A grandmother in Beijing's Haidian district sits for a photo of her eye. Two minutes later, a result is on a doctor's screen. By the end of the day, an ophthalmologist has reviewed a batch of them. No hospital trip, no specialist wait — just a camera in the local clinic and an AI (人工智能) that reads retinas.
This is not a pilot. According to the Beijing Municipal Health Commission, the district's retinal-AI screening platform has already run more than 210,000 screenings.
The setup
Haidian's "眼底人工智能诊疗协作中心平台" (retinal AI diagnostic collaboration center) covers all 49 community health centers in the district. The workflow the Commission describes is bluntly efficient: one minute to shoot, two minutes for a result, and an ophthalmologist completes batch review within half a day.
The system behind it was presented by Beijing Airdoc (鹰瞳科技), a medical-AI company, and it is built for the bottleneck it targets: too few eye doctors for too many at-risk residents.
The numbers, as reported
From the Health Commission's March 2026 briefing:
- 210,800 screenings completed (21.08万人次).
- Over 24,000 high-risk positive cases flagged (检出高风险阳性病例2.4万余例).
- 390 residents referred onward for treatment.
- 99 eye surgeries or treatments completed.
The loop is closed: community screening → referral → surgery → community follow-up. That last step matters, because follow-up is where screening programs usually fall apart.
Why the eye is a good first front
The retina is one of the few places a camera can see small blood vessels directly, which makes it useful for catching diabetic retinopathy and glaucoma early — before vision loss. In an aging society with rising diabetes, that is exactly the kind of high-volume, protocol-driven check an AI can do consistently, leaving the doctor to judge the hard cases.
Diabetic eye disease is a leading cause of preventable blindness, and most of its damage is silent until it is advanced. A camera in every clinic that flags the worried cases is a structural answer to a specialist shortage — not a replacement for the specialist, but a force-multiplier for one.
Haidian is pairing this with other AI tools
The retinal program is one tile in a larger mosaic. The same Health Commission briefing describes a pediatric large model (大模型) — the "福棠·百川" AI pediatrician, community version — an AI doctor assistant inside the "Haidian Health" mini-program, and voice-assisted medical records in clinics. The pattern is the same everywhere: AI handles the repeatable first pass, humans handle the decision.
What "AI in action" means here
- The AI does not diagnose. It screens and prioritizes; an ophthalmologist confirms.
- The value is access: 49 clinics, not one hospital, become the point of care.
- The metric that counts is the referral-and-treatment chain, not the screen count alone.
What it takes to replicate this
The Haidian model is replicable only where the boring parts already exist: a clinic with a camera, a stable internet link to a backend, and — critically — an ophthalmologist willing to do batch review. The AI is the cheap, scalable part; the specialist's attention is the scarce resource the system is designed to protect. Communities without that referral backbone will struggle to close the loop, which is why the Health Commission's emphasis on the full screening-to-follow-up chain matters more than the screen count.
The quiet cost advantage
A community-clinic camera plus a cloud model costs a fraction of a specialist visit, and because the AI runs the same protocol on every patient, the per-screen marginal cost keeps falling as volume rises. That is the arithmetic that makes 210,000 screenings plausible in one district: the expensive resource — the ophthalmologist — is spent only on the cases the AI escalates, not on the full population. It is the opposite of the usual healthcare cost curve, where more screening means more specialist hours. Here, more screening means more automation of the routine pass, and the human time is reserved for judgment.
Honest limitations
- This is one district in Beijing (Haidian), an affluent, tech-dense area — not a national average. Results may not transfer to under-resourced regions without similar clinic infrastructure.
- Screening is not diagnosis. A "high-risk positive" triggers review; we do not know final confirmed-disease rates from the briefing.
- The operating numbers come from a government briefing that cites the vendor (Airdoc) presenting the platform; we did not see an independent clinical audit of sensitivity/specificity.
- Long-term outcomes — did earlier detection actually preserve sight or cut costs — are not yet measured in the public data.
What readers can do now
- If you are in or near Haidian, ask your community health center whether retinal-AI screening is available; it is a low-cost check worth doing if you have diabetes or a family eye-disease history.
- For health-system watchers: the model to study is "community capture + AI triage + specialist batch review + closed follow-up" — the follow-up closure is the hard part.
- For builders: this is a template for any high-volume screening (retina, lung CT, skin) where the bottleneck is specialist time, not camera hardware.