Abstract
Purpose :
To evaluate the accuracy of a novel multilevel model in estimating focal rates of change of standard automated perimetry (SAP) using population data from the Bascom Palmer Glaucoma Repository.
Methods :
In this retrospective study, we extracted SAP data from patients with or suspected of glaucoma from the Bascom Palmer Glaucoma Repository. Reliable 24-2 and 30-2 SAP tests were defined as those with false positives ≤15% and fixation losses ≤33%. Only eyes with ≥5 reliable tests were used. We analyzed point sensitivity values to calculate overall rates of change, with random effect components at the eye and cluster level, the latter defined by glaucoma hemifield test (GHT) sectors. We developed a novel two-fold random slopes model based on the theory of two-fold small area estimation. The multilevel nature of the model allows for borrowing strength across individuals and connected focal clusters when estimating rates. To account for spatial structure, we included a two-dimensional fixed effects surface smoother, resulting in a two-fold random slope generalized additive mixed model (GAMM). We modeled the data using ordinary least squares (OLS) regression for comparison. Right (OD) and left (OS) eyes were analyzed separately to maintain spatial anatomy. Mean squared errors (MSEs) of the rates at the cluster level were estimated for OLS and GAMM.
Results :
We modeled focal visual loss over time using the first 5 tests of 250 OD and 250 OS eyes. OLS MSEs were 20.92 (OD) & 21.75 (OS). For the spatial GAMM, MSEs were 1.59 (OD) & 2.17 (OS) representing a substantial reduction of 93% & 90% for OD & OS respectively (p<0.001). Contour plots of the estimated spatial structure for OD & OS are shown in Fig. 1. Side-by-side 95% confidence interval plots comparing OLS and GAMM estimates for two sample GHT clusters are shown in Fig. 2. These plots demonstrate shrinkage of intervals due to the two-level random effects in the GAMM.
Conclusions :
The GAMM demonstrated clear superiority in estimating rates of change for focal clusters with SAP testing. It also captured meaningful, physiologic residual spatial structures not identified by traditional OLS.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.