Assessment of Glaucoma Progression Using the Dynamic Structure-Function Model with Permutation Analysis

Abstract

 

Purpose

 

The purpose of this study was to develop an individualized approach to assess glaucoma progression using structural and functional data jointly.

 

Methods

 

The dynamic structure-function (DSF) model was used: centroids estimating the current state of the disease and velocity vectors estimating the rate of change over time of paired SF data were calculated (Figure 1). We ran the DSF model on 798 eyes from 452 glaucoma patients and suspects from the Diagnostic Innovations in Glaucoma Study or the African Descent and Glaucoma Evaluation Study who had longitudinal paired rim area (RA) and mean deviation (MD) data. Each eye had 7 SF pairs (mean follow-up of 6.9±1.3 years). We used permutation analysis to test for progression at a siginificance level of 5%. Thus, progression was defined as velocity vectors that pointed towards the third quadrant (decrease in MD and RA) and whose length was outside the 80% CI (to reach target significance of 5%). We compared the results of the DSF to those obtained with PoPLR which applies permutation analysis to functional data only (O’Leary et al, IOVS, 2012;53:6776-84), and to conventional trend analyses for MD, Pattern Standard Deviation (PSD), Visual Field Index (VFI), and RA. Global indices and PoPLR results were obtained with the R open-source package visualFields. Pairwise agreement analysis between the DSF and each approach was assessed using the Kappa statistic (κ; 95% CI obtained with bootstrap).

 

Results

 

The percentage of eyes showing progression was 8% (DSF), 10% (PoPLR), 23% (MD), 15% (PSD), 15% (VFI) and 25% (RA). Poor agreement was found between the DSF and PoPLR (κ=0.04; -0.03, 0.12), MD (κ=0.13; 0.06, 0.20), PSD (κ=0.05; -0.02, 0.12), VFI (κ=0.02; -0.04, 0.10) and RA (κ=0.01; -0.04, 0.07). The Venn diagrams presented in Figure 2 illustrate the agreement.

 

Conclusions

 

While only poor agreement was found between the DSF with permutations and other methods, the DSF model is consistent with these methods in that it identifies only a relatively small percentage of eyes as progressing over time. Future work will focus on predicting which eyes are likely to progress and at which rate.

 

 

Figure 1. An illustration of the DSF model is presented with paired SF points over time (T), centroids (C), velocity vectors (V) and quadrants (Q).

 

 

Figure 2. Venn diagrams showing the agreement expressed in percentages between each of the progression methods.

 

Keywords: 473 computational modeling