Abstract
Purpose :
To develop an artificial map for monitoring glaucoma progression using manifold learning and unsupervised clustering
Methods :
We acquired 31,591 VF tests from 8,077 subjects using the Humphrey Field Analyzer. We selected 13,231 VFs at the baseline visit for training the machine learning framework. We first applied principal component analysis (PCA) to linearly reduce the number of dimensions of the VFs. We then applied manifold learning using t-distributed stochastic neighbor embedding (tSNE) to identify VFs with similar local characteristics on the tSNE manifold. We then developed unsupervised clustering to identify clusters of eyes densely located on the tSNE map (see Fig. 1). These clusters represent eyes at different stages of glaucoma as well as different patterns of VF loss (see Fig. 2). We used this map as a reference artificial screen to stage and monitor glaucoma. Finally, we evaluated the method using subjective visualization of clusters and objective validation using mean deviation (MD) and pattern standard deviation (PSD).
Results :
The mean age of subjects selected for training the machine learning framework was 60.6 years (SD= 17.9). Sample eyes with confirmed progression from an independent datasets with over eight years of follow-up with 16 VF tests consistently followed the direction of glaucoma progression on the artificial tSNE map (subjective evaluation). Other sample eyes with confirmed non-progression (eyes tested once a week for 10 consecutive weeks) consistently represented no progression on the artificial tSNE map (subjective evaluation). The artificial tSNE map showed a greater dynamic range (from normal to advanced glaucoma) compared with the MD versus PSD map (see Fig. 1).
Conclusions :
We demonstrate that glaucoma functional severity and progression can be readily monitored using an artificial-intelligence-enabled map. The proposed tool could be highly useful in clinical practice and glaucoma research for staging and monitoring glaucoma on a user-friendly map.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.