Author: Claude (AI), from PubMed abstracts · Reviewed by: Dr Mina Abdelmseih · Reviewed: 2026-11-15 · Review due: 2027-11-15

Educational content only. Not personalised medical advice. Clinical decisions remain the responsibility of the treating clinician.

Diabetic retinopathy screening programmes

Diabetic retinopathy (DR) screening aims to detect retinopathy before it threatens sight so it can be treated in time [S2] [S4]. Retinal photography, graded by trained staff or artificial intelligence (AI), is accurate enough for programmes in low-resource settings [S3] [S10]. The weakest link is usually attendance: adherence was 39.3% in low- and middle-income countries [S7].

Key points - Non-mydriatic two-field photography, dilating only those with ungradable images, is a pragmatic way to start screening in low-income settings [S3] [S6]. - AI grading in developing countries had pooled sensitivity 0.933 and specificity 0.903 for referable DR [S10]. - Quality improvement interventions increased screening attendance by 12% over usual care (low-certainty evidence), but none of the 66 RCTs was from a low- or middle-income country [S1]. - Health education and knowledge about DR were linked to higher uptake; lack of skilled staff and imaging infrastructure were the main provider barriers [S4]. - Low-risk patients may be screened less often: intervals of 24 to 48 months for no or mild non-proliferative DR are supported by one national guideline [S2].

Evidence last checked: 2026-10-17 (PubMed abstracts).

Red flags: signs that need urgent action

• In a person: proliferative DR or diabetic macular oedema (DMO) on a screening image; globally, 6% of people with diabetes had proliferative DR and 11% sight-threatening DR [S8].
• In a person: ungradable images; dilate and re-image rather than report as normal [S3].
• In a programme: low attendance, especially in Africa (36.1%) [S7].
• In a programme: groups left behind; younger age, lower education, rural residence, lower income and no health insurance were linked to lower adherence [S7].
• In a programme: referrals with nowhere to go; lack of skilled human resources and imaging infrastructure are the main provider barriers [S4].

Recognise it: who to screen

• All people with diabetes. Pooled prevalence from retinal photography studies: any DR 23%, proliferative DR 6%, DMO 5%, sight-threatening DR 11% [S8].
• Children and adolescents with type 1 diabetes. Mean prevalence 9.5%, higher in low- and lower-middle-income countries; most cases occurred after five years of diabetes and after 11 years of age, and HbA1c ≥9% was linked to higher prevalence [S12]. Sight-threatening retinopathy was rare and largely confined to older adolescents with longer duration [S12].
• Those least likely to attend (see red flags) need targeted effort [S7].

Confirm it: the screening test

• Grading scale: the International Clinical Retinopathy Disease Severity Scale is recommended by the Danish guideline [S2].
• Fields: sensitivity and specificity rise with the number of photographic fields [S6]. For referable DR, sensitivity ranged from 76% with one field to 93% with wide-angle photography [S6].
• Pupil dilation: after excluding ungradable images there was no difference in sensitivity between non-mydriatic and mydriatic methods (86%) [S3]. In children, mydriatic screening detected more retinopathy than non-mydriatic (11.5% vs 4.4%) [S12]. The Danish guideline requires at minimum mydriatic two-field disc- and macula-centred images [S2]; these sources differ on mydriasis.
• Macular oedema: optical coherence tomography (OCT) can add diagnostic accuracy when clinically significant DMO is suspected [S2].
• Smartphone imaging: pooled sensitivity 87% and specificity 94% for any DR, 91% and 89% for referral-warranted DR, but only 39% sensitivity for mild NPDR [S5].
• AI grading: in prospective real-world studies, AUROC was 0.959 for referable DR and 0.974 for vision-threatening DR [S11]. In developing countries, AI pooled sensitivity was 0.891 and specificity 0.936 for vision-threatening DR, with a very wide confidence interval for sensitivity (0.393 to 0.990) [S10]. Camera type affected performance [S10].

Manage it: running a programme at each resource tier

Before acting on this section, check the national and local guidelines, product information and regulatory approvals that apply where you practise. See the medical disclaimer.

Tier A (primary care / no specialist)

• Offer screening at the diabetes clinic; nearby screening facilities enable uptake [S4].
• Trained non-ophthalmologists and general physicians using portable fundus cameras showed good agreement with reference standards [S10].
• Educate every person with diabetes about DR: health education was linked to higher uptake (OR 4.3) [S4].
• Use reminders and other quality improvement strategies; both screening-specific and general diabetes care interventions worked, especially where attendance was low [S1].

Tier B (general eye clinic)

• Grade images, or supervise AI output, and run the referral pathway [S10] [S11].
• Dilate and re-image when images are ungradable [S3].
• Consider individualised intervals for low-risk patients only if local guidance allows: 24 to 48 months for no or mild NPDR, and 12 to 24 months for moderate NPDR with HbA1c ≤53 mmol/mol and blood pressure ≤130/80 mmHg [S2].

Tier C (specialist centre and planners)

• Treat referable and sight-threatening DR; confirm DMO with OCT [S2].
• Plan staff, training and imaging infrastructure, the main provider-side barriers [S4].
• AI and telemedicine screening were cost-effective in seven studies from China, Thailand, Brazil and India [S9]; in middle-income countries results were mixed, depending on compliance, glycaemic control and initial costs [S11].
• Validate any screening method locally, since most diagnostic accuracy data come from high-income settings [S3].

Refer when

Pitfalls

• Reporting ungradable images as normal [S3].
• Relying on single-field images: sensitivity for referable DR can fall to 76% [S6].
• Trusting smartphone imaging to detect mild disease [S5].
• Deploying AI without checking camera type and image quality locally [S10] [S11].
• Building screening without staff and treatment capacity downstream [S4].
• Assuming high-income attendance evidence transfers unchanged [S1].

Evidence notes

• Sources conflict on mydriasis: equal sensitivity without it in adults after excluding ungradable images [S3], better detection with it in children [S12], and the Danish guideline requires it [S2].
• The Danish guideline judged automated screening "not ready for implementation" in 2018 [S2]; later meta-analyses report high AI accuracy [S10] [S11].
• Diagnostic accuracy data come mostly from high-income settings [S3] [S4]; insights into low-income countries were limited [S11].
• No randomised trials compared screening strategies in children [S12].
• Global prevalence trends are hard to interpret because of differences in study design [S8].
• Not covered by these sources: treatment of DR and DMO, specific referral timeframes, and grader training standards.

Related articles

• Proliferative Diabetic Retinopathy: Laser, Anti-VEGF and Vitrectomy
• Artificial Intelligence in Retina and OCT
• Smartphone and portable fundus imaging

References

See the numbered source list below. Each marker [S#] links to its PubMed and DOI record.