in which clinical decisions are highly dependent on a continuous influx of new clinical trial data and evolving guidelines, stand to gain immensely from such advancements. It is therefore of critical importance to benchmark these models and describe their performance characteristics to guide their safe application to clinical oncology. Accordingly, the primary objectives of this work were to conduct comprehensive evaluations of LLMs in the field of oncology and to identify and characterize strategies that medical professionals can use to bolster their confidence in a model's response.
Methods: This study tested five publicly available LLMs (LLaMA 1, PaLM 2, Claude-v1, generative pretrained transformer 3.5 [GPT-3.5], and GPT-4) on a comprehensive battery of 2044 oncology questions, including topics from medical oncology, surgical oncology, radiation oncology, medical statistics, medical physics, and cancer biology. Model prompts were presented independently of each other, and each prompt was repeated three times to assess output consistency. For each response, models were instructed to provide a self-appraised confidence score (from 1 to 4). Model performance was also evaluated against a novel validation set comprising 50 oncology questions curated to eliminate any risk of overlap with the data used to train the LLMs.
Results: There was significant heterogeneity in performance between models (analysis of variance, P<0.001). Relative to a human benchmark (2013 and 2014 examination results), GPT-4 was the only model to perform above the 50th percentile. Overall, model performance varied as a function of subject area across all models, with worse performance observed in clinical oncology subcategories compared with foundational topics (medical statistics, medical physics, and cancer biology). Within the clinical oncology subdomain, worse performance was observed in female-predominant malignancies. A combination of model selection, prompt repetition, and confidence self-appraisal allowed for the identification of high-performing subgroups of questions with observed accuracies of 81.7 and 81.1% in the Claude-v1 and GPT-4 models, respectively. Evaluation of the novel validation question set produced similar trends in model performance while also highlighting improved performance in newer, centrally hosted models (GPT-4 Turbo and Gemini 1.0 Ultra) and local models (Mixtral 8×7B and LLaMA 2).
Conclusions: Of the models tested on a standardized set of oncology questions, GPT-4 was observed to have the highest performance. Although this performance is impressive, all LLMs continue to have clinically significant error rates, including examples of overconfidence and consistent inaccuracies. Given the enthusiasm to integrate these new implementations of AI into clinical practice, continued standardized evaluations of the strengths and limitations of these products will be critical to guide both patients and medical professionals. (Funded by the National Institutes of Health Clinical Center for Research and the Intramural Research Program of the National Institutes of Health; Z99 CA999999.).