To the editor,
Large language models (LLMs) represented by ChatGPT have shown promising potential in the field of medicine [1, 2]. However, it should be noted that the answers provided by ChatGPT may contain errors [3]. In addition, other companies have launched internet-connected LLMs that can access the latest data, potentially outperforming ChatGPT which was trained on pre-September 2021 data. Prostate cancer(PCa) is the second-most common type of cancer in men globally, with a relatively long survival time compared with other cancer types [4]. Taking PCa as an example, we evaluated whether these LLMs could provide correct and useful information on common problems related to PCa and provide appropriate humanistic care, thus contributing to the democratization of medical knowledge.
We designed 22 questions based on patient education guidelines (CDC and UpToDate) and our own clinical experience, covering screening, prevention, treatment options, and postoperative complications (Table 1). The questions ranged from basic to advanced knowledge of PCa. A total of five state-of-the-art LLMs were included, including ChatGPT (Free and Plus version), YouChat, NeevaAI, Perplexity (concise and detailed model), and Chatsonic. The quality of the answers was primarily evaluated based on their accuracy, comprehensiveness, patient readability, humanistic care and stability.
Table 1.
No. | Questions | Difficulty level |
---|---|---|
1 | What is prostate cancer? | Basic |
2 | What are the symptoms of prostate cancer? | Basic |
3 | How can I prevent from prostate cancer? | Basic |
4 | Who is at risk of prostate cancer? | Basic |
5 | How is prostate cancer diagnosed? | Basic |
6 | What is a prostate biopsy? | Basic |
7 | How is prostate cancer treated? | Basic |
8 | How long can I live if I have prostate cancer? | Basic |
9 | How often do I need get a PSA test? | Basic |
10 | What is prostate-specific antigen? | Basic |
11 | What is screening for prostate cancer? | Basic |
12 | Should I get screened for prostate cancer? | Basic |
13 | My father had prostate cancer. Will I have prostate cancer too? | Hard |
14 | I have a high PSA level. Do I have prostate cancer? | Hard |
15 | What does a PSA level of 4 mean? | Hard |
16 | What does a PSA level of 10 mean? | Hard |
17 | What does a PSA level of 20 mean? | Hard |
18 | The doctor said my prostate is totally removed by surgery. Why my PSA is still high after surgery? | Hard |
19 | I have localized prostate cancer. Which is better, the radiation therapy or the surgery? | Hard |
20 | Should I have robotic surgery or laparoscopic surgery if I have prostate cancer? | Hard |
21 | What is the best medicine for Castration-resistant prostate cancer? | Hard |
22 | Which is better for prostate cancer? Apalutamide or Enzalutamide? | Hard |
The accuracy of most LLMs’ responses was above 90%, except for NeevaAI and Chatsonic (Fig. 1A). For basic information questions with definite answers, most LLMs could achieve a high accuracy. Nevertheless, the accuracy decreased in questions associated with specific scenario, or in questions that involved summary and analysis (e.g., Why the PSA is still high after surgery?). Among these LLMs, ChatGPT had the highest accuracy rate, and the free version of ChatGPT was slightly better than the paid version.
Evaluations of comprehensiveness show that LLMs performs well in answering most questions (Fig. 1B). For example, they can effectively highlight different PSA level significance, remind patients that PSA is not the final diagnostic test, and suggest further examination. They can also compare treatment options in detail, outlining the pros and cons, and provide helpful references for patients to make informed decisions. In addition, it is commendable that most responses point out the need for patients to consult their doctors for more advice. The readability of responses from most LLMs, except NeevaAI, was satisfactory (Fig. 1C). We believe that patients can understand the information conveyed in LLMs’ responses in most cases. All LLMs could provide humanistic care when discussing expected lifespan, informing patients about the relatively long survival time of PCa, which eased anxiety. However, they did not exhibit humanistic care when answering other inquiries. LLMs’ responses were generally stable, but inconsistent outcomes were detected in some instances (Fig. 1D).
