ChatGPT and Lacrimal Drainage Disorders: Performance and Scope of Improvement.

Abstract

This study aimed to report the performance of the large language model ChatGPT (OpenAI, San Francisco, CA, U.S.A.) in the context of lacrimal drainage disorders. A set of prompts was constructed through questions and statements spanning common and uncommon aspects of lacrimal drainage disorders. Care was taken to avoid constructing prompts that had significant or new knowledge beyond the year 2020. Each of the prompts was presented thrice to ChatGPT. The questions covered common disorders such as primary acquired nasolacrimal duct obstruction and congenital nasolacrimal duct obstruction and their cause and management. The prompts also tested ChatGPT on certain specifics, such as the history of dacryocystorhinostomy (DCR) surgery, lacrimal pump anatomy, and human canalicular surfactants. ChatGPT was also quizzed on controversial topics such as silicone intubation and the use of mitomycin C in DCR surgery. The responses of ChatGPT were carefully analyzed for evidence-based content, specificity of the response, presence of generic text, disclaimers, factual inaccuracies, and its abilities to admit mistakes and challenge incorrect premises. Three lacrimal surgeons graded the responses into three categories: correct, partially correct, and factually incorrect. A total of 21 prompts were presented to the ChatGPT. The responses were detailed and were based according to the prompt structure. In response to most questions, ChatGPT provided a generic disclaimer that it could not give medical advice or professional opinion but then provided an answer to the question in detail. Specific prompts such as "how can I perform an external DCR?" were responded by a sequential listing of all the surgical steps. However, several factual inaccuracies were noted across many ChatGPT replies. Several responses on controversial topics such as silicone intubation and mitomycin C were generic and not precisely evidence-based. ChatGPT's response to specific questions such as canalicular surfactants and idiopathic canalicular inflammatory disease was poor. The presentation of variable prompts on a single topic led to responses with either repetition or recycling of the phrases. Citations were uniformly missing across all responses. Agreement among the three observers was high (95%) in grading the responses. The responses of ChatGPT were graded as correct for only 40% of the prompts, partially correct in 35%, and outright factually incorrect in 25%. Hence, some degree of factual inaccuracy was present in 60% of the responses, if we consider the partially correct responses. The exciting aspect was that ChatGPT was able to admit mistakes and correct them when presented with counterarguments. It was also capable of challenging incorrect prompts and premises. The performance of ChatGPT in the context of lacrimal drainage disorders, at best, can be termed average. However, the potential of this AI chatbot to influence medicine is enormous. There is a need for it to be specifically trained and retrained for individual medical subspecialties.