AI Can Spot Early Diabetic Retinopathy in Asymptomatic Patients

An autonomous artificial intelligence (AI) system sufficiently matched human experts in diagnosing nonproliferative referable diabetic retinopathy (DR) in early stages for asymptomatic patients, according to a study published in Eye.

The prospective observational pilot study included ultra-widefield (UWF) color fundus images from patients with asymptomatic diabetes mellitus (types 1 and 2) without a previous diagnosis of DR. 

For each eye, 2 color fundus images (45°) were analyzed by a server-based AI algorithm (IDx-DR, Digital Diagnostics). UWF color fundus imaging was performed by 2 retina specialists who also assessed the International Clinical DR severity score of the images based on both 7-field area projection (7F-mask) grading (according to the early treatment DR study) and on the total gradable area (UWF full-field).

Of 54 patients (33 men and 21 women; 107 eyes) included in the study, 32 patients had type 2 diabetes (11 women) with a mean HbA1c level of 7.5±1.9%, and 20 patients had type 1 diabetes (8 women) with a mean HbA1c level of 7.9±1.6%. Diabetes type was unknown for 2 patients. Patients had a mean age of 55 years (range, 19–80) and a mean best-corrected visual acuity (Snellen charts) of 0.99±0.25. 

The autonomous AI-based system diagnosed 16 patients as DR negative, 28 patients with moderate DR, and 10 patients as having vision-threatening disease (severe or proliferative diabetic retinopathy, diabetic macular edema). Using the 7F-mask grading, the human retina specialists diagnosed 23 patients as DR negative, 11 patients with mild DR, 19 patients with moderate DR, and 1 patient with severe DR. When analyzing UWF full-field, they diagnosed 20 patients as DR negative, 12 patients with mild DR, 21 patients with moderate DR, and 1 patient with severe DR.

The investigators found that in 66.6% of cases, the autonomous AI-based system and 7F-mask gradings matched (95% CI, 0.21-0.58), and in 66.7% of cases, the autonomous AI-based system and UWF full-field gradings matched (95% CI, 0.18-0.58). They also demonstrated that the autonomous AI-based system sensitivity/specificity was 100% (95% CI, 83–100)/47% (95% CI, 30-65) against the 7F-mask grading and was 95% (95% CI, 77–100)/47% (95% CI, 29–65) against UWF full-field grading. 

The investigators suggest that the autonomous AI-based system may be suitable for DR screening and diagnosis in diabetes primary care settings or telemedicine programs.

Limitations of the study included grading using UWF color fundus images (not stereoscopic images), the relatively small sample size, and recruitment of patients from a single tertiary referral center.

Disclosure: Some study authors declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.  

Reference

Sedova A, Hajdu D, Datlinger F, et al. Comparison of early diabetic retinopathy staging in asymptomatic patients between autonomous AI-based screening and human-graded ultra-widefield colour fundus images. Eye (Lond).  Published online February 7, 2022. doi:10.1038/s41433-021-01912-4

This article originally appeared on Ophthalmology Advisor