Clinical Knowledge and Reasoning Abilities of AI Large Language Models in Anesthesiology: A Comparative Study on the American Board of Anesthesiology Examination

Mirana C Angel; Joseph B Rinehart; Maxime P Canneson; Pierre Baldi

doi:10.1213/ANE.0000000000006892

Clinical Knowledge and Reasoning Abilities of AI Large Language Models in Anesthesiology: A Comparative Study on the American Board of Anesthesiology Examination

Anesth Analg. 2024 Apr 19. doi: 10.1213/ANE.0000000000006892. Online ahead of print.

Authors

Mirana C Angel^{1

2}, Joseph B Rinehart³, Maxime P Canneson⁴, Pierre Baldi^{1

2}

Affiliations

¹ From the Department of Computer Science, University of California Irvine, Irvine, California.
² Institute for Genomics and Bioinformatics, University of California Irvine, Irvine, California.
³ Department of Anesthesiology & Perioperative Care, University of California Irvine, Irvine, California.
⁴ Department of Anesthesiology & Perioperative Medicine, University of California Los Angeles, Los Angeles, California.

PMID: 38640076
DOI: 10.1213/ANE.0000000000006892

Abstract

Background: Over the past decade, artificial intelligence (AI) has expanded significantly with increased adoption across various industries, including medicine. Recently, AI-based large language models such as Generative Pretrained Transformer-3 (GPT-3), Bard, and Generative Pretrained Transformer-3 (GPT-4) have demonstrated remarkable language capabilities. While previous studies have explored their potential in general medical knowledge tasks, here we assess their clinical knowledge and reasoning abilities in a specialized medical context.

Methods: We studied and compared the performance of all 3 models on both the written and oral portions of the comprehensive and challenging American Board of Anesthesiology (ABA) examination, which evaluates candidates' knowledge and competence in anesthesia practice.

Results: Our results reveal that only GPT-4 successfully passed the written examination, achieving an accuracy of 78% on the basic section and 80% on the advanced section. In comparison, the less recent or smaller GPT-3 and Bard models scored 58% and 47% on the basic examination, and 50% and 46% on the advanced examination, respectively. Consequently, only GPT-4 was evaluated in the oral examination, with examiners concluding that it had a reasonable possibility of passing the structured oral examination. Additionally, we observe that these models exhibit varying degrees of proficiency across distinct topics, which could serve as an indicator of the relative quality of information contained in the corresponding training datasets. This may also act as a predictor for determining which anesthesiology subspecialty is most likely to witness the earliest integration with AI.

Conclusions: GPT-4 outperformed GPT-3 and Bard on both basic and advanced sections of the written ABA examination, and actual board examiners considered GPT-4 to have a reasonable possibility of passing the real oral examination; these models also exhibit varying degrees of proficiency across distinct topics.