Posters

Comparison of machine learning methods to automatically detect keratoconus based on Pentacam examinations

Poster Details

First Author: I.Ruiz Hidalgo BELGIUM

Co Author(s):    P. Rodriguez Perez   J. Rozema   C. Koppen   M. Tassignon           

Abstract Details

Purpose:

Given the strong clinical similarities between normal eyes and eyes with subclinical or forme fruste keratoconus, the screening of patients for refractive surgery can sometimes be challenging. This problem could be solved by using computer programs that automatically and objectively classify eyes into different groups based on a combination of parameters. This work aims to evaluate and compare the performance of 3 machine learning algorithms that identify keratoconus and forme fruste keratoconus based on data obtained from Pentacam measurements.

Setting:

Department of Ophthalmology, Antwerp University Hospital

Methods:

This case-control study analyzes 541 parameters provided by the Pentacam for 785 eyes (620 subjects), of which 519 eyes were keratoconus (KC), 71 forme fruste (FF), and 195 normal eyes (N). Two computer programs (Matlab and Weka) were used to perform classification tasks using three different algorithms (Naive Bayes (NB), Discriminant Analysis (DA) and Support Vector Machine (SVM). Within each group, 33% of the subjects were randomly selected for testing and the remaining to train the classifiers. Finally, accuracy was estimated applying the classifier to the test group and also using a cross-validation (CV) method on the training set.

Results:

Three different classification tasks were performed. The highest accuracy in discriminating keratoconic from normal eyes was achieved by a SVM algorithm implemented in Matlab (99.6%). Between the FF and N groups the highest accuracy was 85.9% again with SVM in Matlab, and between all three groups (N, FF and KC), the accuracy was 93.6% with a SVM in Weka. In every classification problem the SVM using the full dataset (541 variables) always reached the highest accuracy.

Conclusions:

This study demonstrates a higher diagnostic accuracy of the SVM classifier using a large number of parameters derived from the Pentacam, compared with other algorithms such as NB and LDA with and without dimension reduction. Accuracy, precision, sensitivity, specificity and area under the curve (AUC) obtained in this study are comparable to or higher than the single parameter methods and indices that have been published in the literature. However, in most cases direct comparisons are not possible due to differences in the compositions of the study groups and in the definitions of forme fruste keratoconus.

Financial Disclosure:

NONE