Glaucomatous patterns of binocular visual field loss identified by unsupervised machine learning

Abstract

 

Purpose

 

To identify patterns of glaucomatous defects in constructed binocular standard automated perimetry (SAP) visual fields (VFs) using a previously described unsupervised classifier, the variational Bayesian independent component analysis mixture model (VIM).

 

Methods

 

Monocular VFs from 1,086 eyes of 543 glaucoma patients (with repeatable abnormal SAP results by PSD or GHT in at least one eye, VF MD = -6.15 ± 6.70 dB in worst eye) and 1,120 eyes of 560 healthy participants (with SAP results within normal limits OU, VF MD=-0.54 ± 1.17 dB in worst eye) were included from participants in the Diagnostic Innovations in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES). Binocular VFs were estimated using the summation method for sensitivity between eyes at each location. For all participants, VIM input was 56 threshold test points from the resulting constructed binocular 24-2 test pattern, plus age. The model with the best average combination of sensitivity and specificity of VFs was chosen from 720 generated models.

 

Results

 

The optimal VIM model separated all of the binocular VFs into 5 clusters. Cluster 1 contained primarily VFs from healthy participants (526/560, specificity 94%). The remaining clusters (clusters 2, 3, 4 and 5, combined) contained primarily VFs from participants with glaucoma (436/543, sensitivity 80%) in increasing order of defect severity. The best 5-cluster model was composed of a total of 13 axes, 2 axes each for the first four clusters and 5 axes for the last cluster.

 

Conclusions

 

Without a priori knowledge of class membership, VIM successfully separated binocular visual fields from glaucoma patients and control patients and identified glaucomatous patterns of loss. Because binocular VF defects may be more strongly associated with general task performance than monocular defects, they likely are more representative of overall visual sensitivity and visual disability.

 

 

Defect patterns representing VFs +2 SD from the healthy cluster centroid along the first axis of each glaucoma cluster, showing an obvious increase in severity among clusters from top left to bottom right.

 

Keywords: 642 perimetry • 758 visual fields • 473 computational modeling