Using hierarchical cluster analysis to cluster 52 visual field points

Abstract

Purpose : Hierarchical cluster analysis (HCA) is an unsupervised learning technique that groups similar observations into relatively homogeneous clusters. Our goal was to determine the optimal clusters of visual field (VF) points using HCA when accounting for the association of their threshold values with retinal nerve fiber layer (RNFL) clock-hours.

Methods : Subjects with qualified OCT (Cirrus HD-OCT, Zeiss, Dublin, CA) scans and VF (Humphrey field analyzer, SITA standard 24-2 protocol; Zeiss) tests were included. All records used for analysis had a VF mean deviation (MD) < -2 dB and an average RNFL > 45 mm. HCA was used to cluster individual VF threshold values and their correlation with OCT RNFL clock hours measurements, using their correlation matrix. The optimal number of clusters was estimated via within-cluster sum of squares based on 10-fold cross validation. Manhattan distance was calculated for HCA and the best linkage method was chosen by cophenetic correlation.

Results : A total of 5789 records from 1131 subjects (1851 eyes) were analyzed with an average RNFL of 71.7 +/- 14.0 mm and a median MD of -5.68 (-12.11, -3.27) dB. The mean age was 65.6 +/- 13.4 years. The optimized number of clusters were 4 for the superior hemifield and 3 in the inferior (see Figure).

Conclusions : We reported the clustering of VF points based on the association between their threshold values and RNFL clock hours. The VF points within each cluster had similar correlation with the clock hours in the corresponding hemifield.

This is a 2020 Imaging in the Eye Conference abstract.