Evaluation of a Machine-Learning Classifier for Keratoconus Detection Based on Scheimpflug Tomography. Issue 6 (June 2016)

• Record Type: Journal Article
• Authors: Ruiz Hidalgo, Irene; Rodriguez, Pablo; Rozema, Jos J.; Ní Dhubhghaill, Sorcha; Zakaria, Nadia; Tassignon, Marie-José; Koppen, Carina
• Is Part Of: Cornea. Volume 35:Issue 6(2016)
• Publication Date: 2016-06
• Subjects: cornea -- keratoconus -- pentacam -- detection -- classification; Cornea -- Periodicals; Cornea -- Periodicals; Cornée -- Périodiques; 617.719
• DOI: 10.1097/ICO.0000000000000834
• Languages: English
• ISSNs: 0277-3740

Abstract

Abstract : Purpose: To evaluate the performance of a support vector machine algorithm that automatically and objectively identifies corneal patterns based on a combination of 22 parameters obtained from Pentacam measurements and to compare this method with other known keratoconus (KC) classification methods. Methods: Pentacam data from 860 eyes were included in the study and divided into 5 groups: 454 KC, 67 forme fruste (FF), 28 astigmatic, 117 after refractive surgery (PR), and 194 normal eyes (N). Twenty-two parameters were used for classification using a support vector machine algorithm developed in Weka, a machine-learning computer software. The cross-validation accuracy for 3 different classification tasks (KC vs. N, FF vs. N and all 5 groups) was calculated and compared with other known classification methods. Results: The accuracy achieved in the KC versus N discrimination task was 98.9%, with 99.1% sensitivity and 98.5% specificity for KC detection. The accuracy in the FF versus N task was 93.1%, with 79.1% sensitivity and 97.9% specificity for the FF discrimination. Finally, for the 5-groups classification, the accuracy was 88.8%, with a weighted average sensitivity of 89.0% and specificity of 95.2%. Conclusions: Despite using the strictest definition for FF KC, the present study obtained comparable or better results than the single-parameter methods and indices reported in the literature. In some cases, direct comparisons with the literature were not possible because of differences in the compositions and definitions of the study groups, especially the FF KC. Abstract : Supplemental Digital Content is Available in the Text.