The role of artificial intelligence in detecting the progression of keratoconus

Knowing the progression in advance, as well as early diagnosis of keratoconus disease, is important in determining the treatment decision and follow-up interval. In this study, we aimed to predict the risk of progression by evaluating the data of progressive and non-progressive keratoconus patients at the first examination with artificial intelligence models. Our study included 562 eyes of 325 patients with diagnosis of keratoconus at Eskişehir Osmangazi University Faculty of Medicine Ophthalmology Clinic. Patient files were examined retrospectively. Keratoconus cases were divided into two groups: those showing progression and those not showing progression. The patients' age, gender, follow-up period, history related to possible risk factors, blood vitamin B12, folate and homocysteine levels, and visual acuity, intraocular pressure, refraction, Pentacam and ORA measurement values were recorded. We used Pentacam and ORA numerical data for machine learning models (XGB, LGBM, AVG Blender, Extra Trees, Random Forest and Logistic regression); images of color-coded corneal topography maps of Pentacam in 3 different areas (whole, 5 mm and 3 mm) for deep learning models (YOLOv8L and YOLOv8X). The most successful among machine learning models was the XGB model, which had a ROC/AUC score of 0.683 and an accuracy score of 0.667. In deep learning model, the YOLOv8L model showed 76% success in training on the front elevation map of the whole corneal area. While the test success rate on the posterior axial sagittal curvature map of the YOLOv8L model in 3 mm corneal area was 70% and the sensitivity was 74%; YOLOv8X model on the posterior tangential curvature map was 70% and its sensitivity was 92%. In this study, we achieved artificial intelligence model success with moderate reliability in predicting and/or early detection of keratoconus progression. As a result, higher success can be achieved with artificial intelligence studies that include more reproducible parameters related to keratoconus progression, use more than one image of the same patient in training, and include a large number of patient data.