Detection of Subclinical Keratoconus Using an Automated Decision Tree Classification

• Published in: American journal of ophthalmology Vol. 156; no. 2; pp. 237 - 246.e1
• Main Authors: Smadja, David, Touboul, David, Cohen, Ayala, Doveh, Etti, Santhiago, Marcony R., Mello, Glauco R., Krueger, Ronald R., Colin, Joseph
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.08.2013; Elsevier Limited
• Subjects: Algorithms; Artificial Intelligence; Automation; Case-Control Studies; Classification; Corneal Topography - methods; Decision Support Techniques; Decision Trees; Humans; Keratoconus - classification; Keratoconus - diagnosis; Medical imaging; Ophthalmology; Retrospective Studies; Sensitivity and Specificity; Surgery; Tomography
• ISSN: 0002-9394; 1879-1891; 1879-1891
• DOI: 10.1016/j.ajo.2013.03.034

To develop a method for automatizing the detection of subclinical keratoconus based on a tree classification. Retrospective case-control study. setting: University Hospital of Bordeaux. participants: A total of 372 eyes of 197 patients were enrolled: 177 normal eyes of 95 subjects, 47 eyes of 47 patients with forme fruste keratoconus, and 148 eyes of 102 patients with keratoconus. observation procedure: All eyes were imaged with a dual Scheimpflug analyzer. Fifty-five parameters derived from anterior and posterior corneal measurements were analyzed for each eye and a machine learning algorithm, the classification and regression tree, was used to classify the eyes into the 3 above-mentioned conditions. main outcome measures: The performance of the machine learning algorithm for classifying eye conditions was evaluated, and the curvature, elevation, pachymetric, and wavefront parameters were analyzed in each group and compared. The discriminating rules generated with the automated decision tree classifier allowed for discrimination between normal and keratoconus with 100% sensitivity and 99.5% specificity, and between normal and forme fruste keratoconus with 93.6% sensitivity and 97.2% specificity. The algorithm selected as the most discriminant variables parameters related to posterior surface asymmetry and thickness spatial distribution. The machine learning classifier showed very good performance for discriminating between normal corneas and forme fruste keratoconus and provided a tool that is closer to an automated medical reasoning. This might help in the surgical decision before refractive surgery by providing a good sensitivity in detecting ectasia-susceptible corneas.