Abstract
Purpose: :
Expert evaluation of visual field data in the Ocular Hypertension Treatment Study has identified 17 typical visual field defects. The purpose of this study was to use unsupervised statistical learning techniques to evaluate the intrinsic dimensionality of visual field defects in chronic open angle glaucoma.
Methods: :
Threshold standard automated perimetry data were evaluated from the worse eye (greater Mean Deviation) of 220 participants in a longitudinal evaluation of early glaucoma. Non-linear dimensionality reduction was performed using Isomap, a technique that combines unsupervised clustering, principal component analysis and multidimensional scaling. We computed the residual variance of the data as a function of the number of dimensions used to model them. The number of dimensions needed to represent these data was compared with expert evaluations of these same data.
Results: :
The age distribution of the subjects (median and interquartile range-IQR) was 63 (55 to 72) years with a median follow-up of 72 (36 to 96) months. The distribution of Mean Deviation in this sample (median and IQR) was 0.87 (-0.88 to 2.23). The residual variance with a single dimension was 12%. Four dimensions were needed to account for more than 95% of the variance, which continued to gradually decrease thereafter. Graphical analysis of the clustered data demonstrates good correspondence between spatial location of visual field defects and identified clusters.
Conclusions: :
Considerably fewer than 17 dimensions were needed to capture the characteristic visual field defects in this cohort of subjects with early glaucoma. This analysis can be extended to a longitudinal analysis to track disease progression.
Keywords: visual fields • computational modeling • clinical research methodology