Purchase this article with an account.
Siamak Yousefi, Michael Henry Goldbaum, Ehsan Varnousfaderani Shahrian, Linda M Zangwill, Robert N Weinreb, Felipe A Medeiros, Christopher A Girkin, Jeffrey M Liebmann, Christopher Bowd; Unsupervised machine learning to recognize glaucoma defect patterns and detect progression in RNFL thickness measurements. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):4564.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To recognize glaucomatous defect patterns and to detect glaucomatous progression in a longitudinal series of RNFL thickness (RNFLT) measurements, using the previously described unsupervised Gaussian mixture model using expectation maximization (GEM, Yousefi et al., 2014, IEEE TBME)
RNFLT measurements from 2274 eyes of 1227 participants in the Diagnostic Innovations in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES) were obtained using the Spectralis RNFL Circle Scan protocol (768 A-scans). For each eye, 64 RNFL sectors were generated by averaging groups of 12 A-scans. 853 eyes had ≥ 3 contiguous sectors with RNFLT outside of normal limits (i.e., abnormal eyes, = 1% cut-off based on RNFL data from 485 DIGS and ADAGES eyes recruited and screened as normal) and 1421 eyes had ≤ 2 contiguous sectors outside of normal limits (designated normal eyes). GEM assigned cross-sectional abnormal and normal RNFL measurements to clusters, decomposed each cluster into independent axes and recognized glaucomatous patterns of RNFL loss within each axis. Simulated deviation plots, overlaid on en face images, were used to display the observed loss patterns (Figure 1). Variability within each axis (95% stability confidence limit) was determined in 97 simulated stable glaucoma eyes (patients tested 5 times over 5 weeks). To test for progression detection, 84 progressed eyes (by SAP GPA) were employed and the sequence of RNFLT measurements for each eye was projected onto each axis. Progression was designated in an eye if the progression rate along any axis was greater than the stability confidence limit of the stable eyes. Otherwise, non-progression was designated.
Sensitivity was 70% for placing eyes with abnormal RNFLT in an abnormal cluster and specificity was 82% for placing eyes with normal RNFLT in the normal cluster. Sensitivity for progression detection was 78% at 95% specificity. Sensitivity for linear regression (using the same method to define progression) of average RNFLT measurements was 52%.
GEM separated RNFL thickness from abnormal and normal eyes with reasonable accuracy. Distinctly different patterns of glaucomatous RNFL defects were identified, and changes in these patterns over time detected known glaucomatous progression with reasonable sensitivity.
This PDF is available to Subscribers Only