Predicting Glaucoma before Onset Using Deep Learning

Abstract

Purpose: To assess the accuracy of deep learning models to predict glaucoma development from fundus photographs several years before disease onset.

Design: Algorithm development for predicting glaucoma using data from a prospective longitudinal study.

Participants: A total of 66 721 fundus photographs from 3272 eyes of 1636 subjects who participated in the Ocular Hypertension Treatment Study (OHTS) were included.

Main outcome measures: Accuracy and area under the curve (AUC).

Methods: Fundus photographs and visual fields were carefully examined by 2 independent readers from the optic disc and visual field reading centers of the OHTS. When an abnormality was detected by the readers, the subject was recalled for retesting to confirm the abnormality and for further confirmation by an end point committee. By using 66 721 fundus photographs, deep learning models were trained and validated using 85% of the fundus photographs and further retested (validated) on the remaining (held-out) 15% of the fundus photographs.

Results: The AUC of the deep learning model in predicting glaucoma development 4 to 7 years before disease onset was 0.77 (95% confidence interval [CI], 0.75-0.79). The accuracy of the model in predicting glaucoma development approximately 1 to 3 years before disease onset was 0.88 (95% CI, 0.86-0.91). The accuracy of the model in detecting glaucoma after onset was 0.95 (95% CI, 0.94-0.96).

Conclusions: Deep learning models can predict glaucoma development before disease onset with reasonable accuracy. Eyes with visual field abnormality but not glaucomatous optic neuropathy had a higher tendency to be missed by deep learning algorithms.

Figure 1.

Flowchart of glaucoma identification and labeling. The eye is labeled as glaucoma based on either glaucomatous optic neuropathy (GON) or visual field abnormality. Reading center assignment should be further confirmation by an endpoint committee. Three datasets were selected from fundus photographs based on glaucoma onset date of each eye; one dataset for glaucoma diagnosis and two datasets for glaucoma prediction.

Figure 2.

Eyes without any sings of glaucomatous optic neuropathy (GON) or visual field abnormality were followed for about 10 years and fundus photographs were collected annually. The onset time represents when a sample eyes was identified as glaucoma based on GON or visual field abnormality. Green arrow corresponds to fundus photographs collected 4–7 years prior to the date of glaucoma onset, yellow arrow represents fundus photographs that were collected 1–3 years prior to the date of glaucoma conversion, and red arrow corresponds to fundus photographs collected on or after the time that glaucoma onset was identified.

Figure 3.

Activation maps representing regions that are most promising for the deep learning model to make a diagnosis. First row: fundus photographs from non-glaucoma eyes. Second row: activation maps of fundus photographs of non-glaucoma eyes (shown in the first row). Third row: fundus photographs from eyes with glaucoma. Fourth row: activation maps of fundus photographs of eyes with glaucoma (shown in the third row).

Figure 4.

Receiver operating characteristic (ROC) curves of the prediction and diagnosis models. Green curved belongs to the model that predicts glaucoma 4–7 years prior to the onset of the disease, red curve predicts glaucoma 1–3 years prior to the onset of the disease, and blue curve shows diagnosing glaucoma on or after onset.