Abstract
Purpose :
Retinitis pigmentosa (RP) is a well-documented retinal degeneration. However, there have been few systematic evaluations of the patterns of visual field (VF) loss of people with RP. We applied unsupervised machine learning to a large sample to calculate representative patterns of RP related VF loss.
Methods :
VFs were measured using the 30-2 and 30/60-1 patterns of the Humphrey Field Analyzer. We analyzed 138 VF locations (combining the two patterns) from 411 eyes from 214 patients (aged 18 to 55 years) using Mixture of Gaussian Clustering, which is a probabilistic cluster analysis method. The data set was partitioned into K different subsets ("clusters") in which the similarity within each cluster was higher than the similarity between the clusters. An optimal set of representative clusters was determined by the Bayesian Information Criterion, a conservative criterion to prevent "overfit" (too many clusters). Each patient’s VF can be found from the combination of the K patterns using a linear combination of that patient’s coefficients for each pattern.
Results :
There was an optimal number of K=13 clusters. The two figures show the corresponding 13 representative patterns (cluster means) and the number of eyes for which that pattern was the main component in our sample. Many patterns were variants of “tunnel vision” with different diameters and others included peripheral islands of residual vision. Most patterns were consistent with clinical experience. Cluster 5 showed a superior peripheral island that was unexpected in its frequency.
Conclusions :
This method of decomposing the VF components provides insight into the degenerative process of RP. For example, pattern 2 would be expected to transition into pattern 6. As our scheme of components allows the quantitative decomposition of newly measured VFs, it will allow for tracking of change over time. The proportion assigned to each component for each patient is expected to change with time as RP progresses. It will allow the quantification of the value of VF components to measures of vision-related ability.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.