Two Machine Learning Models to Economize Glaucoma Screening Programs: An Approach Based on Neural Nets.

Abstract

Background: In glaucoma screening programs, a large proportion of patients remain free of open-angle glaucoma (OAG) or have no need of intraocular eye pressure (IOP)-lowering therapy within 10 years of follow-up. Is it possible to identify a large proportion of patients already at the initial examination and, thus, to safely exclude them already at this point? Methods: A total of 6889 subjects received a complete ophthalmological examination, including objective optic nerve head and quantitative disc measurements at the initial examination, and after an average follow-up period of 11.1 years, complete data were available of 585 individuals. Two neural network models were trained and extensively tested. To allow the models to refuse to make a prediction in doubtful cases, a reject option was included. Results: A prediction for the first endpoint, 'remaining OAG-free and no IOP-lowering therapy within 10 years', was rejected in 57% of cases, and in the remaining cases (43%), 253/253 (=100%) received a correct prediction. The second prediction model for the second endpoint 'remaining OAG-free within 10 years' refused to make a prediction for 46.4% of all subjects. In the remaining cases (53.6%), 271/271 (=100%) were correctly predicted. Conclusions: Most importantly, no eye was predicted false-negatively or false-positively. Overall, 43% all eyes can safely be excluded from a glaucoma screening program for up to 10 years to be certain that the eye remains OAG-free and will not need IOP-lowering therapy. The corresponding model significantly reduces the screening performed by and work load of ophthalmologists. In the future, better predictors and models may increase the number of patients with a safe prediction, further economizing time and healthcare budgets in glaucoma screening.