Large language model (LLM)-driven chatbots for neuro-ophthalmic medical education

The subspecialty of neuro-ophthalmology is currently facing a human resource issue globally, with few entering the subspecialty. This has led to access problems for neuro-ophthalmic care worldwide. While the primary goal of ophthalmology is to save a patient’s vision, a good neuro-ophthalmic diagnosis can potentially save lives. Neuro-ophthalmology is a unique subspecialty of ophthalmology that involves going beyond the eye, relying heavily on principles of general medicine but also the skill and expertise of a neuro-ophthalmologist. While becoming a neuro-ophthalmologist can be considered a long and arduous journey, we believe that stronger neuro-ophthalmic expertise can benefit a variety of clinicians, from ophthalmologists to neurosurgeons. In this paper, we examine how GPT-4 can potentially benefit neuro-ophthalmic education.

GPT-4 is a model based on transformers. The model underwent pretraining using an extensive dataset of both publicly available data as well as data procured from third-party providers [1]. During this phase, GPT-4 acquired the ability to anticipate the subsequent token in a sequence [2]. After the pretraining phase, GPT-4 went through a fine-tuning process that employs reinforcement learning methods [2]. The enhances the efficacy of the model, making it more reliable and responsive to various tasks [2]. GPT-4 has already been shown to be useful in writing ophthalmic surgical notes [3]. In contrast to its previous generations which had a primary emphasis on textual inputs, GPT-4 now has the capability to process both textual and visual inputs [4, 5].

GPT-4 can provide a personalized learning experience tailored to the exact needs and abilities of a learner [1]. The interactive nature of GPT-4 can make learning more engaging and GPT-4 can provide updated information on neuro-ophthalmic conditions instantly. This can include management of various conditions, neuro-ophthalmic procedures, and physical exams. We decided to examine GPT-4’s ability to generate an accurate neuro-ophthalmic management plan for optic neuritis (Fig. 1).

Fig. 1.

Generated from the prompt “how do you manage optic neuritis”.

Optic neuritis is painful, subacute vision loss resulting from demyelination [6]. The management plan provided by GPT-4 was both specific and accurate, and included both medical and supportive management. The LLM-generated management plan included the administration of high-dose corticosteroids, pain relief with non-steroidal anti-inflammatory drugs, and the treatment of any underlying conditions. GPT-4 also correctly recommended to attend regular follow-up appointments with an ophthalmologist or neurologist. However, it is important to note that GPT-4 omitted an important aspect of the management, the evaluation of neuromyelitis optica, which can be vision threatening. Omissions such as this show the dangers of an overreliance on LLMs for patient care.

ChatGPT can also help to provide brief summaries new studies in the neuro-ophthalmic literature. With the constant stream of new publications in neuro-ophthalmology a learner could not possibly have time to keep up with reading the entirety of this literature. Although keeping up with the latest neuro-ophthalmic developments would be highly beneficial for learning. We asked GPT-4 to summarize a systematic review by Yu et al. [7] on anaemia and idiopathic intracranial hypertension (Fig. 2).

Fig. 2.

Generated from the prompt: Summarize the study “Anemia and Idiopathic Intracranial Hypertension: A Systematic Review and Meta-analysis”.

GPT-4 correctly summarized the study by describing the correct aim and study type. The conclusion was also summarized correctly, that the study found that anaemia may be considered as a potential risk factor for the development of IIH. GPT-4 also correctly states that while this study suggests an association but this does not provide causation and that further research is required to understand this mechanism. The only error provided with the response was that “a total of X studies” while 5 observational or case control studies were included in the actual systematic review and meta-analysis.

Generative AI and GPT technologies have also been previously shown to be able to help learners understand complex neuro-ophthalmic visual phenomena by generating visual representations from text [8]. These generated visual representations may potentially help clinicians better understand abstract visual phenomena for conditions such as Charles Bonnet syndrome or oscillopsia in multiple sclerosis. A better understanding of these conditions may also potentially foster an increased sense of clinician empathy.

It is also important to acknowledge the limitations of GPT-4 and other large language models (LLMs). LLMs can have a deficiency of human-like understanding of certain topics and can occasionally provide incorrect information. Hence it is possible for GPT-4 to produce responses that are absurd or incorrect in certain scenarios or when presented with specific challenges. Concerns have previously been made with ChatGPT regarding unintentional plagiarism [5]. Future improvements can be made to GPT-4 such as providing references, similar to the LLM from Bing, which does this. By providing referencing, the trustworthiness of the sources used to synthesize information can be critically evaluated and plagiarism can be better avoided. All things considered, we believe that GPT-4 has the potential to bring significant benefits to neuro-ophthalmic medical education and to the entire field of medicine.

Author contributions

EW: writing, figure development. JO: writing, figure development. MM: writing. and AGL: review, intellectual support.

Competing interests

The authors declare no competing interests.