Abstract
Purpose :
Artificial intelligence (AI) has been widely used to identify subgroups of patients with different disease prognosis. In this study, we aimed to identify and characterize a subtype of glaucoma patients with the highest risk of more rapid visual field (VF) progression.
Methods :
We applied a k-meansclustering algorithm to visual fields (VFs) collected from 331 eyes of 258 patients with glaucoma to identify different clusters of eyes with similar VF patterns at the glaucoma onset visit. We employed Silhouette metricto find the optimal number of clusters. To assess the stability of clusters, we randomly selected 80% of data 100 times and performed clustering and computed the cluster memberships each time then crosschecked the outcome with that of the initial clustering. We then computed the slope of mean deviation (MD) of the eyes in each cluster and characterized clusters (Fig. 1). We identified MD thresholds that separated the cluster with more rapid progression from clusters with slow progression using the Bayes minimum errorclassifier (Fig. 2).
Results :
We identified three clusters with mean MD of 0.0 dB, -2.2 dB, and -6.0 dB at the onset visit (p-value <= 0.05). The first two clusters had a significantly lower rate of MD decline compared to the third cluster (p-value <= 0.05). Bayes identified that MD threshold of -3.9 dB can separate eyes in the third cluster from eyes in the other two clusters. This may suggest patients with MD worse than -3.9 dB at the conversion visit are at higher risk of more rapid progression and subsequent vision loss.
Conclusions :
We identified a subtype of patients with glaucoma that were at higher risk of more rapid VF progression. The identified MD threshold may assist clinicians to identify patients with higher likelihood of future vision loss.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.