Purpose
Glaucomatous visual field loss follows characteristic patterns which are related to the retinal nerve fiber geometry. Over the past decades, numerous mostly qualitative classification schemes for the patterns have been proposed. Here, we try to find quantitative mathematical models for the classification of glaucomatous functional damage.
Methods
We apply the statistical learning procedures Independent Component Analysis (ICA), Cluster Analysis (ClA), and Archetypal Analysis to a set of 13,231 reliable (FL≤33%, FN≤20%) Humphrey Visual Fields (HFV 24-2) of glaucoma patients or suspects from a large clinical glaucoma practice. Left eye locations were mirrored for data analysis to match right eye locations. No repeated measurements over same eye and patient were included. We compare our mathematical classification schemes to the 17 patterns defined in the Ocular Hypertension Treatment Study (OHTS; Keltner et al., Arch. Ophthalmol. 121, 2003, 643-50).
Results
ICA and ClA yielded patterns not consistent with clinical observations. However, when patterns of visual field loss are learned on the convex hull of the data space (Archetypal Analysis), they closely resemble the clinically derived visual field classification scheme of OHTS. In our solution with 17 subtypes, 16 of them match the exact definitions of patterns given in OHTS (Fig. 1), while the remaining subtype represents the normal visual field. Unlike the OHTS scheme, however, our approach can serve as a framework to quantify the various subtypes of glaucomatous visual field loss. Fig. 2 illustratively shows that the model we proposed can be used to decompose any HVF 24-2 into these 17 subtypes quantitatively.
Conclusions
We show that typical patterns observed by clinical practitioners can be extracted by purely mathematical procedures that are agnostic to the ophthalmological background. Our approach can serve as an adjunct to global indices and provides a way to objectively quantify functional loss in glaucoma.
Keywords: 758 visual fields