Purchase this article with an account.
Siamak Yousefi, Michael Henry Goldbaum, Felipe A Medeiros, Linda M Zangwill, Robert N Weinreb, Christopher A Girkin, Jeffrey M Liebmann, Christopher Bowd; Recognizing glaucomatous defect patterns and detecting progression from visual field measurements using Gaussian mixture model and expectation maximization. Invest. Ophthalmol. Vis. Sci. 2014;55(13):985.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To recognize glaucomatous defect patterns and to detect glaucoma progression from longitudinal series of Standard Automated Perimetry (SAP) visual fields (VFs).
We obtained SAP thresholds (24-2 SITA) at 52 test points from 2085 eyes of 1398 participants in the Diagnostic Innovations in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES). 939 eyes had abnormal SAP results (PSD = 5%, GHT outside normal limits) and 1146 eyes had normal SAP results. First, we employed an unsupervised Gaussian Mixture Model using the Expectation Maximization (GEM) method to assign cross-sectional abnormal and normal VFs to clusters. We trained 600 models and chose the model with the best average combination of sensitivity and specificity. We then used principal component analysis to decompose each cluster into several axes. Next, in an independent dataset of 97 stable glaucoma eyes (from patients tested 5 times over 5 weeks), we computed the variability within each axis to determine the 95% confidence limits used to define progression. To test glaucoma progression detection, we employed a dataset of 76 progressing eyes (defined by stereophotograph assessment), projected the sequence of fields for each eye onto each axis, and assigned progression if the progression rate along any axis was greater than the 95% confidence limit (corrected for number of axes) of the stable eyes. Otherwise, non-progression was assigned.
GEM clustering was 87% sensitive and 96% specific for correctly clustering VFs. Progression detection accuracy at 95% specificity using GEM axes was 29% sensitive. Sensitivities for linear regression (using the same criterion to define progression) of MD, PSD and VFI were 17%, 14% and 14%, respectively.
A progression detection framework was developed using GEM, that could identify glaucomatous visual field defect patterns and could detect glaucomatous progression from baseline and a sequence of follow-up SAP measurements, with higher sensitivity than regression of global SAP indices.
This PDF is available to Subscribers Only