ThailandHealthcare and life sciences2018 – ongoing

Thailand Ministry of Public Health (national diabetic retinopathy screening programme)

Deep-learning eye screening at primary care clinics, built with Google Health

An AI eye-screening tool matched retina specialists in 7,651 Thai patients, but nurses in rural clinics struggled with image quality and slow connections

In progressIndependent study

Problem

Diabetic retinopathy can cause blindness, and patients in Thailand often waited four to five weeks for an ophthalmologist's result. Fewer than half of diabetic patients were screened each year.

Approach

Google Health's deep-learning system graded retinal photos taken by nurses at primary care clinics in Thailand's national screening programme, with retina specialists over-reading each image as a safety check. A prospective study ran at nine sites from December 2018 to March 2020. A separate field study observed nurses at eleven clinics in Pathum Thani and Chiang Mai.

Results

91.4% vs 84.8%sensitivity for vision-threatening disease, deep-learning system vs retina specialist over-readers, in the prospective studyIndependent study[1]

7,651patients eligible for analysis at nine primary care sites (31.5% referred)Independent study[1]

200 to 100patients screened in a day at one clinic after a two-hour internet outage, per TechCrunch's reading of the field studyIndependent study[3]

Takeaway

Accuracy held up in the clinics, but the field study found that a lack of sufficiently dark rooms, blurry images and weak internet limited use. The authors said workflows must be fitted to each setting before scaling.

Sources

Real-time diabetic retinopathy screening by deep learning in a multisite national screening programme: a prospective interventional cohort study
The Lancet Digital Health (via Google Research)2022Research paper or report
A Human-Centered Evaluation of a Deep Learning System Deployed in Clinics for the Detection of Diabetic Retinopathy
ACM CHI Conference on Human Factors in Computing Systems (via Google Research)2020Research paper or report
Google medical researchers humbled when AI screening tool falls short in real-life testing
TechCrunch27 Apr 2020News article

Source summaries

[1]The Lancet Digital Health (via Google Research)2022

Researchers conducted a prospective study at nine Thai primary care sites from December 2018 to March 2020. Of 7,940 patients screened, 7,651 were eligible. For vision-threatening disease the system had 91.4% sensitivity and 95.4% specificity, against 84.8% and 95.5% for retina specialists. The authors say it matched specialists and that workflows matter when scaling.

[2]ACM CHI Conference on Human Factors in Computing Systems (via Google Research)2020

Researchers observed and interviewed nurses at eleven Thai clinics. Resource-limited clinics struggled to produce images that met the model's quality thresholds, and screening workflows and patient experience shaped how the tool performed.

[3]TechCrunch27 Apr 2020

TechCrunch summarised the field study. Only two of the clinics had a room dark enough for good images, uploads took 60 to 90 seconds on slow connections, and a two-hour outage halved one clinic's screening from 200 to 100 patients. The system rejected blurry or dark images, which frustrated nurses.

Full details are in the original sources.

Added 25 Oct 2026 · Updated 25 Oct 2026