Inter-Eye Symmetries of Visual Field Defects in Glaucoma

Abstract

Purpose : To study the inter-eye symmetries of visual field (VF) defects through artificial intelligence.

Methods : From a multi-centered dataset, we selected the most recent reliable 24-2 Humphrey VF pairs from both eyes tested on the same date with mean deviation (MD) ≥ -12 dB. VF defects were quantified by decomposing VFs into 16 archetype (AT) patterns consisting of 1 normal (AT 1) and 15 defect patterns (ATs 2 to 16), as detailed in our prior work (Fig. 1a). Coefficients describing the contribution of each VF AT to the overall VF pattern were correlated between worse and better eyes as defined by MD. The presence of VF defects (AT coefficients > 10%) in the better eye was modeled from the AT coefficients of the worse eye by logistic regression. The optimal models selected by stepwise regression were evaluated by the area under the receiver operating characteristic curve (AUC). We specifically provide model details for inter-eye prediction of the most common type of glaucomatous central vision loss: superior paracentral loss (AT14).

Results : 64,447 pairs of VFs from 64,447 patients (61.0 ± 16.6 years) were selected. The MDs for the worse and better eyes (r = 0.77, p < 0.001) were -3.4 ± 3.2 dB and -1.6 ± 2.3 dB, respectively. The inter-eye correlations of AT coefficients ranged from -0.34 to 0.70 (Fig. 1b), with strongest correlations either with the same or normal VF pattern. The AUCs predicting the presence of 15 VF defects in the better eye based on ATs of the worse eye ranged from 0.69 (AT 3, superonasal step) to 0.92 (AT 6, near total loss). Fig. 2a shows that AT 14’s presence in the better eye can be predicted by 9 AT features in the worse eye (AUC = 0.89): AT 7 (central scotoma), AT 8 (superior altitudinal defect), AT 14 (superior paracentral loss), AT 15 (nasal hemianopia) and AT 16 (inferior paracentral loss) from the worse eye were positively correlated with AT 14 in the better eye, while AT 1 (normal VF), AT 2 (superior peripheral defect), AT 4 (temporal wedge) and AT 11 (concentric peripheral defect) showed negative correlations. Fig. 2b shows an example of high likelihood of AT 14 in the better eye predicted by the worse eye for every reliable VF pair in a 3 VF series. While the first and last VFs of the better eye have 14.1% and 10.4% AT 14, the defect (1.0% AT 14) was missed in the second VF.

Conclusions : VF ATs from the patient’s worse eye can be used to predict and therefore confirm a VF defect in the better eye.

This is a 2020 ARVO Annual Meeting abstract.

 

Fig. 1

 

Fig. 2