Validation Of Deep Learning-Based Clinical Decision Support System To Assist Evaluation Of Dry Eye Disease Severity

To evaluate the efficacy of a deep learning-based clinical decision support system (CDSS) in grading the severity of dry eye disease (DED)

Seoul National University Hospital, Seoul National University Bundang Hospital, Seoul Metropolitan Government Seoul National University Boramae Medical Center

A total of 48 anterior segment images, evaluated by three cornea specialists using the National Eye Institute (NEI) scale, were selected as the validation set. The images represented each NEI score from 0 to 15, with 3 images per score. To validate the utility of the developed deep learning system as a CDSS, a reader study was conducted. A total of 13 ophthalmologists evaluated the 48 anterior segment images based on the NEI scale in two sessions, separated by a one-week washout period. In session 1, DED grading was performed using only the images, while in session 2, grading was conducted with the assistance of the CDSS.

The ICC values increased across all zones, with improvements ranging from 0.0638 to 0.1424. Accuracy also improved significantly with CDSS in all zones, ranging from 0.0481 to 0.1090. When analyzed by zone, the highest enhancement of ICC was observed in zone 5 (inferior part of the cornea), while the greatest improvement in accuracy was seen in zone 1 (center part of the cornea). Additionally, when comparing ICC and accuracy between senior (clinical experience ≥5Y) and junior ophthalmologists (clinical experience <5Y), no significant difference was found in ICC and accuracy between session 1 and session 2.

The CDSS improves accuracy and inter-examiner consistency by reducing variability in individual assessments when evaluating the severity of DED. This suggests that the CDSS could contribute to the standardization of DED severity grading, thereby enhancing clinical decision-making for DED treatment.