Cornea. 2017 Jun;36(6):689-695. doi: 10.1097/ICO.0000000000001194.
To validate a recently developed program for automatic and objective keratoconus detection (Keratoconus Assistant [KA]) by applying it to a new population and comparing it with other methods described in the literature.
KA uses machine learning and 25 Pentacam-derived parameters to classify eyes into subgroups, such as keratoconus, keratoconus suspect, postrefractive surgery, and normal eyes. To validate this program, it was applied to 131 eyes diagnosed separately by experienced corneal specialists from 2 different centers (Fondation Rothschild, Paris, and Antwerp University Hospital [UZA]). The agreement of the KA classification with 7 other indices from the literature was assessed using interrater reliability and confusion matrices. The agreement of the 2 clinical classifications was also assessed.
For keratoconus, KA agreed in 92.6% of cases with the clinical diagnosis by UZA and in 98.0% of cases with the diagnosis by Rothschild. In keratoconus suspect and forme fruste detection, KA agreed in 65.2% (UZA) and 100% (Rothschild) of cases with the clinical assessments. This corresponds with a moderate agreement with a clinical assessment (κ = 0.594 and κ = 0.563 for Rothschild and UZA, respectively). The agreement with the other classification methods ranged from moderate (κ = 0.432; Score) to low (κ = 0.158; KISA%). Both clinical assessments agreed substantially (κ = 0.759) with each other.
KA is effective at detecting early keratoconus and agrees with trained clinical judgment. As keratoconus detection depends on the method used, we recommend using multiple methods side by side.