Purchase this article with an account.
Kouros Nouri-Mahdavi, Vahid Mohammadzadeh, Alessandro Rabiolo, Joseph Caprioli, Siamak Yousefi; Machine Learning for Prediction of Visual Field Progression. Invest. Ophthalmol. Vis. Sci. 2019;60(9):5573.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a machine learning classifier that can identify visual field (VF) progression based on baseline clinical and demographic factors and baseline and longitudinal OCT parameters in a cohort of patients with evidence of central field damage.
104 eyes of 104 patients with ≥2 years of follow-up and 5 or more 24-2 VF exams were enrolled. We defined 24-2 VF progression at the end of follow-up with pointwise linear regression based on presence of ≥3 test locations with a significant (p<0.01) rate of change (<–1.0 dB/year). A naive Bayes machine learning classifier with 10-fold cross-validation was used to predict VF progression with the following predictors: baseline clinical and demographic factors, baseline global and sectoral peripapillary retinal nerve fiber layer (pRNFL), macular ganglion cell/inner plexiform layer (GCIPL) and full thickness (FMT) measurements and global and sectoral macular (Figure 1) and pRNFL rates of change during the follow-up. The most discriminative subset of attributes was selected based on the Classifier Subset Evaluator and a Greedy Stepwise Forward Search. Accuracy and area-under-ROC curves were used to compare subsets.
The average (SD) follow-up time and number of VFs were 4.5 (0.9) years and 8.7 (1.6), respectively. Progression on 24-2 VFs was detected in 23 eyes (22%). The following subset of variables best predicted VF worsening: type of glaucoma, rates of FMT change at 5.6°eccentricity, rates of GCIPL change in the superior and inferior hemiretina, and rates of GCIPL change in sectors C and F. The prediction accuracy of this subset of variables was 81.7% with an area-under-ROC curve of 0.75 (Figure 2).
We demonstrate that VF progression can be detected with a machine learning approach with reasonable and clinically relevant accuracy. Further refinement of this approach could provide clinicians with a valuable tool for predicting functional progression and clinical decision making in glaucoma.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only