Comprehensive Summary
This study, presented by Shin et. al., examined the application of AI system Glandy CAS on diagnosing active thyroid eye disease (TED) using images of the participant's face to determine Clinical Activity Scores. This machine learning model was tested using a sample of 756 periocular photos while three primary ophthalmologists also analyzed the data to provide a relative comparison of accuracy. With a marker of CAS greater than or equal to three, Glandy CAS determined that 207 of the 756 participant photos had active TED such that the reported F1 score was higher for Glandy CAS analysis (F1 = 0.88) than ophthalmologist inferences (F1 = 0.57). Glandy CAS also had much greater sensitivity scores with an identification accuracy of 87.9% compared to the ophthalmologists who correctly named 60.4% of cases. The AI system provided much more accurate and effective diagnostic ability compared to traditional ophthalmologists who had five or less years of experience in the field. Glandy CAS has high potential for improving interpretability of symptoms for TED patients and narrowing the black box in diagnosing patients.
Outcomes and Implications
This new machine learning method presents a diagnostic tool that could be potentially more effective and accurate than physician inference, especially for patients suffering from active TED. The study remains very clinically relevant given that the device has been authorized for use (Product Licence No. 24-93) in medicine by the Korea Ministry of Food and Drug Safety. The ability for early detection of TED and the relatively inexpensive method of collecting data (facial photographs) are promising indicators of Glandy CAS’s use in the real world.