Real-time diabetic retinopathy screening by deep learning in a multisite national screening programme: a prospective...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Public Health

Background:

• Diabetic retinopathy is a major cause of preventable blindness globally, particularly in low- and middle-income countries (LMICs).
• Deep-learning (DL) systems offer potential to improve diabetic retinopathy screening in resource-limited settings.
• Prospective studies evaluating the real-world usability and performance of DL systems for this purpose are limited.

Purpose of the Study:

• To assess the real-world performance and feasibility of deploying a DL system for diabetic retinopathy screening in Thailand.
• To compare the diagnostic capabilities of the DL system with those of human retina specialists.

Main Methods:

• A prospective interventional cohort study was conducted at nine primary care sites in Thailand.
• 7651 eligible patients with diabetes underwent screening using a DL system, with fundus images interpreted in real-time.
• Performance metrics (accuracy, sensitivity, specificity, PPV, NPV) of the DL system were compared against an adjudicated reference standard set by fellowship-trained retina specialists.

Main Results:

• The DL system achieved 94.7% accuracy, 91.4% sensitivity, and 95.4% specificity for vision-threatening diabetic retinopathy.
• Performance was comparable to retina specialist over-readers (93.5% accuracy, 84.8% sensitivity, 95.5% specificity).
• The DL system showed a positive predictive value (PPV) of 79.2% and a negative predictive value (NPV) of 95.5%.

Conclusions:

• A DL system can provide real-time diabetic retinopathy detection capabilities comparable to retina specialists in community-based screening settings.
• Successful implementation of DL systems in large-scale screening programs in LMICs necessitates consideration of socioenvironmental factors and existing workflows.