Progression of Patterns Using Non-negative Matrix Factorization and Iso-severity Clustering of Glaucoma Patterns Improves Detection of Glaucoma Progression

Abstract

 

Purpose:

 

We evaluated progression of patterns (POP) based on patterns as axes derived by non-negative matrix factorization (NMF) and iso-severity clustering of glaucoma patterns (ICG) for identification and quantification of progression in glaucomatous visual fields.

 

Methods:

 

The previous version of POP was developed on axes derived by the variational Bayesian independent component mixture model (VIM) environment. The environments created by NMF and ICG were designed to orient their axes to the directions along which progression actually takes place. NMF and ICG separated total deviation plots of SAP-SITA visual fields (VFs) from a set of 2085 normal and glaucomatous eyes into 4 (NMF) or 15 (ICG) axes of different patterns of VF defects. A second, independent set of 55 stable eyes with 5 VFs collected within 4 weeks (by D.R. Anderson and F. Goñi) was used to simulate stability, and a third set of 4,186 VFs from 628 eyes from the longitudinal Diagnostic Innovations in Glaucoma Study and the African Descent and Glaucoma Evaluation Study were tested for progression. The sequence of visits had a mean+sd of 6.7+1.7 VFs followed for 4.0+1.4 years. For post-hoc analysis, eyes were labeled as glaucomatous by visual field diagnostic criteria and as progressing by the assessment of progressive glaucomatous optic neuropathy (PGON) in sequential stereoscopic disk photographs. POP used linear regression (LR) to quantitate change along a sequence of fields projected onto each axis generated by VIM, NMF, and ICG. The probability of POP-NMF and POP-ICG progression in the third set of eyes was compared to the Humphrey Visual Field Index (VFI) and POP-VIM scores. Percentage progressing was compared at 95% specificity.

 

Results:

 

Percentage of progressing glaucomatous fields was shown in glaucomatous eyes and in eyes showing progression of glaucomatous optic neuropathy (Table).

 

Conclusions:

 

The data-driven automatic algorithm in POP-NMF and POP-ICG identified more progressing eyes in the PGON group than VFI or POP-VIM. POP-NMF and POP-ICG and their display methods show promise for improving our ability to detect visual field progression.  

 

Clinical Trial:

 

http://www.clinicaltrials.gov NCT00221897

 

Keywords: perimetry • computational modeling • detection