Glaucoma Monitoring Using Manifold Learning and Unsupervised Clustering

We developed an artificial-intelligence-enabled software for monitoring eyes with glaucoma using manifold learning and unsupervised clustering. A total of 31,591 visual fields (VF) measurements from 8,077 subjects were acquired using the Humphrey Field Analyzers instrument. The two locations closest to the blind spot were excluded from each VF. The number of remaining VFs were 13,231 with 52 VF test locations (features). We first applied principal component analysis (PCA) to linearly reduce the number of dimensions from 52 to four significant principal components. We then developed a manifold learning algorithm to identify VFs with similar patterns of VF loss. Manifold learning preserved the local characteristics of the input principal components and nonlinearly reduced the dimensions further. Finally, we developed an unsupervised density-based clustering to identify clusters at different stages of glaucoma as well as different patterns of VF loss. We evaluated the quality of learning using both subjective visualization of clusters and objective validation using global VF parameters including mean deviation (MD) and pattern standard deviation (PSD). The proposed tool could be highly useful in clinical practice and glaucoma research for monitoring and staging glaucoma.