Purchase this article with an account.
Siamak Yousefi, Tobias Elze, Louis Pasquale, Osamah Saeedi, Mengyu Wang, Lucy Shen, Sara Wellik, C Gustavo De Moraes, Jonathan S Myers, Michael V Boland; Clinical utility of the artificial intelligence enabled dashboard for glaucoma monitoring. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4526.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To validate an artificial intelligence enabled dashboard for monitoring glaucoma using automated visual field (VF) tests
We developed the initial dashboard using 13,231 cross-sectional VF tests and applied principal component analysis (PCA) and t-distributed stochastic neighbor embedding (tSNE) to extract global and local characteristics of the patterns of VF loss (Yousefi et al., Image and Vision Computing New Zealand; IVCNZ, 2018:1-6). To validate the dashboard and investigate its clinical utility, we used VF sequences from two independent datasets. The first dataset included 133 eyes of glaucoma patients (each with 10 VF tests collected every week over 10 consecutive weeks) at Moorfields Eye Hospital, London, UK, that were likely stable in such a short time. The second dataset included 154 eyes of glaucoma patients (average 9.3 tests) from the Glaucoma Research Network (GRN) with a consensus by six different algorithms that they were progressing. We projected the sequences of VFs of the eyes in these benchmark datasets and computed the specificity (using Moorfields dataset; Fig. 1) and sensitivity (using GRN dataset; Fig. 2) of the dashboard based on the trajectory direction towards glaucoma worsening and its length.
A total of 125 eyes from the Moorfields benchmark dataset with confirmed non-progression consistently represented no progression on the dashboard, equivalent to specificity of 94%. A total of 119 eyes from GRN with confirmed progression consistently represented progression on the dashboard, equivalent to sensitivity of 77%.
We validated the clinical utility of the artificial intelligence enabled dashboard for glaucoma monitoring using longitudinal VFs from independent datasets. The dashboard had an excellent specificity and reasonable sensitivity. The proposed dashboard could be useful in clinical practice and glaucoma research for monitoring glaucoma on a user-friendly screen.
This is a 2020 ARVO Annual Meeting abstract.
Figure 1. VF sequences of 133 stable eyes were tracked on dashboard. Red color indicates likely progression, blue represents possibly progression (borderline), and black indicates no progression. Green circles represent the first visit of the patient.
Figure 2. VF sequences of 154 eyes with consensus on progression based on six different algorithms were tracked and presented on the dashboard.
This PDF is available to Subscribers Only