[Skip to Navigation]
Comment & Response
February 20, 2020

Considerations When Using a Deep Learning System to Diagnose Glaucomatous Optic Neuropathy

Author Affiliations
  • 1School of Medicine, Imperial College London, London, United Kingdom
JAMA Ophthalmol. 2020;138(4):419-420. doi:10.1001/jamaophthalmol.2020.0023

To the Editor I congratulate Liu et al1 for their exciting study on the use of machine learning in identifying patients with glaucomatous optic neuropathy. While the study has succinctly presented its major findings, there are 2 areas that interested readers may have found helpful if these had been included in the study.

First, glaucomatous optic neuropathy, especially cases resulting from the common disease primary open-angle glaucoma, progresses slowly and is largely asymptomatic until later stages of the disease.2 Therefore, I wonder what proportion of the training set images were from patients who had asymptomatic glaucomatous optic neuropathy and what proportion from those who had more advanced or symptomatic disease. This might determine the performance of the trained model in different clinical settings. A model that is able to accurately identify glaucomatous optic neuropathy in patients with no symptoms might improve screening efficiency in high-risk populations and prevent glaucomatous optic neuropathy from causing irreversible damage before detection.2 Likewise, a model trained with images from advanced disease might have better performance in the acute setting. Further information about the nature of the fundus images might give the readers better appreciation of the potential application of the trained model.

Add or change institution