Google DR screening in Thailand matches specialists at 94.7%

AI system

System: Google Health deep-learning diabetic retinopathy screening system
Model: Not assessed
Release: Exact release not established.
Application developer: Google (Google Health)
Role: The model graded retinal images for signs of diabetic retinopathy and recommended referral.
Configuration: The deployment used the Automated Retinal Disease Assessment system, ARDA.
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: Fundus photography at primary-care sites. Real-time deep-learning grading with referral recommendations. Retina-specialist over-reading as a safety mechanism.
Human oversight: Not assessed
Relation to AI: The deep-learning grades and referral recommendations were delivered in real time with specialist over-reading as a safety net. Accuracy figures come from the vendor-funded, vendor co-authored prospective study. An independent news account of Google's field study records substantial image rejection and nurse frustration.
AI contribution: The deep-learning grades and referral recommendations were delivered in real time with specialist over-reading as a safety net. Accuracy figures come from the vendor-funded, vendor co-authored prospective study. An independent news account of Google's field study records substantial image rejection and nurse frustration.

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: 94.7% accuracy for vision-threatening DR, 2,412 referrals
Comparison: regional retina specialist over-readers, measured against an adjudicated reference standard from fellowship-trained retina specialists
Results: Reference: an accuracy of 93·5 (91·7-95·0; p=0·17), a sensitivity of 84·8% (79·4-90·0; p=0·024), and specificity of 95·5% (94·1-96·7; p=0·98). Result: an accuracy of 94·7% (95% CI 93·0-96·2), sensitivity of 91·4% (87·1-95·0), and specificity of 95·4% (94·1-96·7).
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Participants: Developer: Google (Google Health). Operator: Thailand national diabetic retinopathy screening program (Ministry of Public Health guidelines) with Rajavithi Hospital, Bangkok.
Jurisdictions: Benefit: Thailand (nine primary-care sites). Developer: United States. Operator: Thailand.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: Image quality sufficient for grading (a fifth of images were rejected in the 2020 field study). Reliable connectivity. Referral pathways able to absorb referrals.
Replication attempts: 2 follow-up entries. Outcome unresolved: 1. Benefit recurred: 1.
Later findings: 23 September 2026. A Nature Medicine comment of 23 September 2026 by the Google and partner-site authors describes scaling the same deep-learning screening tool across India, Thailand and Australia; an October 2026 AI Weekly alert summarizing it says the model, ARDA, has now screened more than one million patients, with Perceptra licensed to run screening in Thailand.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: People. Organizations. Society.
Depth: People: 2. Organizations: 1. Society: Not assessed.
Reasons: People: 7,651 patients received real-time grading with specialist-level accuracy and 2,412 were referred. Organizations: Real-time results in primary care; field reporting notes speed gains when the system worked and rejections and connectivity problems when it did not. Society: No population-level vision outcome reported.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently reported

Kind: Not assessed
Related entries: Not assessed
Field: Health; ophthalmology. People with diabetes due for retinopathy screening.
Purpose: Real-time automated grading of fundus photographs for diabetic retinopathy in a national screening program
Dates: Date basis unassessed: 2022 (precision unassessed). Updated: 21 October 2026.
Stage: Not assessed
Version: SHA-256: 0adfb8866006…