Abstract
Purpose :
To evaluate whether the identification of distinct subpopulations among glaucoma patients improves the estimates of future perimetric loss.
Methods :
Eyes with ≥5 reliable visual fields were identified in the Duke Glaucoma Registry. Standard automated perimetry (SAP) mean deviation (MD) and visual field index (VFI) values were collected with associated timepoints. The dataset was split at the patient level, with 80% used for model training and 20% for testing. Rates of change were modeled using latent class mixed models (LCMM) and ordinary least square (OLS) regression. LCMM can identify subpopulations clustered around different average rates. Bayesian Information Criteria was used to identify the LCMM with the optimal number of classes. The LCMM was presented with the first 3, 4, 5, & 6 visual fields of eyes from the test set to predict subsequent SAP MD or VFI values. Model performance was compared to OLS with mean square prediction error (MSPE) using Wilcoxon signed-rank test. Analysis was completed in R.
Results :
The full dataset contained 52,900 visual fields from 6,558 eyes of 3,981 subjects, with an average of 8.1±3.7 visual fields per eye. The optimal LCMM for SAP MD contained 3 classes with rates of -0.08, -0.17, & -1.33 dB/year (84.0%, 11.4%, & 4.6% of the population respectively). The optimal LCMM for SAP VFI contained 4 classes with rates of -0.21, -0.66, -2.96, & -6.77%/year (82.6%, 10.1%, 5.7%, & 1.5% of the population respectively). For both SAP MD and VFI, MSPE was significantly lower with LCMM compared to OLS regardless of the number of tests used (p<0.001 for all comparisons; Figure 1). Notably, MSPE from LCMM was significantly lower than that of OLS when using only the first 3 visual fields of test eyes (4.90 vs. 60.19 for MD respectively, p<0.001; Figure 2). MSPE of fast progressors was significantly lower with LCMM versus OLS (27.30 & 79.59 respectively, p= 0.01).
Conclusions :
LCMM successfully identified different groups of progressors in a large glaucoma population and significantly improved the prediction of future test estimates. LCMM had superior accuracy for future prediction even when using a small number of visual fields, suggesting that it could more easily identify eyes at risk for significant visual field loss.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.