Visual Field (VF) change-based archetype analysis for early-stage glaucoma detection

Abstract

Purpose : In order to diagnose glaucomatous visual field (VF) loss, summary measurements are used that do not reveal any information about particular glaucomatous patterns. Here we investigate the possible patterns in visual field changes that may help to detect Glaucoma in early stages.

Methods : We use Archetype Analysis (AA) to find possible patterns in visual field loss (1^st minus 2^nd visit) which can be the indicator of Glaucoma in early stages. Total and mean deviations (TD, MD) of the most recent reliable VFs (SITA Standard 24-2) were retrospectively chosen among all patients from Mass. Eye and Ear glaucoma service. Along with reliability check, patients with at least three visits were selected when their first VF was almost normal (MD in [-3dB:+1dB] and GHT normal/borderline) and their second VF occurred within 2 years with an MD decline <3dB. Early-glaucoma label for each patient determined based on GHT flip to outside normal limits in future visits. VF change archetypes were calculated for a model with 6 (6-AT) and a model with 9 (9-AT) archetypes (ATs). The dataset was split into a training (75%) and testing (25%) set. In addition, random forest classifier (RFc) for early-glaucoma detection was performed and also used for permutation-based feature importance analysis to identify important ATs.

Results : 2,106 eyes of 1852 patients fulfilled our selection criteria. Figures 1A/1B show the solutions with six/nine ATs from the training set. The 9-AT solution contains all patterns from the 6-AT solution extended by three additional patterns. Furthermore, Figures 1C/1D show the corresponding archetype importance computed by RFc. While the accuracy of the trained classifier for detecting early-glaucoma is different regarding to f1-score over the test-set (6-AT: 69.04% vs 9-AT: 73.19%, p < 0.001), the AT importance is almost the same between them over the train-set. AT3 in 6-AT and ATs 6,7 in 9-AT are the most ATs because of preserving the mean. Interesting patterns are AT1 and AT6 (in 6-AT, correspondingly AT2 and AT8 in 9-AT) which were found important in both scenarios.

Conclusions : We show that archetype analysis of VF changes within 2 years may be beneficial for detecting and classifying the onset of glaucomatous VF loss after only two visits. We also identified two relevant VF-change ATs predictive of the future onset of VF loss. A larger multi-clinical dataset is planned for future work and verification.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.