This paper is in the following e-collection/theme issue:

Research Letter (4) Artificial Intelligence (AI) in Medical Education (50) New Methods and Approaches in Medical Education (192) New Resources for Medical Education (53) Artificial Intelligence (117) Chatbots and Conversational Agents (458) Clinical Information and Decision Making (1314) Generative Language Models Including ChatGPT (201)

Published on in Vol 10 (2024)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/54283, first published .
The Performance of ChatGPT-4V in Interpreting Images and Tables in the Japanese Medical Licensing Exam

The Performance of ChatGPT-4V in Interpreting Images and Tables in the Japanese Medical Licensing Exam

The Performance of ChatGPT-4V in Interpreting Images and Tables in the Japanese Medical Licensing Exam

Authors of this article:

Soshi Takagi1 Author Orcid Image ;   Masahide Koda2 Author Orcid Image ;   Takashi Watari3, 4 Author Orcid Image
Article Authors Cited by Tweetations (2) Metrics

    1Faculty of Medicine, Shimane University, , Izumo, , Japan

    2Co-learning Community Healthcare Re-innovation Office, Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama University, , Okayama, , Japan

    3General Medicine Center, Shimane University Hospital, , Izumo, , Japan

    4Integrated Clinical Education Center, Kyoto University Hospital, , Kyoto, , Japan

    *all authors contributed equally

    Corresponding Author:

    Takashi Watari, MHQS, MD, PhD




    OpenAI’s ChatGPT, a leading large language model (LLM), has shown promise for medical purposes. The program can pass the United States Medical Licensing Examination (USMLE) and the Japanese Medical Licensing Exam (JMLE) [1-3]. However, previous studies regarding this software have focused on its text-based capabilities. ChatGPT-4 Vision (ChatGPT-4V), announced on September 25, 2023, includes image input features, potentially expanding the medical applications of the program [4]. To assess the multimodal performance of ChatGPT-4V in medicine, its performance on JMLE questions involving clinical images and tables was tested.


    Overview

    ChatGPT-4V was used to complete the 117th JMLE in the Japanese language (Figure S1 in Multimedia Appendix 1). Its responses were compared to the passing criteria and mean human examinee score of the JMLE. This study, conducted from October 12 to 14, 2023, used the September 25, 2023, version of the LLM (ChatGPT-4V) with a knowledge cutoff date of January 2022 (Multimedia Appendix 2 [5]). Human examinees’ correct response rates were obtained from statistics based on reports from actual JMLE examinees, calculated by medu4, a preparatory school for the JMLE [5,6].

    Statistical Analysis

    The mean and 95% CIs of the test scores are provided. A one-sample proportion test was used to compare the correct response rate of the human examinees with that of ChatGPT-4V. Statistical significance was set at P<.05 for all 2-tailed tests. All statistical analyses were conducted using Stata statistical software (version 17; StataCor).

    Ethical Considerations

    This study used previously available web-based data and did not include human participants. Therefore, Shimane University’s Institutional Review Board did not mandate ethics approval.


    Evaluation Outcomes

    The responses to 386 questions from the 117th JMLE were used in this study. Using the Ministry of Health, Labor, and Welfare criteria, GPT-4V scored 85.1% on the essential knowledge section and 76.5% on the other sections of the JMLE, meeting the passing criteria [6]. For text-only questions, ChatGPT-4V achieved a correct response rate of 84.5%, similar to the mean human examinee score (Table 1). The correct response rate for questions with images was 71.9% for ChatGPT-4V, 13.1 points below the mean human examinee score (P<.001). The correct response rate for questions with tables (including figures) was 35.0% for ChatGPT-4V, which was significantly lower than the mean human examinee score (83.9%; P<.001).

    Table 1. Correct response rates of ChatGPT-4 Vision (ChatGPT-4V) and human examinees on the Japanese Medical Licensing Examination (JMLE).
    CharacteristicsTotal, n (%)Examineesa, meanGPT-4V, mean95% CIDifferenceP value
    All questions386 (100)84.978.274.1-82.4−6.7.003
    Question category
    Essential knowledge96 (24.9)89.683.375.9-90.8−6.3.04
    General clinical knowledge144 (37.3)83.170.863.4-78.3−12.3<.001
    Specific diseases146 (37.8)83.582.276.0-88.4−1.3.67
    Type
    General190 (49.2)84.678.973.2-84.7−5.7.03
    Clinical149 (38.6)84.177.270.4-83.0−6.9.02
    Clinical sentence47 (12.2)88.578.767.0-90.4−9.8.04
    Imaging and table questions
    Text only252 (65.3)84.984.580.1-89.0−0.4.87
    With images114 (29.5)85.071.963.7-80.2−13.1<.001
    With tables20 (5.2)83.935.014.1-55.9−48.9<.001

    aThe correct response rates of human examinees are based on a survey of actual human examinees, reported by medu4, a preparatory school for the JMLE [5].


    Principal Results

    Although ChatGPT-4V demonstrated proficiency in text-centric questions, the correct response rates were significantly lower for image and table-oriented questions. ChatGPT-4V may have poorer text comprehension skills compared to ChatGPT-4, even when image processing is not required [7]. Additionally, a language bias may obscure the image context when interpreting images and texts simultaneously, potentially leading to an overreliance on prior text information, even when it contradicts the image context, a phenomenon called “hallucination” [8]. These factors may have led to ChatGPT-4V’s lower rate of correct responses to questions involving images.

