Algorithm detected glaucoma in older adults’ eye photos
The algorithm detected 70.6% of affected eyes while correctly classifying 95% of eyes without glaucoma in an external population-based validation.
Independent replication
Limit: Incomplete optic nerve head visibility was an important source of false positives
Edited by Boris Djordjevic · AgingResearch.News
An observational validation study tested a deep-learning algorithm on eye photographs from 9,837 Rotterdam Study participants aged 50–80 years. Among 19,088 eyes, 201 had perimetric glaucoma, meaning glaucoma with visual-field loss. The algorithm had been trained on a separate set of photographs.
It detected 70.6% of affected eyes at a setting that correctly classified 95% of eyes without glaucoma. Overall, there were 5.1 referrals per detected glaucoma case. Correctly detected eyes had more advanced visual-field damage than missed cases. Among correctly identified cases with earlier photographs available, 33 of 61 (54%) had been flagged before diagnosis, typically 4–7 years earlier. Incomplete views of the optic nerve’s exit point from the eye were an important source of false alarms.
Why it matters
For glaucoma screening in older populations, the authors see photograph-based deep learning as a scalable way to identify people needing further assessment.
Caveats
This observational validation assessed detection accuracy, not whether algorithm-guided screening improved vision. The participants were aged 50–80, so these results do not establish performance outside that age range.
Comments
Be the first to comment.