Feasibility of Identifying High-Risk Glaucoma Patients Before the Onset of Disease

Abstract

Purpose : This study investigates the potential for identifying high-risk glaucoma patients despite the presence of normal mean deviations (MD) on visual field (VF) tests, presenting a challenging prognosis scenario.

Methods : Our study retrospectively analyzed SITA Standard 24-2 visual field data from patients at the Mass. Eye and Ear glaucoma service. To be included, patients required at least three visits with reliable tests and normal GHT and MD values (-2.1 to +2.1dB) in their first two visits. Patients with subsequent GHT values that fell outside the normal range or became borderline were categorized as high-risk or converters. This study focuses on identifying converters using only the information from the first two visits, employing two well-known methods for tabular data: XGBoost and TabNet, a deep learning technique. The dataset was split 75/25 for training and testing, with a 20% validation sub-set. Feature importance analyses were performed on both the training set (using built-in methods) and the test set (using permutation importance) to identify the most influential features. As shown in Figure 2A, the predictive features from the baseline and 2nd visits are race, values of age, MD, PSD and archetypes decomposition coefficients of baseline total deviations (TDs) and their changes. Archetype analysis was employed to identify potential representative patterns within the baseline TDs (8 archetypes) and their changes (2^nd visit-1^st visit, 11 archetypes).

Results : A total of 6,686 eyes from 5,467 patients met our study criteria, of which 1,874 eyes were identified as converters. Figure 1 illustrates the characteristics of the data. Using all features, converter identification performance (F1-score TabNet 65% vs. XGBoost 64%, Fisher's exact p-value <0.05 (Figure 2A)), demonstrates the potential for identification despite the inherent challenges. However, significant discrepancies in the feature importance rankings were observed between methods, as depicted in Figures 2B to 2E. This highlights the complexity of the task and the need for further investigation to identify the most crucial factors for converter prediction.

Conclusions : Our study shows potential for early identification of high-risk glaucoma patients, but further research to improve accuracy (e.g., by incorporating genetics, retina images, and IOP measurements/events), investigate uncertainty, and address variations in feature importance is needed.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.