Purchase this article with an account.
Stephanie Reeves, Tobias Elze, Michael Sandberg, Carol Weigel-DiFranco, Russell L Woods; Patterns of Visual Field Loss in Retinitis Pigmentosa found using Unsupervised Machine Learning of Goldmann Perimetry. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4322.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Retinitis pigmentosa (RP) is a well-documented degenerative retinal disease that results in substantial visual field (VF) loss with characteristics patterns. There have been few systematic evaluations of the patterns of VF loss of people with RP. We applied unsupervised machine learning to a large sample of Goldmann data to identify representative patterns of VF loss.
Retrospective analysis of monocular VFs measured using a Goldmann Perimeter with the V4e stimulus (maximal sensitivity). We obtained 4,661 measured VFs of 435 subjects with RP (18 to 85 years) who participated in four previous studies and were followed for up to 15 visits over 12.6 years. The VFs were divided into 208 discrete sectors (Fig. 1) for analysis using Mixture of Gaussian Clustering, a probabilistic cluster analysis method. The model was trained using a subset of 866 exams, with each subject contributing only one VF from each eye. An optimal set of representative clusters was determined by the conservative Bayesian Information Criterion, which prevents "overfit" (too many clusters). Logistic mixed-effect models were implemented to investigate changes in cluster allocation with age and hereditary pattern.
There was an optimal solution of 8 clusters. Fig. 2 shows the corresponding representative patterns (cluster means). Each VF was assigned to one of the eight clusters. For the tunnel-vision clusters, the likelihood of being assigned to patterns #6, 7 and 8 increased with increasing age (p<0.001), but not for cluster #3 (p=0.56). For the peripheral-island patterns, #1 and 2, likelihood increased with increasing age (p<0.001). For the large-area patterns, #4 and 5, the likelihood decreased with increasing age (p<0.001). X-linked subjects had a lower likelihood of being assigned to clusters #1 and 4, and the likelihood changed more slowly with increasing age. Whereas, X-linked subjects had higher likelihood and likelihood increased more quickly with age for patterns #2 and 3.
This method of describing VF patterns provides insight into the vision available to people with RP and allows statistical analyses of change over time and between groups. This allows for a more nuanced tracking of change or evaluations of differences than total area alone.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only