Official ESCRS | European Society of Cataract & Refractive Surgeons


Comparison of machine learning algorithms in diagnosis of subclinical keratoconus

Search Title by author or title

Session Details

Session Title: Keratoconus

Session Date/Time: Monday 16/09/2019 | 14:00-16:00

Paper Time: 15:48

Venue: Free Paper Forum: Podium 2

First Author: : D.Smadja ISRAEL

Co Author(s): :    D. Chen   B. Chiu   M. Santhiago   A. Abulafia   D. Zadok   J. Young              

Abstract Details


To compare the performance of different machine learning algorithms in differentiating subclinical keratoconus from subjects with keratoconus and normal corneas.


Ophthalmology Department, Refractive Unit, Vision Institute, Shaare Zedek Medical Center, Jerusalem, Israel New York University, Department of Ophthalmology, New York, NY, USA


Retrospective case-control study using data from 177 normal eyes, 47 eyes with forme fruste, and 148 keratoconus were included. Corneal measurements were performed using the GALILEI Dual Scheimpflug Analyzer System. Attribute parameters were extracted and analyzed using the WEKA machine learning platform (University of Waikato). Algorithms compared included J48 (C4.5) decision tree classifier, Random Forest and Support Vector Machine (SMO) in the classification of keratoconus, subclinical keratoconus, and normal corneas.


Among machine learning algorithms, Random Forest reached the highest discriminating performance by correctly classifying 339 instances and incorrectly classified 33 instances, with an accuracy of 91.1%. The pruned J48 classifier tree correctly classified 334 instances and incorrectly classified 38 instances, with an accuracy of 89.8%. SMO Support Vector Machine correctly classified 332 instances and incorrectly classified 40 instances, with an accuracy of 89.2%.


Machine learning algorithms can diagnose keratoconus, subclinical keratoconus, and normals with high levels of accuracy. In our dataset, the Random Forest algorithm provided the highest accuracy in identifying subclinical keratoconus, possibly due to its tendency to minimize overfitting of data, though differences among all the algorithms were very small.

Financial Disclosure:

receives consulting fees, retainer, or contract payments from a company producing, developing or supplying the product or procedure presented, travel has been funded, fully or partially, by a company producing, developing or supplying the product or procedure presented

Back to previous