Detection of subclinical keratoconus using a novel combined tomographic and biomechanical model based on an automated decision tree.

Early detection of keratoconus is a crucial factor in monitoring its progression and making the decision to perform refractive surgery. The aim of this study was to use the decision tree technique in the classification and prediction of subclinical keratoconus (SKC). A total of 194 eyes (including 105 normal eyes and 89 with SKC) were included in the double-center retrospective study. Data were separately used for training and validation databases. The baseline variables were derived from tomography and biomechanical imaging. The decision tree models were generated using Chi-square automatic interaction detection (CHAID) and classification and regression tree (CART) algorithms based on the training database. The discriminating rules of the CART model selected metrics of the Belin/Ambrósio deviation (BAD-D), stiffness parameter at first applanation (SPA1), back eccentricity (Becc), and maximum pachymetric progression index in that order; On the other hand, the CHAID model selected BAD-D, deformation amplitude ratio, SPA1, and Becc. Further, the CART model allowed for discrimination between normal and SKC eyes with 92.2% accuracy, which was higher than that of the CHAID model (88.3%), BAD-D (82.0%), Corvis biomechanical index (CBI, 77.3%), and tomographic and biomechanical index (TBI, 78.1%). The discriminating performance of the CART model was validated with 92.4% accuracy, while the CHAID model was validated with 86.4% accuracy in the validation database. Thus, the CART model using tomography and biomechanical imaging was an excellent model for SKC screening and provided easy-to-understand discriminating rules.