    Furthermore, responding to questions with tables requires interpreting the Japanese characters within the tables. OpenAI has verified that its GPT-4V model misrecognizes symbols, including image characters [4]. Previous studies have noted that GPT-4V relies on text-based information rather than an analysis of tables when answering questions [8]. In addition, the program’s performance diminishes when interpreting characters in non-Latin languages [9]. These factors may explain the observed decline in performance when interpreting tables containing Japanese characters.

    The multimodal LLM GPT-4V is unreliable in interpreting information presented in image or tables, especially for medical purposes [4]. Further development of the program is required for diagnostic applications.

    Limitations

    This study has several limitations. First, different results may be obtained even when using the same methods owing to the inherent randomness of ChatGPT or version changes in ChatGPT. A report indicates that test results can vary with repeated responses from ChatGPT [10]. Furthermore, when providing images to ChatGPT, we did not remove blank spaces, indicating that the quality of images sent to ChatGPT could also affect the outcomes. Second, the JMLE includes options that, if selected twice or more, will result in failure. However, these options are not publicly disclosed, making them unaccounted for in this study [5]. Finally, although this study focused on ChatGPT, ongoing advancements in other multimodal LLMsshould also be considered.

    Conclusions

    ChatGPT-4V successfully passed the 117th JMLE, demonstrating proficiency in handling including image- and table-based questions. However, more developments are needed to improve its ability to interpret tables. Further research should assess the safety and efficacy of ChatGPT-4V as a multimodal LLM in supporting medical practice, facilitating learning in clinical environments and advancing medical education.

    Acknowledgments

    We would like to thank Dr Kota Sakaguchi, Shimane University Hospital, for his careful support throughout this study. We would also like to thank Dr Sanjay Saint, a professor at the University of Michigan, for his numerous contributions and support in this work.

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Additional statistics.

    DOCX File, 2285 KB
    Multimedia Appendix 2

    Detailed methodology.

    DOCX File, 17 KB
    1. Introducing ChatGPT. OpenAI. 2022. URL: https://openai.com/blog/chatgpt/ [Accessed 2023-11-30]
    2. Gilson A, Safranek CW, Huang T, et al. How does ChatGPT perform on the United States Medical Licensing Examination? The implications of large language models for medical education and knowledge assessment. JMIR Med Educ. Feb 8, 2023;9:e45312. [CrossRef] [Medline]
    3. Takagi S, Watari T, Erabi A, Sakaguchi K. Performance of GPT-3.5 and GPT-4 on the Japanese Medical Licensing Examination: comparison study. JMIR Med Educ. Jun 29, 2023;9:e48002. [CrossRef] [Medline]
    4. GPT-4V(Ision) system card. OpenAI. 2023. URL: https://cdn.openai.com/papers/GPTV_System_Card.pdf [Accessed 2023-10-26]
    5. Announcement of successful passage of the 117th National Medical Examination (Japanese) [Article in Japanese]. Ministry of Health. 2023. URL: https://www.mhlw.go.jp/general/sikaku/successlist/2023/siken01/about.html [Accessed 2023-10-26]
    6. Searching questions [Article in Japanese]. Medu4. 2023. URL: https://medu4.com/quizzes/search [Accessed 2023-10-26]
    7. Wu Y, Wang S, Yang H, et al. An early evaluation of GPT-4V(ision). arXiv. Preprint posted online on Oct 25, 2023. URL: https://arxiv.org/abs/2310.16534 [Accessed 2024-05-14]
    8. Liu F, Lin K, Li L, Wang J, Yacoob Y, Wang L. Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning. arXiv. Preprint posted online on Jun 26, 2023. URL: https://arxiv.org/abs/2306.14565 [Accessed 2024-05-14]
    9. Shi Y, Peng D, Liao W, Lin Z, Chen X, Liu C, et al. Exploring OCR capabilities of GPT-4V(ision): a quantitative and in-depth evaluation. arXiv. Preprint posted online on Oct 25, 2023. URL: https://arxiv.org/abs/2310.16809 [Accessed 2024-05-14]
    10. Zhu L, Mou W, Yang T, Chen R. ChatGPT can pass the AHA exams: open-ended questions outperform multiple-choice format. Resuscitation. Jul 2023;188:109783. [CrossRef] [Medline]

    ‎
    ChatGPT-4V: ChatGPT 4 Vision
    JMLE: Japanese Medical Licensing Examination
    LLM: large language model
    USMLE: United States Medical Licensing Examination

    Edited by Gunther Eysenbach, Taiane de Azevedo Cardoso; submitted 06.11.23; peer-reviewed by Fuxiao Liu, Lingxuan Zhu, Tianyu Ma; final revised version received 09.04.24; accepted 22.04.24; published 23.05.24.

    Copyright

    © Soshi Takagi, Masahide Koda, Takashi Watari. Originally published in JMIR Medical Education (https://mededu.jmir.org), 23.5.2024.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Education, is properly cited. The complete bibliographic information, a link to the original publication on https://mededu.jmir.org/, as well as this copyright and license information must be included.