Use of a Support Vector Machine for Keratoconus and Subclinical Keratoconus Detection by Topographic and Tomographic Data

Purpose To define a new classification method for the diagnosis of keratoconus based on corneal measurements provided by a Scheimpflug camera combined with Placido corneal topography (Sirius, CSO, Florence, Italy). Design Retrospective case series. Participants We analyzed the examinations of 877 eyes with keratoconus, 426 eyes with subclinical keratoconus, 940 eyes with a history of corneal surgery (defined as abnormal), and 1259 healthy control eyes. Methods For each group, eyes were divided into a training and a validation set. A support vector machine (SVM) was used to analyze the corneal measurements and classify the eyes into the 4 groups of participants. The classifier was trained to consider the indices obtained from both the anterior and posterior corneal surfaces or only from the anterior corneal surface. Main Outcome Measures Symmetry index of front and back corneal curvature, best fit radius of the front corneal surface, Baiocchi Calossi Versaci front index (BCVf) and BCV back index (BCVb), root mean square of front and back corneal surface higher order aberrations, and thinnest corneal point were analyzed. The diagnostic performance of the classifier was evaluated. Results The accuracy of the classifier was excellent both with and without the data generated from the posterior corneal surface and corneal thickness because the number of true predictions was greater than 95% and 93%, respectively, in all classes. Precision improved most when posterior corneal surface data were included, especially in cases of subclinical keratoconus. Using the data from both anterior and posterior corneal surfaces and pachymetry allowed the SVM to increase its sensitivity from 89.3% to 96.0% in abnormal eyes, 92.8% to 95.0% in eyes with keratoconus, 75.2% to 92.0% in eyes with subclinical keratoconus, and 93.1% to 97.2% in normal eyes. Conclusions The classification algorithm showed high accuracy, precision, sensitivity, and specificity in discriminating among abnormal eyes, eyes with keratoconus or subclinical keratoconus, and normal eyes. Including the posterior corneal surface and thickness parameters markedly improved the sensitivity in the diagnosis of subclinical keratoconus. Classification may be particularly useful in excluding eyes with early signs of corneal ectasia when screening patients for excimer laser surgery. Financial Disclosure(s) Proprietary or commercial disclosure may be found after the references. To define a new classification method for the diagnosis of keratoconus based on corneal measurements provided by a Scheimpflug camera combined with Placido corneal topography (Sirius, CSO, Florence, Italy). Retrospective case series. We analyzed the examinations of 877 eyes with keratoconus, 426 eyes with subclinical keratoconus, 940 eyes with a history of corneal surgery (defined as abnormal), and 1259 healthy control eyes. For each group, eyes were divided into a training and a validation set. A support vector machine (SVM) was used to analyze the corneal measurements and classify the eyes into the 4 groups of participants. The classifier was trained to consider the indices obtained from both the anterior and posterior corneal surfaces or only from the anterior corneal surface. Symmetry index of front and back corneal curvature, best fit radius of the front corneal surface, Baiocchi Calossi Versaci front index (BCVf) and BCV back index (BCVb), root mean square of front and back corneal surface higher order aberrations, and thinnest corneal point were analyzed. The diagnostic performance of the classifier was evaluated. The accuracy of the classifier was excellent both with and without the data generated from the posterior corneal surface and corneal thickness because the number of true predictions was greater than 95% and 93%, respectively, in all classes. Precision improved most when posterior corneal surface data were included, especially in cases of subclinical keratoconus. Using the data from both anterior and posterior corneal surfaces and pachymetry allowed the SVM to increase its sensitivity from 89.3% to 96.0% in abnormal eyes, 92.8% to 95.0% in eyes with keratoconus, 75.2% to 92.0% in eyes with subclinical keratoconus, and 93.1% to 97.2% in normal eyes. The classification algorithm showed high accuracy, precision, sensitivity, and specificity in discriminating among abnormal eyes, eyes with keratoconus or subclinical keratoconus, and normal eyes. Including the posterior corneal surface and thickness parameters markedly improved the sensitivity in the diagnosis of subclinical keratoconus. Classification may be particularly useful in excluding eyes with early signs of corneal ectasia when screening patients for excimer laser surgery.