Early detection of visual field conversion using an unsupervised machine-learning analysis of retinal nerve fiber layer thickness measurements

Abstract

Purpose : To determine if glaucomatous visual field conversion can be detected by machine-learning analysis of retinal nerve fiber layer thickness (RNFLT) measurements prior to glaucomatous visual field conversion

Methods : Study participants were recruited from the Diagnostic Innovations in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES). Longitudinal RNFLT measurements (Spectralis RNFL circle scan protocol, 768 A-scans; 64 RNFL sectors were generated by averaging 12 contiguous A-scans to reduce noise and any possible misalignment) from 74 eyes of 68 patients followed clinically (minimum 3 visits before conversion date) for suspicion of glaucoma were analyzed using Gaussian mixture model using expectation maximization (GEM) progression of patterns (POP) prior to conversion to glaucoma (conversion was defined as ≥ 3 consecutive abnormal SAP results by GHT or PSD ≤ 0.05). Previous cross-sectional GEM analyses of RNFLT measurements identified 7 glaucomatous RNFL patterns (axes) of loss (Yousefi et al., ARVO 2015). To define the conversion limit for GEM, longitudinal RNFLT measurements of 83 stable glaucoma eyes (imaged once a week for approximately 5 weeks, approximately 120 permuted series per eye) were analyzed across each axis and the rate of change was approximated using linear regression (LR). The conversion limit for each axis was adjusted to result in an overall 95th percentile conversion limit. To detect conversion, the longitudinal RNFLT measurements of patient eyes were analyzed across each GEM axis and the rate of change was approximated by LR. Conversion was assigned if the rate of change in an eye, along any GEM axis, was greater than the conversion limit for that axis; otherwise, no conversion was assumed. The early detection rate (defined as percentage of convert eyes correctly identified) of GEM was compared to that of LR of RNFLT global average, using the same method for setting conversion limits.

Results : GEM detected 88% of converted eyes prior to the first of 3 consecutive abnormal SAP results and linear regression of average RNFLT detected 62%. Fifty-seven percent of converted eyes were detected by both methods.

Conclusions : Analysis of patterns of RNFLT using GEM can detect glaucoma in more eyes than linear regression of global average RNFLT before the onset of repeatable SAP abnormality.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.