We are amidst a revolution in artificial intelligence (AI). In this paper, we aim to share our early experience with creating custom AI-powered chatbots in surgical education. While our versions are prototypes, we hypothesize that similar chatbots will one day play a significant role in the field.
In brief, AI is the ability of a computer to perform human-like tasks. Large-language models (LLMs) are a type of AI in which algorithms, trained on large collections of textual data (like books, websites, and papers) can generate authentic writing, images, and sounds. These capabilities have earned LLMs the sobriquet of "Generative AI".1 Generative AI was introduced to the public in November 2022 with the debut of ChatGPT - a LLM built by OpenAI (https://openai.com) with a user-friendly interface that allows users to converse and receive human-like responses. Over 100 million people were using ChatGPT within two months of launch - asking the chatbot to plan their vacations, draft their emails, and even write code.2 Since ChatGPT’s release, OpenAI has introduced improved models – such as GPT-4 – with greater capabilities, increased accuracy, and more up-to-date knowledge.
Thus far, surgeons have approached Generative AI with interest but skepticism. In a recent X (formerly Twitter) poll, a majority (52.5%) of surgeons felt uncomfortable with healthcare professionals or patients using ChatGPT to answer medical questions.3 Despite this, one study found that GPT provided accurate responses to patient questions on colorectal cancer a majority of the time with high inter-rater reliability.4 Moreover, another study found that a sample of adults from the United States had difficulty distinguishing between human and GPT-generated responses to questions.5 While the utility of LLMs in patient education requires further exploration, LLMs clearly exhibit a strong command of declarative knowledge for physicians. For example, GPT-4 was able to pass the United States Medical Licensing Exam and the written neurosurgery board certification exams.6-8
Within surgical education, one exciting application of LLMs is the ability to build custom chatbots that can create and disseminate educational content. Previously, making a surgery specific chatbot required choosing an LLM such as GPT-4, supplying an initialization prompt explaining what the bot was supposed to do, and building a chat-based user interface. We encountered two major obstacles with this approach. First, development of the chatbot required coding, making it inaccessible to non-technical researchers. Moreover, while we were able to take advantage of ChatGPT's immense existing knowledge base, we were unable to supplement it with domain-specific knowledge unique to our use case. While methods exist to do so, they are currently beyond our technical abilities.9
On November 6, 2023, OpenAI unveiled new technology to develop custom chatbots that may address both issues.10 GPTs can be built and iteratively modified using natural language instead of code, and users can upload text documents to provide domain-specific knowledge. The user is first asked what type of chatbot they want to create and then is asked to further specify the chatbot's content, behavior, and tone. The user can upload domain-specific knowledge using a drag-and-drop interface. GPTs can even accept voice and image inputs. These chatbots can then be tested and modified in real time. For example, the user can type "make the chatbot's tone more formal" or "prompt the user by asking more questions."
The potential novel uses of this technology are expansive. Within the first week of its release, we designed chatbots to help users perform thematic analyses, design surveys, build curricula using Kern's Six-Step Framework, simulate American Board of Surgery In-Training Exam questions, and write research protocols according to institution-specific templates. We also developed a patient education chatbot to support patients after bariatric surgery (Table 1).
In our early experience, the GPTs we developed showed promise despite having significant limitations. For example, in our unpublished oral boards simulation GPT, the chatbot referenced pearls from the uploaded knowledge but often gave the user too much information up-front without being prompted. In other cases, the chatbot's performance was more readily usable. In our thematic analysis GPT, the chatbot intuited codes and themes similar to those created from manual review. Overall, GPTs performed better at offering general knowledge rather than adhering to a specific format. The ability to adjust additional parameters such as response-length and format will make these chatbots even more powerful in the future. This also highlights the importance of prompt engineering - the crafting of natural language prompts that the LLM can understand and implement. In fact, a recent study showed that general purpose LLM accuracy can be significantly increased using high-quality prompts without additional domain-specific fine-tuning.11
The ability for any user to make custom chatbots raises some concerns. First, how can the accuracy of these chatbots be verified? LLMs are known to hallucinate - that is, confidently say things that are made-up or false. However, each version of GPT claims to make fewer hallucinations than its predecessors, and the ability to integrate domain-specific knowledge may improve accuracy further still. Second, when we talk about improved accuracy, against what goalpost are we measuring? Currently, there are no rubrics with validity evidence to assess chatbot accuracy, and traditional methods such as Delphi consensus are unlikely to be fast enough to respond to the pace and breadth of chatbots developed. Novel frameworks are needed to determine chatbot accuracy at scale. Third, GPTs were found to leak information about their initialization prompts and training data, raising concerns about the use of private or proprietary information in their construction.12 Finally, the OpenAI models used by GPTs are proprietary, meaning we do not know specifically how they work. Further development of open-source models will help alleviate this issue.
AI is a profound and rapidly evolving technology. It may seem reasonable to say that this technology is not accurate or safe enough today and that we should keep it at arm's length until it is. However, failure to use this resource risks the majority of surgeons, residents, and students being left out of the innovation process. Prior to ChatGPT, the majority of AI-related work was performed by those with a technical background. The only way surgeons could participate was to provide their own data to train models. With the advent of generative AI tools, non-technical users are able to participate in the AI revolution for the first time as partners. While there is significant discussion surrounding the regulation of AI, we must be wary of placing well-intentioned guardrails around its use that end up restricting access to those with prestige or position.
