Purchase this article with an account.
P. A. Sample, G. Jang, T.-P. Jung, J. Hao, C. Bowd, L. M. Zangwill, J. Liebmann, C. A. Girkin, R. N. Weinreb, M. Goldbaum; Unsupervised Machine Learning With Independent Component Analysis Identifies Patterns of Glaucomatous Visual Field Loss in SITA Fields. Invest. Ophthalmol. Vis. Sci. 2009;50(13):5283.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Our goal is to develop a more sensitive method to identify glaucoma progression. We previously showed that variational Bayesian independent component analysis-mixture model (VIM), a type of unsupervised machine learning, could separate normal full threshold visual fields from fields with different patterns of loss and by severity in patients with glaucoma1, which then allowed development of a promising progression algorithm.2 Because SITA, the current clinical standard, uses a different thresholding protocol that could impact the convergence of VIM, we tested VIM on a six times larger dataset of SITA fields from the Diagnostic Innovations in Glaucoma Study (DIGS) and African Descent and Evaluation Study (ADAGES).
One field was used from each of 939 patients with repeatable abnormal SITA-standard fields (glaucoma Hemifield test "outside normal limits" or pattern standard deviation at 5% or worse) and 1,146 normal eyes with normal fields. The input set into VIM was 52 absolute threshold values in dB from program 24-2 plus age. Without knowledge of class identity, VIM separated the fields into clusters while simultaneously positioning a set of statistically independent axes through the mean of each cluster. The optimum number of axes was determined and a limit was set beyond which no additional information was observed.
VIM separated the fields into 3 clusters, Cluster N contained primarily normals (1089/1146; specificity 95%) and clusters G1 and G2 that together contained primarily glaucomatous fields (835/939; sensitivity 88.9%). For Clusters N and G1 the optimum number of axes was 2 each and for G2 it was 5. Post-hoc analysis found the patterns generated along axes in G1 were very mild and diffusely affected. Patterns at + 2 and -2 standard deviation directions from the G2 mean revealed 10 more specific patterns similar to those identified previously1 and by experts as indicative of glaucoma1. SITA fields assigned to a given axis showed increasing severity as they were located farther from the normal mean.
VIM successfully identified glaucomatous patterns of loss in SITA fields. Pattern severity captured by VIM provided the information needed to update the individualized unsupervised progression algorithm2, Progression of Patterns (POP), for use with SITA.1. IOVS 46:3676-83, 2005, 2. IOVS 46:3684-92, 2005
This PDF is available to Subscribers Only