Abstract
Purpose :
Visual field monitoring of patients with primary open-angle glaucoma (POAG) is critical to identify disease progression. The statistical power to detect visual field progression from POAG is influenced by long-term test variability, which is not consistent among patients. Predicting long-term test variability for a given patient would help determine the frequency of visual field testing needed for that patient. Baseline visual field mean deviation (MD) is associated with long-term test variability. We hypothesized that a Bayesian model accounting for the association between baseline MD and MD variability would better predict a patient’s long-term visual field test variability than typically used ordinary least-squares (OLS) linear regression.
Methods :
We identified patients with POAG and at least 6 Standardized Automated Perimetry (SAP) visual field tests in the Duke Glaucoma Registry. We iteratively divided the study population into 80% training groups and 20% testing groups. For the training groups, we estimated the long-term variability by calculating the OLS variance (mean squared error) for each patient using all their data points. We trained a Bayesian model where the prior for variance was based on a regression equation including baseline MD and baseline MD squared using all observations. For the testing groups, we established ground truth variance for each patient by estimating the OLS variance using all of their data points. We then used only the first 3 observations for the patients to predict the variance using both the Bayesian and OLS models. We compared predicted variances to the ground truth variance by calculating absolute error (AE). We compared mean AE for the Bayesian and OLS models and used bootstrapping to test statistical significance.
Results :
We identified 4719 tests of 586 eyes of 434 patients with POAG that met inclusion criteria. The mean AE for the Bayesian estimates was 0.38 dB (standard deviation [SD]: 0.84 dB, median 0.17 dB) compared to 0.91 dB (SD: 1.09 dB, median 0.57 dB, p<0.001) for OLS regression.
Conclusions :
Accounting for the association of baseline MD with long-term variability using a Bayesian model improved prediction of long-term variability. This could improve our ability to determine the frequency of visual field testing needed for a given patient with fewer measures than with traditional OLS models and lead to earlier detection of POAG progression.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.