Assessment of ChatGPT in the preclinical management of ophthalmological emergencies – an analysis of ten fictional case vignettes

Dominik Knebel, Siegfried Priglinger, Nicolas Scherer, Jakob Siedlecki, Benedikt SchwormmedRxiv 2023 · DOI

Abstract

Background: The AI platform ChatGPT has shown promise in general medical triage, but its performance in specialized fields like ophthalmology remains under-evaluated. Purpose: To assess ChatGPT’s triage accuracy, appropriateness of preclinical recommendations, and potential for harm in ophthalmological emergencies. Methods: Ten fictional case vignettes were entered into ChatGPT five times each using a standardized interaction pathway. Results: Triage accuracy was 87.2%, but 32% of responses contained potentially harmful advice. Conclusion: ChatGPT is not yet safe as a stand-alone source for acute ophthalmological symptoms.

Introduction

In ophthalmological emergencies, timely professional consultation is critical to preventing vision loss. As patients increasingly turn to AI for medical advice, understanding the reliability of tools like ChatGPT is essential. This study evaluates whether a text-based Large Language Model (LLM) can accurately triage urgent eye conditions and provide safe preclinical guidance.

Methodology

The researchers developed a standardized testing framework using ten fictional case vignettes. These vignettes were short, one-sentence descriptions of acute ocular symptoms, ranging from elective conditions to critical emergencies.

Table 1: Fictional case vignettes and assigned urgency levels used in the study.

To account for the stochastic nature of AI outputs, each vignette was entered five times following a standardized three-step interaction pathway. Responses were evaluated by experts for triage accuracy, diagnostic correctness, and the Appropriateness of Preclinical Measures (APM) using a 5-point ordinal scale.

Results

The study found a significant disconnect between the AI's general accuracy and its clinical safety.

Triage and Diagnostic Accuracy

ChatGPT demonstrated a high triage accuracy of 87.2%, correctly identifying the need for professional care in most cases. However, diagnostic accuracy was lower at 61.5%. While the AI often recommended seeing a doctor "as soon as possible," it frequently lacked the nuance to distinguish between varying levels of urgency. Notably, treatment advice was 100% accurate only when the diagnosis was correctly identified.

Potential for Harm and the Disclaimer Paradox

A critical finding was that 32% of responses contained potentially harmful advice, such as inadequate instructions for chemical burns. Furthermore, while 62% of responses included medical disclaimers, statistical analysis revealed no correlation between these disclaimers and the safety or accuracy of the advice.

Figure 1: Scatter plots showing the lack of correlation between disclaimers and accuracy or harm.

Discussion

The visual nature of ophthalmology presents a unique challenge for text-based LLMs. The study highlights that ChatGPT struggles with conditions requiring physical inspection and lacks the nuance necessary for high-stakes triage. While the AI is an impressive tool, its current potential for harm makes it unsuitable as a primary source for emergency eye care.

Conclusion

ChatGPT should not be used as a stand-alone source for managing acute ophthalmological symptoms. Regular reassessment is required as AI technology continues to evolve.

Summary updated

Abstract

Abstract Background/Aims The artificial intelligence (AI) based platform ChatGPT (Chat Generative Pre-Trained Transformer, OpenAI LP, San Francisco, CA, USA) has gained an impressing popularity over the past months. Its performance on case vignettes of general medical (non-ophthalmological) emergencies has priorly been assessed with very encouraging results. The purpose of this study is to assess the performance of ChatGPT on ophthalmological emergency case vignettes in terms of the main outcome measures triage accuracy, appropriateness of recommended preclinical measures and overall potential to inflict harm to the user/patient. Methods We wrote ten short, fictional case vignettes describing different acute ophthalmological symptoms. Each vignette was entered into ChatGPT five times with the same wording and following a standardized interaction pathway. The answers were analysed in a standardised manner. Results We observed a triage accuracy of 87.2%. Most answers contained only appropriate recommendations for preclinical measures. However, an overall potential to inflict harm to users/patients was present in 32% of answers. Conclusion ChatGPT should not be used as a stand-alone primary source of information about acute ophthalmological symptoms. As AI continues to evolve, its safety and efficacy in the preclinical management of ophthalmological emergencies has to be reassessed regularly.
BibTeX
@misc{knebel2023assessment,
  author = {Dominik Knebel and Siegfried Priglinger and Nicolas Scherer and Jakob Siedlecki and Benedikt Schworm},
  title = {{Assessment of ChatGPT in the preclinical management of ophthalmological emergencies – an analysis of ten fictional case vignettes}},
  year = {2023},
  doi = {10.1101/2023.04.16.23288645},
  howpublished = {medRxiv},
}
Figure