We then analyzed the reasons for the poor performance of LLMs in some responses. The most common issue was the mixture of outdated or incorrect information in the answers, including claims that open surgery is a more common choice for prostate cancer radical prostatectomy than robot-assisted surgery [5], and inaccurate responses regarding the approved indications when comparing apalutamide and enzalutamide. Inadequate comprehensiveness was mainly due to lack of specific details or omission of key points. For instance, Perplexity missed screening as an important measure in preventing PCa. Regarding the frequency of PSA testing, some answers only recommended a case-by-case approach, without specifying testing frequency for different age groups. LLMs sometimes misunderstand background information and provide inaccurate answers, such as mechanically suggesting that “PSA testing is not the final diagnostic test for PCa,” but monitoring PSA after prostatectomy is clearly not for the purpose of diagnosing PCa. It must be noted that some AI models based on search engines such as NeevaAI tend to simply provide the content of literature without summarizing and explaining, leading to poor readability. While we anticipated that the internet-connected LLMs would surpass ChatGPT, they failed to do so. This suggests that model training may be more important than real-time internet-connection.
Although not yet perfect, LLMs can provide correct answers to basic questions that PCa patients are concerned about and can analyze specific situations to a certain extent. LLMs have the potential to be applied in patient education and consultation, providing patient-friendly information to help them understand their medical conditions and treatment options, enabling shared decision-making. More importantly, LLMs can help democratize medical knowledge, providing timely access to accurate medical information regardless of geographic or socioeconomic status. This is especially important for underserved populations in medical deserts, and those facing longer waiting times for medical care during the pandemics like COVID-19. We believe that LLMs have unlimited potential with the rapid development of AI.
However, current LLMs are not yet capable of completely replace doctors, as they may contain errors or omit key points in responses, still have significant shortcomings in analyzing questions in specific contexts and cannot ask patients additional questions to gather more information. Moreover, they still cannot comfort patients like humans.
Acknowledgements
Not applicable.
Author contributions
LZ: conceptualization, methodology, investigation, formal analysis, writing—original draft, visualization; WM: conceptualization, investigation, visualization, writing—review and editing; RC: supervision, funding acquisition, writing—review and editing, conceptualization, methodology. All authors read and approved the final manuscript.
Funding
This study is supported by the Rising-Star Program of Science and Technology Commission of Shanghai Municipality (21QA1411500), Natural Science Foundation of Science and Technology Commission of Shanghai (22ZR1478000), and the National Natural Science Foundation of China (82272905). The funding source had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Availability of data and materials
The data that support the findings of this study are available on request from the corresponding author upon reasonable request.
Declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential competing interests.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Lingxuan Zhu and Weiming Mou have contributed equally to this work and share first authorship
References
- 1.Johnson SB, King AJ, Warner EL, Aneja S, Kann BH, Bylund CL. Using ChatGPT to evaluate cancer myths and misconceptions: artificial intelligence and cancer information. JNCI Cancer Spectr. 2023;7:pkad015. doi: 10.1093/jncics/pkad015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Hopkins AM, Logan JM, Kichenadasse G, Sorich MJ. Artificial intelligence chatbots will revolutionize how cancer patients access information: ChatGPT represents a paradigm-shift. JNCI Cancer Spectr. 2023;7:pkad010. doi: 10.1093/jncics/pkad010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Sarraju A, Bruemmer D, Van Iterson E, Cho L, Rodriguez F, Laffin L. Appropriateness of cardiovascular disease prevention recommendations obtained from a popular online chat-based artificial intelligence model. JAMA. 2023 doi: 10.1001/jama.2023.1044. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Rawla P. Epidemiology of prostate cancer. World J Oncol. 2019;10:63–89. doi: 10.14740/wjon1191. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Crew B. Worth the cost? A closer look at the da Vinci robot’s impact on prostate cancer surgery. Nature. 2020;580:S5–7. doi: 10.1038/d41586-020-01037-w. [DOI] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The data that support the findings of this study are available on request from the corresponding author upon reasonable request.