Glaucoma progression detection using structural retinal nerve fiber layer measurements and functional visual field points
Siamak Yousefi, Michael H. Goldbaum, Madhusudhanan Balasubramanian, Tzyy-Ping Jung, Robert N. Weinreb, Felipe A. Medeiros, Linda M. Zangwill, Jeffrey M. Liebmann, Christopher A. Girkin, and Christopher Bowd University of California San Diego, University of Memphis, New York University, University of Alabama
Volume 61, Issue 4, Page:1143-1154
Machine Learning Classifiers (MLCs) are widely used in biomedical research and clinical applications, including detection of glaucoma and glaucomatous progression. This work employs various supervised MLCs to enhance detection of glaucomatous progression using optical imaging (spectral domain optical coherence tomography, SDOCT) and visual function (standard automated perimetry, SAP) measurements.
MLCs (Bayesian Net, Lazy K Star, Meta-classification Regression, Meta-ensemble Selection, Alternating Decision Tree, Random Forest Tree, CART) were trained and tested on seven SDOCT retinal nerve fiber layer (RNFL) thickness parameters (six sectors surrounding the optic nerve and average thickness) and 54 SAP sensitivity parameters (sensitivity at 52 test points and two global indices) from progressed (n=107, defined based on standard clinical reference assessment) and stable glaucoma (n=73, tested multiple times over five weeks) eyes, using 10-fold cross-validation. Glaucomatous progression was detected in a study eye when at least 50% of individual follow-up exams or at least two consecutive follow-up exams were identified as progressing from baseline by the MLCs. Diagnostic accuracy was assessed using several performance metrics including Mathews correlation coefficient, area under the ROC curve and area under the precision recall curve. Accuracy of RNFL features alone was better than accuracy of SAP features alone. The accuracy of combined RNFL and SAP features was similar to the accuracy of RNFL features alone. This was the case for all classifiers investigated. Performance of the simple Bayesian Net classifier was as good as that of the more complicated Meta- or Tree-based classifiers.
Further, all RNFL and SAP features were combined and performance of features was ranked using both independent and dependent feature selection methods. By both feature-ranking methods, the same RNFL features were ranked 1-4 of the 10 best total features, providing further evidence that, in our sample, RNFL measurements are superior to SAP sensitivity for identifying eyes showing glaucomatous progression.