This research introduces a groundbreaking deep-learning system designed to predict and stratify the risk of glaucoma onset and progression using color fundus photographs (CFPs). This AI model stands out for its exceptional diagnostic abilities, showcasing high accuracy in identifying the incidence and progression of glaucoma. A notable aspect of the study is the exploration of using smartphones to capture CFPs, which could significantly enhance disease screening accessibility.
The AI model’s robust generalizability was confirmed through validation with external population cohorts, underlining its potential in diverse clinical settings. This advancement marks a significant contribution to developing a practical deep-learning tool for diagnosing and predicting glaucoma involving a substantial dataset of 17,497 eyes.
However, the study acknowledges certain limitations in current deep-learning models for glaucoma prediction, such as the absence of progression prediction, limited data from community populations, and insufficient external validation. The researchers suggest that incorporating various data modalities could enhance future predictive performance. Moreover, the study’s focus on high-quality CFPs limits applicability to cases with media opacities.
Addressing the challenge of a few glaucoma cases, the researchers utilized a deep-learning model with fewer parameters, ensuring effective training and validation. Despite variations in sensitivity and specificity across AI models, they consistently demonstrated high Area Under the Receiver Operating Characteristic (AUROC) values, indicating predictive solid power.
The study points out that increasing the volume of training data can further refine predictive performance. It also underscores the need for further validation in populations beyond China to solidify the model’s global applicability.
For those interested in replicating or extending this research, the study makes its data and code available at https://www.r-project.org, encouraging further exploration and development in this vital Area of medical AI.
F. Li et al., “A deep-learning system predicts glaucoma incidence and progression using retinal photographs,” Journal of Clinical Investigation, vol. 132, no. 11, p. e157968, Jun. 2022, doi: .
