Purchase this article with an account.
Jessica Nevins, David Danks, Gadi Wollstein, Hiroshi Ishikawa, Larry Kagemann, Ian Sigal, Joel Schuman; Machine Classifier Clustering of Ocular Structure Measurements Poorly Corresponds with Longitudinal Functional Performance in Glaucoma. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5916.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Structure and function measurements are both used to assess glaucomatous damage. The purpose of this study was to evaluate the association between longitudinal structural and functional measurements.
5330 visits from 694 healthy, glaucoma suspect, and glaucomatous eyes were used. Each eye had at least two visits with visual field (VF) tests and at least two ocular imaging devices obtained in each visit: scanning laser polarimetry (GDx with enhanced corneal compensation and with variable corneal compensation), confocal scanning laser ophthalmoscopy (HRT), time-domain (Stratus) and spectral-domain (RTVue) optical coherence tomography. Four different machine learning classifier methods were used to cluster the eyes based on all structural parameters. Two methods (k-means and k-medoids) find compact, convex clusters defined by center points in parameter space. The other two methods (hierarchical agglomerative and divisive) build clusters through iterative grouping and splitting, and so can yield complex, non-convex clusters. These clusters were tested for correlation with functional measurements from VF tests.
The mean baseline age of the study population was 59.2 (range: 16.5-85.1) years with a mean follow-up length of 5.0 (0.4-7.8) years and a mean of 8 (2-24) visits. Clustering within each device was stable with a minimal correlation coefficient =0.85 and half the correlations >0.95 by comparing clusters created when using all the data compared to subsamples of the data. Clustering between devices was highly correlated with a minimal correlation coefficient =0.80 and over half the correlations >0.90. There was poor correspondence with predicting the VF measurements at the next visit or the rates of VF measurement change with a maximum correlation =0.50 and over half the correlations <0.30.
The presence of structural changes was confirmed by the good agreement among clusters based on structure, regardless of the algorithm used for clustering. Nevertheless, structure clusters were dissociated from longitudinal functional changes. It is unknown if this is because of a lack of association between structure and function or because different clusterings are needed between them.
This PDF is available to Subscribers Only