Purchase this article with an account.
Stuart Keith Gardiner, Shaban Demirel; Comparison between global, cluster, and pointwise trend analyses for detecting visual field change. Invest. Ophthalmol. Vis. Sci. 2017;58(8):2863.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
It is vital to be able to measure the rate at which the visual field deteriorates in eyes with glaucoma, but the best rate metric is unclear. Global measures such as Mean Deviation (MD) may miss localized defects; yet pointwise analyses are more variable. This study assesses an intermediate possibility, using cluster trend analysis.
Trend-based criteria for change were defined using global analysis (MD worsening with p<CritMD), cluster analysis (mean total deviation within ten clusters used by the Octopus perimeter; ≥n clusters worsening with p<Critnc), and pointwise analysis (total deviation at ≥n locations worsening with p<Critnp), for different n. Critical p-values were chosen so that each criterion had specificity exactly 95% in a test-retest dataset: 45 eyes of 23 participants, tested 5 times within a few weeks, using all 120 possible re-orderings of tests per eye. The criteria were applied to a separate longitudinal dataset: 506 eyes of 256 participants tested every ~6 months, mean 14 visits, range 5-24. The time to detect change using each criterion was found, and compared between criteria using Cox proportional hazards models.
For cluster analysis, the quickest criterion to detect change in ≥25% of eyes was “5 clusters worsening with p<0.259”. For pointwise analysis, the quickest was “9 locations worsening with p<0.139”, and this detected change significantly sooner than cluster or global criteria (both p<0.001; see figures). However, once 2 more fields after initial change detection were included, fewer than half of these eyes still met the criterion. Proportions confirmed were higher for cluster and global criteria (both p<0.001). Indeed if changes were only accepted if they were confirmed after 2 more fields, the cluster criterion detected change slightly, albeit non-significantly, sooner than global or pointwise criteria (p=0.16 and p=0.58 respectively). The same was true if changes had to be confirmed using the next 4 fields.
For equal specificity, pointwise analysis detects change more rapidly than global or cluster trend analyses. However, a high proportion of these ‘changing’ series are likely false positives. By contrast, global trend analysis using MD detected change later, but it was more often subsequently confirmed. Cluster trend analysis appears a good compromise, detecting change sooner than global indices without too many false positives.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only