- AI model predicts heart damage in Kawasaki patients
- Helps doctors act sooner for high-risk children
- Not yet in hospitals — still being tested
This tool could help doctors spot which kids face the highest risk of heart damage from Kawasaki disease — before it happens.
Every year, hundreds of children show up at hospitals with raging fevers, red eyes, and swollen hands. No infection explains it. Parents worry. Doctors race to diagnose. Often, it’s Kawasaki disease — a rare but serious illness that can hurt the heart.
Most kids recover fully. But some develop lasting heart problems. The danger lies in the coronary arteries — the tiny blood vessels that feed the heart muscle. If they swell or form aneurysms, the child could face heart attacks even as a young adult.
Right now, doctors treat all severe cases with strong medications. But not every child needs the strongest drugs. And by the time heart damage shows up on an ultrasound, it may be too late to stop it.
What if doctors could know — early — who’s most at risk?
The surprising shift
For years, doctors relied on basic blood tests and clinical signs to guess who might develop heart complications. Tools like the Kobayashi score helped, but they weren’t perfect. Many kids got aggressive treatment just in case. Others slipped through the cracks.
But here’s the twist: a new AI-powered tool may do a much better job.
Using data from over 1,000 children with Kawasaki disease, researchers built a machine-learning model that predicts which patients are most likely to develop coronary artery lesions (CAL) — with high accuracy.
This isn’t just another checklist. It learns from patterns in real patient data — things like age, fever length, platelet count, and liver enzyme levels — and weighs them in ways humans can’t easily see.
What scientists didn’t expect
The model didn’t just match current methods — it outperformed them.
The best-performing algorithm, called CatBoost, correctly identified high-risk patients 95% of the time in initial testing. Even in outside hospitals’ data, it stayed strong — predicting risk with 87–89% accuracy.
That’s a big deal. Because catching risk earlier means doctors can act sooner.
Imagine a child arrives with fever and rash. The doctor enters their symptoms and lab results into a hospital system. Within seconds, the AI returns a risk score.
High risk? The team starts stronger treatment immediately.
Low risk? They avoid unnecessary drugs and side effects.
It’s like a weather forecast for the heart — warning of storms before they hit.
Think of the body like a city’s traffic system. Blood vessels are roads. The heart is downtown. In Kawasaki disease, inflammation acts like a massive traffic jam — blocking key routes and damaging the roads themselves.
Doctors want to know: which jams will cause permanent road damage?
The AI model acts like a smart traffic monitor. It doesn’t just look at one camera. It pulls in data from hundreds of past traffic events — weather, time of day, accident reports — and finds hidden patterns.
In this case, the “cameras” are 41 different patient details: from white blood cell count to how long the fever lasted.
The model weighs each factor, learns which combinations lead to heart damage, and builds a prediction.
Some factors mattered more than expected. For example, certain liver enzyme changes and early platelet drops were strong signals — even when they looked normal to doctors.
Better predictions, faster
The study used records from 1,037 children with Kawasaki disease treated at a major hospital in China. Researchers split the data: 70% to train the model, 30% to test it.
They tested 10 different AI models. CatBoost won — not just in speed, but in accuracy.
In the internal test, it correctly flagged 72% of high-risk kids (sensitivity) and correctly ruled out 85% of low-risk ones (specificity). In external testing — using data from other hospitals — it did even better at ruling out risk, with 95% specificity.
That means fewer false alarms — and fewer kids getting unnecessary treatment.
This doesn’t mean this treatment is available yet.
But there’s a catch.
The model was trained on Chinese children. Kawasaki disease can act differently in other populations. What works in沈阳 may not work in Chicago.
Also, this was a retrospective study — meaning it looked at past data. It hasn’t been tested in real time, with live patients.
And while the AI makes predictions, it doesn’t replace doctors. It’s a tool — like a GPS for treatment decisions — but the driver still needs to steer.
Where this fits
Experts say tools like this are the future — but only if they’re used wisely.
“This kind of model could help standardize care, especially in hospitals without pediatric heart specialists,” said one independent researcher familiar with the work.
Right now, only a few major centers have the experience to catch subtle signs of heart risk. An AI tool could level the playing field — helping smaller hospitals make smarter calls.
But it must be tested across diverse groups first.
If your child has Kawasaki disease, this tool isn’t available yet — and won’t be for a while.
But it offers hope: one day, doctors may predict heart risk faster and more accurately.
For now, early diagnosis and IVIG treatment within 10 days remain the best defense.
Talk to your doctor about your child’s risk factors. Ask about echocardiograms. And know that research like this is moving us closer to personalized care.
Not perfect — but promising
The study had limits. It only included Chinese patients. It used past data, not real-time decisions. And it didn’t include genetic or imaging data that might improve accuracy.
Still, the results are among the strongest seen for this type of prediction.
Researchers plan to test the model in real-time trials and adapt it for use in other countries. If proven effective, it could become part of hospital software within the next few years — helping doctors protect children’s hearts before damage ever starts.