Artificial Intelligence Chatbot-Driven Detection of Keratoconus From Corneal Tomography

To evaluate the performance of ChatGPT-5 in detecting keratoconus (KC) from corneal tomography maps.

ChatGPT-5 (Instant mode) was prompted in September 2025 with an open-source dataset of corneal tomography maps from Scheimpflug imaging systems (Pentacam), graded by 3 corneal specialists as KC, KC suspect, or normal. The primary analysis evaluated the model's ability to distinguish KC from normal eyes, with reported metrics including sensitivity, specificity, accuracy, positive predictive value, and negative predictive value. Secondary analyses assessed its ability to distinguish (1) KC or KC suspect eyes from normal eyes and (2) KC eyes from KC suspect or normal eyes.

A total of 200 KC, 173 KC suspect, and 200 normal eyes were analyzed. In the primary analysis (KC vs. normal), ChatGPT-5 achieved a sensitivity of 96.5%, specificity of 77.0%, accuracy of 86.8%, positive predictive value of 80.8%, and negative predictive value of 95.7%. In the secondary analysis, when KC and KC suspect eyes were grouped together and compared with normal eyes, sensitivity and specificity declined to 74.3% and 75.5%, respectively. In contrast, when distinguishing KC eyes from either KC suspect or normal eyes, the chatbot exhibited relatively higher sensitivity (93.0%) but lower specificity (62.5%).

ChatGPT-5 demonstrated promising performance in rapidly distinguishing KC from normal eyes based on corneal tomography maps; however, its performance declined when KC suspects were included, suggesting greater reliability in detecting definitive cases. Future research should aim to validate the chatbot's utility for KC screening within frameworks that prioritize expert oversight and patient privacy.