Identifying Clinically Useful Markers in Glaucoma Suspects and Primary Open Angle Glaucoma Patients Using a Machine Learning J48 Decision Tree

Abstract

Purpose : Machine learning offers the potential to identify subclinical parameters that extend beyond the threshold of human pattern recognition. Here, we utilize a J48 decision tree to determine if specific variables from optical coherence tomography (OCT) and Humphrey Visual Fields (HVF) can assist in classifying glaucoma suspects and eyes with primary open angle glaucoma.

Methods : A retrospective analysis of 1,530 eyes from glaucoma suspects (GS) and 1,050 eyes with primary open angle glaucoma (POAG) was conducted using WEKA (University of Waikato, New Zealand) machine learning software and a J48 classifier were used to generate a decision tree. OCT retinal nerve fiber layer (RNFL), ganglion cell (GC), and macular thickness measurements were obtained using commercial spectral-domain OCT (Cirrus HD-OCT, Zeis, Dublin, CA). Eyes were excluded if OCT imaging signal strength was <7, HVF fixation losses, false positives, or false negatives were >33%, or if pathology other than glaucoma was present.

Results : There were statistically significant differences (p<0.05) in visual field index (VFI) and thicknesses of the RNFL, GC, central macula, and outer retina (Table 1). The data included 56 attributes and were randomly split into 90% training and 10% testing. A J48 pruned decision tree with 17 leaves and a minimum leaf size of 30 correctly classified 1,649 out of 2,580 instances (overall accuracy of 63.95%, Figure 1). In the setting of a thinner RNFL at clock hour 11 (≤64 µm) on OCT and a thinner inferonasal GC layer (≤66 µm), the algorithm was able to correctly identify 127 out of 162 (78.4%) eyes with POAG. For eyes with a thicker RNFL at clock hour 11 (≥64 µm) on OCT and a lower VFI (≤60), the tree was able to utilize the superotemporal outer retinal thickness to correctly classify 33 out of 40 eyes (82.5%) with POAG.

Conclusions : A J48 machine learning decision tree using HVF and OCT data was generated to classify eyes as GS or POAG with an overall accuracy of 63.95%. The tree also identified particular sets of objective variables that had a larger impact on the classification process. Further machine learning studies with longitudinal data are needed for higher predictive values.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Table 1: OCT and HVF 24-2 data for glaucoma suspects and eyes with primary open angle glaucoma.

 

Figure 1: A J48 pruned decision tree demonstrating how eyes were classified as GS or POAG.