A toddler sits in front of a tablet. The device watches where the child looks, how the face moves, how long attention holds. A few minutes later, a risk score appears. No questionnaire marathon, no months on a specialist waitlist.
For autism spectrum disorder (孤独症), speed is the whole game. The earlier a child is flagged, the earlier therapy can begin — and the better the odds of closing the gap with peers.
The four-year gap
China is home to more than 13 million people on the autism spectrum, by figures cited by Chinese disability and health authorities, with roughly 7 per 1,000 of children aged 0–6 affected. The painful part is timing: the average age of diagnosis in China is around four years, well past the golden window. Most experts agree meaningful intervention works best before age three, when the young brain is still most plastic.
The bottleneck is not ignorance but access. Specialist child psychiatrists are concentrated in big cities; standardized diagnostic assessments are slow and scarce. A fast, cheap, first-pass screen could change who gets seen.
What the camera actually measures
Autism's early signs are visual and behavioral: reduced eye contact, atypical attention, unusual facial expression, repetitive movement. An AI system built on computer vision can score these from ordinary video or an eye-tracking session — turning a 30-minute specialist evaluation into a 3-to-5-minute screen that a non-specialist can run.
One academy, two tools
The most developed example comes from the Institute of Medicine (Hangzhou), Chinese Academy of Sciences (中国科学院杭州医学研究所), which released an autism AI agent called StellarCareAI (星启AI). Built on a large model (大模型) trained on tens of thousands of research papers and guidelines, it is meant to support doctors, families, and researchers rather than replace them. Alongside it, the institute built an eye-tracking screening system that finishes an initial screen in 3–5 minutes, with pilot data reporting accuracy above 85% and sensitivity and specificity above 83%. In trials run through a maternity hospital in Quanzhou, the team collected data from more than 1,300 children and has since served 3,000-plus in pilot deployments, pushing the diagnosis window toward 18 months.
Separately, Xinhua Hospital affiliated with Shanghai Jiao Tong University built an AI screening tool, the iTracker "早瞰," that works from an ordinary camera in about 2 minutes, using eye-movement, micro-expression, and craniofacial features — a deliberately low-cost design meant to spread into community clinics.
The bottleneck is not the algorithm
Having a good screen is necessary but not sufficient. The hard parts are the ones technology does not solve:
- Reach. A tool only helps if it sits in the clinics children actually visit — maternal-and-child health centers, community hospitals, kindergartens.
- Follow-through. A risk score means nothing without a confirmed diagnosis and therapy behind it, and China's intervention capacity is thin.
- Trust. Families need to understand what a screen does and does not say, so a "higher risk" result leads to help, not panic.
Why screen so early
The brain's plasticity is highest in the first three years, which is precisely when autism is hardest for parents to spot and easiest for a camera to notice. A screen that flags risk at 18 months, rather than confirming at four years, can mean two extra years of therapy during the window when it changes the most. The tool does not have to be perfect to be useful; it only has to be fast and cheap enough to sit in front of every child.
The equity angle is real. Most of China's autism specialists sit in tier-one cities, while the bulk of children are born elsewhere. A camera-and-AI screen is cheap enough to place in a township maternal-and-child health center — exactly the setting where the four-year diagnosis gap is widest. Technology alone will not close it, but it removes one of the most stubborn bottlenecks: the very first look.
Trend reading
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, eye-tracking screening is most useful not as a verdict but as a fast, cheap funnel that routes the right children to confirmation and early therapy.
Honest limitations
The performance figures here — 85% accuracy, above-83% sensitivity and specificity, diagnosis pushed to 18 months — come from the institute's own pilot reporting and Chinese state-media coverage (CNR, China News), not from independent multi-center trials we verified. We have not confirmed the size or representativeness of the validation cohorts, how the tools perform across regions and subgroups, or their false-positive rate in real community settings, where over-referral can overload scarce diagnostic clinics. "Screening" is also explicitly not "diagnosis": a positive result should trigger confirmation by a specialist, not a label. We have not assessed regulatory clearance or how broadly either tool has actually deployed beyond pilots.
What readers can do now
- If you are a parent or carer, ask your maternal-and-child health center whether a quick AI or eye-tracking autism screen is offered at the 18–24 month check, instead of waiting for obvious signs.
- Treat any screen result as a prompt for confirmation, not a judgment — follow a higher-risk score with a specialist developmental assessment.
- Push for early intervention the moment risk appears, since the evidence for better outcomes is strongest in the youngest children, regardless of which tool raised the flag.