A deep-learning system predicts glaucoma incidence and progression using retinal photographs.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning is established for glaucoma diagnosis but lacks validated algorithms for incidence and progression prediction.
• Current methods for predicting glaucoma onset and progression require further clinical validation.

Purpose of the Study:

• To develop and clinically validate a deep-learning system for predicting glaucoma onset and progression risk.
• To assess the system's performance using color fundus photographs (CFPs) in diverse patient cohorts.

Main Methods:

• Developed artificial intelligence (AI) models using CFPs from 17,497 eyes across longitudinal cohorts (3-5 year follow-up).
• Trained AI to predict glaucoma incidence and progression, validated against ophthalmologist diagnoses.
• Evaluated model performance using area under the receiver operating characteristic (AUROC) curve, sensitivity, and specificity.

Main Results:

• The glaucoma incidence prediction model achieved an AUROC of 0.90 in the validation set and 0.88-0.89 in external test sets.
• The glaucoma progression prediction model achieved an AUROC of 0.91 in the validation set and 0.87-0.88 in external test sets.
• Both models demonstrated good generalizability and outstanding predictive performance across external cohorts.

Conclusions:

• Deep learning algorithms are feasible for the early detection of glaucoma.
• The developed AI system shows significant potential for predicting glaucoma progression.
• Clinical validation in external cohorts supports the system's reliability for risk stratification.