Loading...
By Joe Elia
Edited by David G. Fairchild, MD, MPH, and Lorenzo Di Francesco, MD, FACP, FHM
A computer algorithm, based on "deep machine learning," evaluated retinal fundus photographs for diabetic retinopathy with an accuracy exceeding guideline recommendations, according to a JAMA study conducted by Google.
The algorithm was developed with a set of over 125,000 retinal images that were each evaluated by several ophthalmologists. The algorithm was "tuned" to deliver results similar to the ophthalmologists'. The algorithm then evaluated a separate set of over 10,000 images, and its accuracy was compared against the results of ophthalmologists' interpretations.
The algorithm detected referable diabetic retinopathy with high sensitivity and specificity, ranging from 87% upwards. (Editorialists point out that guidelines generally call for percentages above 80%.)
Comment
LINK(S):
JAMA article (Free)
JAMA editorial #1 (Free)
JAMA editorial #2 (Free)
JAMA Viewpoint article (Free)
Background: Physician's First Watch coverage of use of telemedicine in detecting diabetic retinopathy (Free)