- Design
- Systematic review and diagnostic accuracy meta-analysis
- Population
- 17 studies reviewed; 4 pooled for screening
- Primary outcome
- Accuracy of AI for diabetic retinopathy on ultrawide-field images
- Effect
- Sensitivity 85.0%; specificity 72.5%; AUC 0.870
A systematic review identified 17 studies applying artificial intelligence to ultrawide-field fundus images for diabetic retinopathy, four of which were pooled for screening accuracy. All used one manufacturer's imaging platform.
Pooled sensitivity was 85.0% and specificity 72.5%, with an area under the curve of 0.87. Ultrawide-field imaging can be done without dilating the pupil and captures more of the peripheral retina, which suits screening.
With specificity around 73%, roughly one in four people without retinopathy would be flagged, which would load referral pathways. The evidence base is small and from a single platform, so generalisability is uncertain.
For programmes in India and elsewhere considering AI-assisted screening, the lesson is to validate any system locally and plan for the referral volume its false positives will create.
- Validate any AI screening system locally before relying on it.
- Plan referral capacity for the false positives a 73% specificity implies.
- Use human grading of positives where resources allow.
- Do not assume results transfer across imaging platforms.
Why it matters
Low specificity can overwhelm referral clinics, eroding the efficiency AI is meant to bring.
Don't overread it
Only four pooled studies, all from one platform; the estimates may not hold elsewhere.
The statistics, in plain English
Sensitivity of 85% means about 15 in 100 people with retinopathy would be missed. Specificity of 72.5% means about 27 in 100 without it would be wrongly flagged.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for ophthalmology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free