- AI spots a rare but serious Sjögren’s subtype
- Helps people with dry eyes and mouth get better care
- Still in testing — not yet in clinics
This could help doctors catch dangerous complications earlier.
Dry eyes. Dry mouth. Fatigue. For many with Sjögren’s syndrome, these are daily struggles. But for some, there’s more hiding underneath. A small group also has a rare antibody that raises their risk for lung and blood vessel problems. Now, a new AI-powered tool may help doctors find these high-risk patients faster — before serious issues arise.
Sjögren’s syndrome affects over 1 million people in the U.S. It’s an autoimmune disease where the body attacks its own moisture-producing glands. Most patients manage symptoms with eye drops or saliva substitutes. But about 10% have anti-centromere antibodies (ACA). These patients face higher odds of developing pulmonary hypertension — a dangerous lung condition. The problem? ACA-positive cases are hard to spot early. Blood tests exist, but they’re not always ordered. And symptoms often don’t show until damage is done.
Doctors used to rely on routine blood panels and patient history to guess who might have ACA. But patterns are subtle. One person may have dry eyes and Raynaud’s (cold fingers that turn white or blue). Another may show odd liver markers or strange immune proteins. Alone, none scream “ACA.” Together, they might — but only if someone connects the dots.
But here’s the twist: computers might now do that better than humans.
Using data from 616 Sjögren’s patients across multiple clinics, scientists trained six different AI models to detect who had ACA. The best one — called GBDT — learned to weigh dozens of clues: age, specific antibodies, immune protein levels, and symptoms like Raynaud’s phenomenon.
Think of it like a lock with many tumblers. No single key opens it. But when the right combo clicks — age plus Raynaud’s plus low IgM plus anti-AMA-M2 — the AI knows the lock is likely ACA-positive.
The model didn’t just memorize facts. It found hidden links doctors had missed. For example:
- Raynaud’s is more telling in older patients
- Anti-AMA-M2 matters most when IgM levels are low
- Certain anti-SSA antibodies add weight only in specific combinations
These aren’t linear rules. They’re complex interactions — the kind our brains struggle with, but AI handles easily.
This doesn’t mean this treatment is available yet.
The AI was tested on real patient records — 81 with ACA, 535 without. It correctly identified ACA-positive cases 81% of the time (AUC 0.811), which is strong for medical prediction tools. It also caught three out of every four ACA cases (75% sensitivity), while avoiding false alarms.
To put that in perspective: current screening might miss half of these cases without targeted testing. This tool could double the detection rate — before lung damage starts.
That’s not the full story.
What makes this tool different isn’t just accuracy. It’s explainability. Many AI systems are “black boxes” — they give answers but don’t say why. This one uses SHAP analysis, a method that shows which factors pushed the decision.
So a doctor sees not just “high risk,” but why: “Raynaud’s + age 58 + low IgM + anti-AMA-M2.” That builds trust — and guides next steps.
Experts say this approach could shift how we manage autoimmune diseases. Instead of waiting for symptoms, we might predict them. “We’re moving from reaction to prediction,” said one researcher not involved in the study. “AI won’t replace doctors — but it can highlight patterns we overlook.”
For patients, this means hope for earlier action. If your doctor knows you’re ACA-positive early, they can:
- Order lung function tests
- Watch for blood pressure changes
- Start preventive care sooner
But here’s the catch: this tool isn’t in clinics yet. It worked on past data — not live patients. And it needs validation in larger, more diverse groups.
Also, it was tested only in people already diagnosed with Sjögren’s. It won’t help find Sjögren’s itself — only spot a risky subtype once diagnosed.
The road ahead includes testing the AI in real-time settings. Can it work in busy clinics? Will it improve outcomes? Trials are needed. Regulatory approval will take time. But researchers are optimistic. With enough validation, this could become a built-in alert in electronic health records — quietly flagging high-risk patients during routine visits.
No one’s claiming a cure. But catching ACA earlier? That could save lives.