Background: The glycemic threshold for the diagnosis of type 2 diabetes (T2D) was determined based upon the prevalence of moderate retinopathy; however, diabetes related complications may occur below that threshold. This study utilized three companies to analyse retinal images using their commercial software and returned data on absence or presence of features of any diabetic retinopathy (DR) on a per-eye basis.

Methods: Ophthalmic retinal images from 11,449 subjects (right and left eyes for each, totalling 22,898 images) not known to have diabetes from the Qatar Biobank were subjected to assessment for detection of features of DR using three Conformitè Europëenne (CE)-marked independent commercial Artificial Intelligence (AI) software systems.

Results: From 11,449 patients: •DR was recorded as absent in both eyes in 7,094 (62.0%) patients by all 3 methods (Test Negative at patient level).•DR was detected by 2 or 3 of the AI systems in at least one eye in a total of 2,408 (21.0%) patients (Test Positive at patient level). Of these, in 1,532 (13.4%) cases DR was detected in just one eye of any given patient and in 876 (7.6%) cases DR was detected in both eyes of a patient. •Overall, 1,947 (17.0%) patients had at least one eye reported as unassessable by 2 or 3 of the AI systems; hence, No Overall Result for DR features could be obtained for these patients. Results from the three AI companies differed markedly: 17.7%, 32.9% and 46.6% ranked Test-Positive results at patient level. Only 304 right eyes and 280 left eyes were positive for all 3 companies.

Conclusion: This is the first direct comparison of three independent AI companies when simultaneously assessing retinal images from more than 10,000 individuals. It showed markedly differing prevalences of diabetic retinopathy between existing AI systems, indicating inconsistencies if wishing to apply these tools to detect retinopathy in subjects not already known to have diabetes, and may differ in patients with known diabetes.


A.E.Butler: None. E.S.Kilpatrick: None. S.C.Hunt: None. S.Atkin: None.

Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at