Unsupervised learning with independent component analysis can identify patterns of glaucomatous visual field defects

Area of Science:

• Ophthalmology
• Machine Learning
• Data Mining

Background:

• Standard automated perimetry (SAP) is crucial for glaucoma diagnosis.
• Unsupervised learning offers potential for analyzing complex SAP data.
• Previous work utilized machine learning classifiers for SAP pattern segmentation.

Purpose of the Study:

• To evaluate information represented by axes derived from independent component analysis (ICA) of SAP patterns.
• To apply unsupervised learning via ICA to segment glaucoma-related visual field patterns.

Main Methods:

• Utilized Humphrey Visual Field Analyzer data from 189 normal eyes and 156 eyes with glaucomatous optic neuropathy (GON).
• Employed variational Bayesian independent component analysis mixture model (vB-ICA-mm) for clustering and axis decomposition.
• Masked review with stereoscopic optic disc photos confirmed GON diagnosis.

Main Results:

• vB-ICA-mm identified two informative clusters: 'G' (68.6% GON eyes) and 'N' (98.4% normal eyes).
• Cluster G optimally contained six axes, revealing glaucoma-indicative defects at varying severity levels.
• Increasing distance along positive axes correlated with increased visual field defect severity.

Conclusions:

• vB-ICA-mm successfully represented SAP fields with patterns meaningful to glaucoma experts.
• The technique effectively captured visual field defect severity.
• vB-ICA-mm is validated as a data mining tool for complex visual field tests.