advertisement

Editors Selection IGR 14-1

Clinical examination methods: Machine classifiers, neural networks and progression

Comment by William Swanson on:

13164 Using unsupervised learning with independent component analysis to identify patterns of glaucomatous visual field defects, Goldbaum MH; Sample PA; Zhang Z et al., Investigative Ophthalmology and Visual Science, 2005; 46: 3676-3683

Find related abstracts

Machine classifiers have the potential to produce alternative new indices for visual fields based on quantifying patterns of glaucomatous damage. Companion papers from the UCSD Hamilton Glaucoma Center provide a first step in this direction. Goldbaum et al. (944) used a machine classifier to analyze visual fields from 345 eyes, half free of disease and half with glaucomatous optic neuropathy. This was 'unsupervised learning', in that the classifier operated on the raw sensitivity values (and the age) without information concerning visual field abnormality or diagnosis. This yielded two clusters: Cluster 1 contained one-third of the eyes (primarily those with visual field abnormalities); and Cluster 2 contained the remaining eyes (primarily those free from visual field abnormality). Component analysis of Cluster 1 yielded six relatively independent axes, or ways of quantifying patterns of visual loss. The primary axis reflected general depression, but also reflected small localized defects.

Machine classifiers may provide new indices for progression

The remaining axes reflected arcuate scotomas, nasal steps and hemifield depressions, with each axis emphasizing some form of asymmetry (superior/inferior, nasal/temporal, macular/peripheral). However, the axes did not distinguish between these different patterns of loss: an axis that is influenced by arcuate defects can also be influenced by nasal steps or hemifield defects.

Sample et al. (941) evaluated the model by using the six axes to assess progression in an independent group of 191 eyes, of which one-third had a diagnosis of glaucoma and the rest were suspects or ocular hypertensives. Patients were followed for a mean of six years, and were considered to have progressed if at least one of the axes showed change greater than the 95% confidence limits while at least two axes did not. This is a lenient criterion, which would be expected to yield a false positive rate of 26% (the authors did not assess false positive rates).

Goldbaum's axes take a value of zero at the mean of Cluster 1, which contains mild losses that tend to increase nasally. Therefore distance away from the mean is not a direct index of degree of abnormality, and a finding of significant change may be difficult to interpret. As the authors note, both improvement and progression can cause change along an axis. However, any significant change was scored as progression, without determining whether the change was due to improvement.

• ← Previous
• Next →