Purpose: Neural networks can recognize patterns and classify complex variables. We assessed the ability of neural networks to discriminate between normal and glaucomatous eyes by using structural and functional measurements.
Methods: Several neural network algorithms were tested with a database of 185 eyes of patients with early glaucomatous visual field loss (average mean defect, 4.5 dB) and 54 eyes of age-matched normal control subjects. The information used included automated visual field indices (mean defect, corrected loss variance, and short-term fluctuation) and structural data (cup/disk ratio, rim area, cup volume, and nerve fiber layer height) from computerized image analysis.
Results: A back propagation network with two intermediate layers assigned an estimated probability of being glaucomatous to each eye and correctly identified 88% of all eyes with 90% sensitivity and 84% specificity. The same neural network trained with only structural data correctly identified 80% of the eyes with 87% sensitivity and 56% specificity, and when trained with functional data only, it correctly identified 84% of the eyes with 84% sensitivity and 86% specificity.
Conclusion: Analysis of several optic nerve and visual field variables by neural networks can help identify early glaucomatous damage and assign an estimated probability that early damage is present in individual patients.