Evaluation of non-ophthalmologist-led and offline AI-assisted models for diabetic retinopathy screening in India: a pragmatic diagnostic accuracy study.

The non-ophthalmologist-led model demonstrated high diagnostic accuracy for detecting diabetic retinopathy at health and wellness centres.

• Sensitivity for DR detection was 86.4% (95% CI 65.1% to 97.1%)
• Specificity for DR detection was 94.3% (95% CI 88.5% to 97.7%)
• The ungradability rate for this model was 8%
• Images were captured using non-mydriatic fundus cameras and graded by two masked human graders with disagreements resolved by a senior retina specialist

The non-ophthalmologist-led model achieved very high sensitivity and specificity for referable diabetic retinopathy (RDR) detection.

• Sensitivity for RDR reached 95.8% (95% CI 78.9% to 99.9%)
• Specificity for RDR was 93.1% (95% CI 88.0% to 96.5%)
• This model was conducted at health and wellness centres (HWCs) in rural Block Boothgarh, Mohali District, Punjab, India
• 200 people with diabetes aged ≥30 years were enrolled in this screening arm out of 600 total participants

The offline AI-assisted smartphone-based community screening model achieved moderate-to-good accuracy for RDR but was limited by a high ungradability rate.

• Sensitivity for RDR was 93.3% (95% CI 68.1% to 99.8%)
• Specificity for RDR was 85.1% (95% CI 76.9% to 91.2%)
• The ungradability rate was 38%, substantially higher than the non-ophthalmologist-led model's 8%
• High ungradability was mainly attributed to cataracts and poor image quality
• The AI provided only binary classification outputs, which was identified as a current limitation

The non-ophthalmologist-led model demonstrated greater accuracy and operational feasibility compared to the offline AI-assisted model.

• Both approaches effectively identified referable cases of diabetic retinopathy
• The non-ophthalmologist-led model had a substantially lower ungradability rate (8% vs 38%)
• The study was conducted in primary healthcare settings including HWCs and community-based screening sites in rural India
• The study enrolled a total of 600 people with diabetes aged ≥30 years across three screening models including a standard referral-based care arm

Offline AI-enabled screening demonstrates potential for community use but is currently limited by image quality issues and binary classification outputs.

• The AI system operated offline using smartphone-based fundus imaging in community settings
• Poor image quality and cataracts were the primary drivers of the 38% ungradability rate
• Binary classification outputs were identified as a limitation of the current AI model
• The authors suggest integrating both approaches may strengthen diabetic retinopathy screening coverage in resource-limited settings

The study was designed as a pragmatic diagnostic accuracy study using specialist grading as the reference standard across three distinct screening models.

• The three models were: (1) non-ophthalmologist-led DRS at HWCs, (2) AI-assisted smartphone-based DRS in the community, and (3) standard referral-based care
• 600 people with diabetes aged ≥30 years were enrolled in total
• The setting was rural Block Boothgarh, Mohali District, Punjab, India
• Retinal images were independently graded by two masked human graders; a senior retina specialist resolved disagreements
• The study was registered under CTRI/2022/10/046283