Researchers at Shanghai Jiao Tong University have built an AI system that outperforms experienced physicians at diagnosing rare diseases. DeepRare, published in Nature this week, correctly identified diseases on its first attempt 64.4% of the time compared to 54.6% for doctors with 10+ years of experience.
Why This Matters
Getting a rare disease diagnosis typically takes years. A survey by the China Alliance for Rare Diseases covering 20,000 patients found 42% experienced misdiagnosis, with an average diagnostic delay of 4.26 years. During that time, patients often receive ineffective treatments or none at all.
Rare diseases affect 300-400 million people worldwide but individually affect so few patients that most doctors rarely encounter them. There are over 7,000 known rare diseases, making it nearly impossible for any physician to maintain expertise across all of them.
How DeepRare Works
Unlike typical medical AI that pattern-matches symptoms to diseases, DeepRare uses what researchers call an “agentic” architecture. It thinks like a doctor: forming hypotheses, testing them against evidence, and revising conclusions.
The system orchestrates over 40 specialized tools that analyze:
- Patient DNA data
- Medical literature databases
- Handwritten clinical notes
- Symptom profiles
A central AI coordinator runs these tools and synthesizes their outputs. Critically, it generates traceable diagnoses with complete evidence chains, so physicians can verify the reasoning rather than just accept a black-box answer.
The Numbers
The team validated DeepRare against 6,401 historical clinical cases, where it outperformed 15 existing diagnostic tools and often identified diseases earlier than the original treating physicians.
In a direct head-to-head comparison on 163 difficult cases:
- DeepRare: 64.4% accuracy on first diagnosis
- Experienced physicians (5 doctors with 10+ years each): 54.6% accuracy
When allowed three guesses instead of one, DeepRare typically had the correct diagnosis in its top three suggestions.
Perhaps more importantly, when 10 rare disease specialists reviewed DeepRare’s step-by-step reasoning, they agreed with its logic 95.4% of the time.
Without genetic sequencing data, using only clinical symptoms, DeepRare achieved 57.18% accuracy - a 23.79 percentage point improvement over the previous best international tool. With genomic data added, accuracy in complex cases exceeded 70.6%, significantly outperforming the widely-used Exomiser tool’s 53.2%.
Already in Clinical Use
DeepRare isn’t just a research project. The system has been deployed on an online diagnostic platform since July 2025, with over 600 medical institutions worldwide now registered.
It’s currently undergoing internal testing at Xinhua Hospital (affiliated with Shanghai Jiao Tong University’s School of Medicine), though the researchers haven’t published results from that real-world deployment yet.
The Fine Print
There are limitations to consider. The head-to-head comparison used retrospective cases where the correct diagnosis was already known, not prospective real-world scenarios where outcomes remain uncertain.
The system also requires structured input - either clean clinical data or genetic sequencing results. Many rare disease patients first present with vague symptoms to general practitioners who may not know what data to collect.
And while DeepRare excels at diagnosis, it doesn’t prescribe treatments. For many rare diseases, diagnosis is just the beginning of a difficult journey with limited therapeutic options.
Still, cutting the average diagnostic delay from 4+ years to potentially weeks or months could transform outcomes for millions of patients currently trapped in diagnostic limbo.