AAO 2023: The Role of ChatGPT and LLMs in Retina Practice

Flavius Beca, MD
Bascom Palmer Eye Institute, Miami, FL

Retina Subday 2 offered fascinating talks on artificial intelligence (AI). The AI session concluded with Dr Raymond lezzi, MD, discussing the role of Large Language Models (LLMs), such as ChatGPT, in modern clinical practice. Dr Iezzi briefly discussed the history of machine learning AI platforms, with the technology being first invented by Google, furthered developed though the OpenAI platform, and subsequently, brought to wider public audience through ChatGPT. The platform grew exponentially from its launch date on November 30, 2022 to over 180 million users today.

A screenshot of a computer Description automatically generated

The talk opened with wide potential applications, from coding to object recognition and from image generation to writing poetry. Indeed, Dr Iezzi noted all of the images used in his presentation were made by the Dall-E 3 image generation platform of ChatGPT.

However, ChatGPT is not without limitations. When the program was given the Ophthalmic Knowledge Assessment Program (OKAPs) examination, it did well in the general medicine section but performed significantly worse overall with a 0% score in the retina section! Meanwhile, confident data “hallucinations” have been popular topics of discussion and a clear limitation of the software. In fact, when it came to drawing a picture of a retinal detachment with localized tears, the image was more nightmare than daydream. Despite multiple attempts to train the model, it remained unsuccessful at producing an appropriate anatomic illustration. Even still, when asked to localize a tear in a detachment between 2 and 6 o’clock, the program accurately applied Lincoff’s rules to answer. Dr Iezzi points out that as a language model, rather than a structured medical training program, ChatGPT is limited by the language-based information it learns from. Bias is built in. If the predominant discussion on the topic occurs on Facebook or Twitter, that can and is reflected in the output.

A screenshot of a computer screen Description automatically generated

Looking to the future, Dr Iezzi finished his discussion with ways to improve and modify LLMs for clinical use. Running platforms locally allows for HIPAA compliance. Limiting data input to the scientific and medical literature while training on curated and relevant data can improve accuracy of outputs and reduce bias. Current applications include patient education and engagement handouts with opportunity for output in just about any language desired. Future potential applications include decreasing time pre-charting, drafting impressions and plans, and even assisting with identifying research study candidates.

Dr Iezzi concluded with a call to our retina colleagues to engage in this era of generative AI by helping produce the HIPAA-compliant retina optimized models that will allow us to realize the full potential of LLMs in the clinic.

A screenshot of a computer screen Description automatically generated