DeepRare, an agentic AI system integrating 40 specialised tools, outperformed medical specialists in identifying rare conditions in a head-to-head study published in Nature.
For millions of people with rare diseases, the path to diagnosis is a labyrinth. Patients bounce between generalist GPs and specialists across years, sometimes decades, piecing together symptoms that fall outside textbook presentations.
Eighty per cent of rare diseases have a genetic origin, yet most go undiagnosed until too much biological damage has occurred. The bottleneck is not lack of data, it’s finding the needle in the medical haystack.
A new study published in Nature this month suggests that artificial intelligence may accelerate that hunt. Researchers at Shanghai Jiao Tong University’s School of Artificial Intelligence and Xinhua Hospital developed DeepRare, an AI system designed to mimic how human doctors reason through diagnostic uncertainty.
In a head-to-head comparison with five experienced physicians, each with more than a decade of practice, the system achieved higher accuracy across the board.
The numbers are striking. DeepRare correctly identified the disease on its first suggestion 64.4 per cent of the time, compared to 54.6 per cent for the doctors. When given three suggestions instead of one, the AI system achieved diagnostic success in 79 per cent of cases versus 66 per cent for the human specialists.
Crucially, the physicians endorsed the AI’s reasoning 95.4 per cent of the time, suggesting the system not only reaches correct conclusions, but does so in ways that experienced clinicians find persuasive and medically sound.
What distinguishes DeepRare from earlier diagnostic AI is its architecture. Rather than applying a black-box classification model, the system integrates 40 specialised digital tools and follows an explicitly reasoned workflow.
It forms diagnostic hypotheses, tests them against patient evidence, searches global medical literature databases, analyses genetic variants, and revises its conclusions iteratively before ranking possibilities.
The process mirrors the cognitive steps a human diagnostician takes, but with access to the entirety of medical knowledge and computational speed humans cannot match.
The system has already moved beyond the laboratory. Since July 2025, DeepRare has been deployed on an online diagnostic platform, with more than 600 medical institutions worldwide registered to use it.
The research team plans to validate the system further using 20,000 real-world cases and to launch a global rare disease diagnostic alliance. Notably, the authors emphasise that the system is not intended to replace clinicians, but to augment diagnostic workflows, a safeguard that acknowledges both the technical limits of AI and the irreducible human element in medicine.
The implications for patients are profound. Approximately 300 million people worldwide are affected by rare diseases, and the average diagnostic odyssey stretches to five years or longer.
Each year of diagnostic delay is a year of uncertainty, wrong treatments, and accumulating organ damage. An AI system that can trim weeks or months from that timeline, and surface possibilities that might otherwise be overlooked, could reshape the early experience of living with a rare condition.
