Medical education with large language models in ophthalmology: custom instructions and enhanced retrieval capabilities

Abstract

Foundation models are the next generation of artificial intelligence that has the potential to provide novel use cases for healthcare. Large language models (LLMs), a type of foundation model, are capable of language comprehension and the ability to generate human-like text. Researchers and developers have been tuning LLMs to optimise their performance in specific tasks, such as medical challenge problems. Until recently, tuning required technical programming expertise, but the release of custom generative pre-trained transformers (GPTs) by OpenAI has allowed users to tune their own GPTs with natural language. This has the potential to democratise access to high-quality bespoke LLMs globally. In this review, we provide an overview of LLMs, how they are tuned and how custom GPTs work. We provide three use cases of custom GPTs in ophthalmology to demonstrate the versatility and effectiveness of these tools. First, we present 'EyeTeacher', an educational aid that generates questions from clinical guidelines to facilitate learning. Second, we built 'EyeAssistant', a clinical support tool that is tuned with clinical guidelines to respond to various physician queries. Lastly, we design 'The GPT for GA', which offers clinicians a comprehensive summary of emerging management strategies for geographic atrophy by analysing peer-reviewed documents. The review underscores the significance of custom instructions and information retrieval in tuning GPTs for specific tasks in ophthalmology. We also discuss the evaluation of LLM responses and address critical aspects such as privacy and accountability in their clinical application. Finally, we discuss their potential in ophthalmic education and clinical practice.

Keywords: Medical Education.

Figure 1. Simplified comparison between generic and custom GPT architectures. Generic GPTs operate linearly: a user prompt is processed by an LLM to generate a response. Custom GPTs integrate custom instructions and domain-specific knowledge into the process. Here, a user prompt is combined with custom instructions (provided by the developer) to steer the LLM, which can also retrieve external knowledge (eg, web searches, clinical guidelines and databases) to produce an aligned response. GPT, generative pretrained transformer; LLM, large language model.

Figure 2. Custom GPTs are built in natural language using a builder chatbot. The GPT builder enables users to input custom instructions using natural language and upload specific knowledge datasets for retrieval. These modifications are implemented in the GPT‘s backend. Additionally, the builder offers the option to enhance or limit functionalities, such as web browsing, image creation and code interpretation and generation, although the latter may be less relevant for our specific use cases. GPT, generative pretrained transformer.

Figure 3. EyeTeacher is a custom GPT built for education. In this enhanced learning scenario, EyeTeacher presents a multiple-choice question about diabetic macular oedema. When a student selects an answer, EyeTeacher elaborates on why the answer is correct and explains why each distractor is incorrect. This approach is crucial because it does not merely validate the student’s choice but educates them about the nuances of each option. This method promotes active learning, as students do not just memorise the right answer; they understand the rationale behind each option. The explanations are free of hallucinations and factually correct. GPT, generative pretrained transformer.

Figure 4. EyeAssistant is a custom GPT built for clinical support. In this clinical scenario, EyeAssistant presents an evidence-based response regarding the management of narrow angles or primary angle closure suspects. From the guideline, it identifies risk factors for angle closure that may warrant prophylactic iridotomy. When prompted to justify the answer and describe the popular Zhongshan Angle Closure Prevention trial, it summarises the findings adequately and contextualises the findings for the London-based ophthalmologist. The explanations are free of hallucinations and factually correct. GPT, generative pretrained transformer

Figure 5. The GPT for GA is a custom GPT built to provide balanced information on the treatments for geographic atrophy. When queried about FDA-approved treatments for GA, The GPT for GA accurately identifies pegcetacoplan (Syfovre) and avacincaptad pegol (Izervay). It appropriately references the pertinent pivotal trials that led to their approval. When prompted to discuss the functional benefits of these treatments, it judiciously cites relevant sources while acknowledging that the associated functional benefits remain limited. Addressing safety concerns, it correctly highlights the elevated risk of macular neovascularisation associated with both drugs and the potential for intraocular inflammation with pegcetacoplan. When asked about the latter, it also provides up-to-date information and cites authoritative sources, such as the American Society of Retina Specialists (ASRS) ReST committee. Of note, it mistakenly refers to the ASRS as the American Society of Retinal Surgeons. FDA, food and drug administration; GA, geographic atrophy; GPT, generative pretrained transformer.