Purchase this article with an account.
Janelle Tong, David Alonso-Caneiro, Michael Kalloniatis, Barbara Zangerl; Applying cluster analysis to macular OCTs for prediction of visual field defects in glaucoma. Invest. Ophthalmol. Vis. Sci. 2021;62(8):3353.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a normative model of the macular ganglion cell-inner plexiform layer (GCIPL) and evaluate its accuracy in prediction of visual field (VF) defects in patients with glaucoma.
Macular OCTs were acquired for 493 healthy and 37 glaucoma participants using the Spectralis OCT. GCIPL measurements from the healthy cohort were used to develop a normative model incorporating hierarchical cluster analysis, to identify locations with similar structural properties under considerations of normal variations in optic disc tilt position and aging characteristics. Glaucoma participants underwent 10-2 and 30-2 VFs using the Humphrey Field Analyzer, and GCIPL locations were binarized to VF-normal and VF-defective based on a pattern deviation cut-off of <2% for the corresponding VF location. Receiver operator characteristic (ROC) analyses based on the normative model were conducted to generate GCIPL cut-off values per cluster using Youden’s criterion. Global and cluster-specific accuracy of the model was assessed by applying derived GCIPL cut-offs to individual locations and comparing areas under the ROC curves (AUROCs) respectively.
The normative model resulted in 10 quasi-concentric clusters (Figure 1A) that largely co-locate despite various fovea to optic disc tilts, reflecting robust classification. The model demonstrated moderate global sensitivity and specificity in predicting VF defects from GCIPL thicknesses, at 0.75 and 0.76 respectively. AUROCs were significantly different between clusters (P<0.0001, Welch’s ANOVA), with central clusters demonstrating higher accuracy than peripheral clusters (P=0.009-0.04, Figure 1B).
Normative models applying cluster analysis principles can predict VF defects from macular GCIPL measurements with moderate accuracy. However, spatial variability in peripheral macular locations were associated with poorer accuracies, suggesting that OCT alone may be insufficient to holistically predict macular function.
This is a 2021 ARVO Annual Meeting abstract.
A. A 10 cluster normative model derived from hierarchical cluster analysis applied to GCIPL measurements in the healthy cohort, where locations of the same color indicate those showing statistically similar structural properties. Multiple data locations reflect various fovea to optic disc tilts. B. ROC curves describing GCIPL measurements corresponding to VF defects in the glaucoma cohort for each cluster, as per the normative model.
This PDF is available to Subscribers Only