Cornea · 2026
Reproducibility of Tear Ferning Test Classification by Human Examiners and Artificial Intelligence Models: A Comparative Study.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Felberg Sérgio, Bernardes Guilherme D'Agosto, Dos Santos Felipe Augusto Casseb, Della Paolera Lucas, Sato Elcio Hideo, Moscovici Bernardo Kaplan, Dantas Paulo
Аннотация
PURPOSE: To compare the reproducibility of tear ferning classification by human evaluators and 3 multimodal artificial intelligence platforms using the 4-grade Rolando scale and a predefined binary scheme, and to clarify its clinical role.
METHODS: Eighty polarized-light microscopy tear-film images were independently graded (Rolando grades I-IV) by 3 ocular surface researchers and 3 platforms (ChatGPT, Claude, Gemini), each evaluated with an identical prompt and image order in single-run conditions. Grades I-II were defined a priori as normal and III-IV as abnormal.
RESULTS: For the 4-grade scale, Fleiss kappa was 0.65 (humans), 0.44 (6 raters), and 0.37 (platforms); binary reclassification raised these to 0.76, 0.66, and 0.66. Against consensus, Claude agreed best (binary agreement 93.8%, kappa 0.87; 4-grade 82.5%, weighted kappa 0.84, 95% confidence interval, 0.75-0.91), within the interhuman range. ChatGPT and Gemini showed lower 4-grade agreement (weighted kappa 0.48 each) and significantly underclassified abnormalities (McNemar P = 0.0004 and 0.00006), whereas Claude showed no such asymmetry (P = 0.3750).
CONCLUSIONS: Binary classification was more reproducible than the 4-grade Rolando scale for humans and platforms, but it is a pragmatic standardization strategy, not a replacement for ordinal grading. Claude best matched the human benchmark, whereas ChatGPT and Gemini were less reproducible and systematically underclassified abnormalities. Platform assistance may have adjunctive value, but clinical use will require external validation, accuracy and repeated-session testing, and safety evaluation.
METHODS: Eighty polarized-light microscopy tear-film images were independently graded (Rolando grades I-IV) by 3 ocular surface researchers and 3 platforms (ChatGPT, Claude, Gemini), each evaluated with an identical prompt and image order in single-run conditions. Grades I-II were defined a priori as normal and III-IV as abnormal.
RESULTS: For the 4-grade scale, Fleiss kappa was 0.65 (humans), 0.44 (6 raters), and 0.37 (platforms); binary reclassification raised these to 0.76, 0.66, and 0.66. Against consensus, Claude agreed best (binary agreement 93.8%, kappa 0.87; 4-grade 82.5%, weighted kappa 0.84, 95% confidence interval, 0.75-0.91), within the interhuman range. ChatGPT and Gemini showed lower 4-grade agreement (weighted kappa 0.48 each) and significantly underclassified abnormalities (McNemar P = 0.0004 and 0.00006), whereas Claude showed no such asymmetry (P = 0.3750).
CONCLUSIONS: Binary classification was more reproducible than the 4-grade Rolando scale for humans and platforms, but it is a pragmatic standardization strategy, not a replacement for ordinal grading. Claude best matched the human benchmark, whereas ChatGPT and Gemini were less reproducible and systematically underclassified abnormalities. Platform assistance may have adjunctive value, but clinical use will require external validation, accuracy and repeated-session testing, and safety evaluation.