While remaining cognizant of their limitations, we should maintain proficiency in the use of these tools and iteratively find safe use cases. Education is likely to be an initial area of focus. By doing so, we will have a skill base and infrastructure in place when the algorithms are accurate enough to be used in broader contexts such as patient care. This holds especially true for medical students and residents, who may be the most comfortable using these tools, but are the least likely to have a seat at the table when the parameters of their use are determined. Despite its risks, generative AI has the potential to be a democratizing technology in surgical education.
| Initial Prompt | Training Knowledge | Link |
|---|---|---|
| "Make me a qualitative researcher who can help analyze focus group data with thematic analysis" | (Kiger and Vapiano, 2020)13 (Stalmeije et. al. 2014)14 | https://chat.openai.com/g/g-ANULT7gg3-qually-the-qualitative-researcher |
| "Make me an expert IRB reviewer that can help draft IRBs according to my institutions' format and help me determine if my study is exempt" | Federal Policy for the Protection of Human Subjects (the Common Rule) (2018 Revision)15 | https://chat.openai.com/g/g-eDGmfjZb3-kirby |
| "Make me a patient support tool for patients who recently underwent bariatric surgery" | Post-operative pamphlets from the University of Texas Medical Branch | https://chat.openai.com/g/g-2Pyw7lbrh-barry |
| "Make me 10 high quality ABSITE-style multiple choice questions on thyroid-related topics." | Surgical Council on Resident Education modules, NBME Item-Writing Guide | https://chat.openai.com/g/g-IAbUj5I2h-high-yield-surgery-question-wizard |
References
1. Karpathy A. [1hr Talk] Intro to Large Language Models. Published November 22, 2023. Accessed November 29, 2023. https://www.youtube.com/watch?v=zjkBMFhNj_g
2. ChatGPT sets record for fastest-growing user base - analyst note. Reuters. https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/. Published February 2, 2023. Accessed November 29, 2023.
3. SURGERY Journal. I am comfortable for both healthcare professionals & patients to use #AI platforms such as #ChatGPT to answer their medical questions@SWexner @juliomayol @SyedAAhmad5 @dr_samehhany81 @mikifreund @nirhoresh @_TylerLoftus @hayfarani @Laparoscopes @luciacolorectal #SoMe4Surgery. Twitter. Published November 5, 2023. Accessed November 20, 2023. https://twitter.com/SurgJournal/status/1721280763998421415
4. Emile SH, Horesh N, Freund M, et al. How appropriate are answers of online chat-based artificial intelligence (ChatGPT) to common questions on colon cancer? Surgery. 2023;174(5):1273-1275. doi:10.1016/j.surg.2023.06.005
5. Nov O, Singh N, Mann DM. Putting ChatGPT’s medical advice to the (Turing) Test. bioRxiv. Published online January 24, 2023. doi:10.1101/2023.01.23.23284735
6. Kung TH, Cheatham M, Medenilla A, et al. Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models. PLOS Digit Health. 2023;2(2):e0000198. doi:10.1371/journal.pdig.0000198
7. Hopkins BS, Nguyen VN, Dallas J, et al. ChatGPT versus the neurosurgical written boards: a comparative analysis of artificial intelligence/machine learning performance on neurosurgical board-style questions. J Neurosurg. 2023;139(3):904-911. doi:10.3171/2023.2.JNS23419
8. Ali R, Tang OY, Connolly ID, et al. Performance of ChatGPT and GPT-4 on Neurosurgery Written Board Examinations. medRxiv. Published online March 29, 2023:2023.03.25.23287743. doi:10.1101/2023.03.25.23287743
9. Lewis P, Perez E, Piktus A, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv [csCL]. Published online May 22, 2020. http://arxiv.org/abs/2005.11401
10. Introducing GPTs. Accessed December 5, 2023. https://openai.com/blog/introducing-gpts
11. Nori H, Lee YT, Zhang S, et al. Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine. arXiv [csCL]. Published online November 28, 2023. http://arxiv.org/abs/2311.16452
12. Burgess M. OpenAI’s Custom Chatbots Are Leaking Their Secrets. Wired. Published online November 29, 2023. Accessed December 3, 2023. https://www.wired.com/story/openai-custom-chatbots-gpts-prompt-injection-attacks/
13. Kiger ME, Varpio L. Thematic analysis of qualitative data: AMEE Guide No. 131. Med Teach. 2020;42(8):846-854. doi:10.1080/0142159X.2020.1755030
14. Stalmeijer RE, Mcnaughton N, Van Mook WNKA. Using focus groups in medical education research: AMEE Guide No. 91. Med Teach. 2014;36(11):923-939. doi:10.3109/0142159X.2014.917165
15. Office for Human Research Protections (OHRP). Revised common rule. HHS.gov. Published January 17, 2017. Accessed December 5, 2023. https://www.hhs.gov/ohrp/regulations-and-policy/regulations/finalized-revisions-common-rule/index.html