Harnessing generative pre-trained transformer technology for clinical decision support in retinal detachment

Aim: Considering the increasing incorporation of artificial intelligence (AI) in healthcare, it is crucial to comprehend the advantages and constraints of these technologies within ophthalmologic settings for their secure and efficient clinical utilization. This study aims to comprehensively assess the efficacy of three leading Generative Pre-trained Transformer (GPT) -based platforms in providing clinical decision-support for retinal detachment (RD). Methods: This cross-sectional comparative study was conducted between April 2024 and May 2024. Fifty questions were created based on the American Academy of Ophthalmology “Retina Book”, specifically targeting RD. The answers were produced by three different platforms and assessed by three independent reviewers who used Likert scales to evaluate their comprehensiveness and accuracy. Six readability metrics, including the Flesch-Kincaid Grade Level (FKGL) and Flesch Reading Ease Score (FRES), average words per sentence, average syllables per word, total sentence count, and total word count, were assessed. Results: Gemini earned the most outstanding results for comprehensiveness (4.11±0.72) and accuracy (1.49±0.61), followed by ChatGPT and Copilot. ChatGPT had superior readability metrics, achieving an FKGL of 15.62±2.85 and a FRES of 62.54±12.34, establishing it as the most accessible platform. ChatGPT demonstrated significantly higher performance compared to other platforms in the metrics of average syllables per word (p=0.0421) and total word count (p=0.0115). At the same time, no significant differences were found among the platforms in the metrics of average words per sentence (p=0.0842) and total sentence count (p=0.1603). Intraclass correlation coefficient (ICC) values indicated strong inter-rater agreement for comprehensiveness (ICC >0.74) and moderate to-high agreement for accuracy (ICC >0.56). Conclusion: Gemini’s detailed and accurate responses position it as a robust tool for professional use, while ChatGPT’s superior readability makes it suitable for patient education. These findings emphasize the synergistic advantages of AI platforms in research and development management and show the necessity for hybrid systems that integrate accessibility with accuracy.