Unsupervised machine learning identifies patterns of glaucoma-related binocular visual field loss

Glaucoma researchers developing machine-learning techniques to identify patterns of glaucoma-related visual field loss, without human supervision, and to detect progression of glaucoma based on these patterns are reporting their experience in applying their methods to binocular visual fields.


By Cheryl Guttman Krader; Reviewed by Christopher Bowd, PhD

San Diego-In previous research, Christopher Bowd, PhD, and colleagues at the University of California, San Diego, Hamilton Glaucoma Center, demonstrated that an unsupervised machine-learning based classifier-the variational Bayesian independent component analysis mixture model (VIM)-could be applied to monocular standard automated perimetry (SAP) visual fields and monocular frequency-doubling technology (FDT) perimetry visual fields to identify patterns of glaucomatous damage without foreknowledge of diagnosis and without human intervention.

Changes in these patterns over time proved to be a promising method for describing glaucomatous progression, and according to recent evidence, VIM, trained on monocular SAP visual fields, was better at detecting glaucomatous progression than current methods.

Taking their investigations forward, they determined that VIM successfully identified binocular patterns of glaucomatous defects in SAP visual fields and successfully separated patients with glaucoma from unaffected controls.

“Our studies show that this method is able to discriminate between glaucomatous and healthy individuals and can automatically find statistically different binocular visual field patterns in glaucoma patients,” said Dr. Bowd, senior research scientist, director of the Hamilton Glaucoma Center-based Visual Field Assessment Center and co-director of the Hamilton Glaucoma Center-based Imaging Data Evaluation and Analysis Center, Department of Ophthalmology, University of California, San Diego.

“Patterns identified by unsupervised classifiers are more objectively determined than those observed and described by experts, so they are less biased by previous experience and rules of thumb,” he said.


He explained that the visual image formed in the brain is based on visual information received from both the eyes. Therefore, it was important to evaluate the performance of the technique for analyzing binocular visual fields because they are more strongly associated with daily activities, such as driving, and also are a better indicator of overall quality of life than the previously studied monocular visual fields.

“Binocular visual fields better reflect the effect of losing part of the entire useful visual field than monocular visual fields and, as the visual field deteriorates, change in binocular visual fields will better reflect the corresponding increase in disability,” Dr. Bowd said.

The study included monocular visual fields from both eyes of 543 glaucoma patients and 560 healthy controls. To be included in the glaucoma cohort, patients had to have repeatable abnormal SAP results by Pattern Standard Deviation or the Glaucoma Hemifield Test in at least one eye.

Study participants in the glaucoma cohort had a visual field mean deviation (MD) = -6.15 ± 6.70 dB in their worst eye. Participants were included as healthy controls if they had SAP results within normal limits in both eyes. Healthy controls had a visual field MD in the worst eye of -0.54 ± 1.17 dB.

The study participants were selected from individuals enrolled in the UCSD-based Diagnostic Innovations in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES).

Binocular visual fields were constructed through averaging of data from fellow eyes at each visual field location. Input for the VIM analysis included the 56 threshold test points from the constructed binocular visual field and age.

Using the data, a total of 720 unsupervised VIM models were generated. The best model had a specificity of 94% and a sensitivity of 80%. That model separated the binocular visual fields into five clusters or classes, one containing primarily the healthy participants and the other four representing glaucoma patients with increasing levels of severity in the defect pattern that may correlate with the patient’s ability to perform everyday tasks.


“Future research will investigate the correlation of cross-sectional and progressing binocular visual field defects with quality of life measures and simulated driving tasks. In addition, we’d like to investigate and compare the performance of VIM with more recently developed unsupervised algorithms,” Dr. Bowd said.



Christopher Bowd, PhD

E: cbowd@ucsd.edu

This article was adapted from Dr. Bowd’s presentation at the 2014 meeting of the Association for Research in Vision and Ophthalmology. Dr. Bowd did not indicate any proprietary interest in the subject matter.