AI and Rare Disease Diagnosis: How DeepRare Is Ending the Diagnostic Odyssey

For 300 million people worldwide living with a rare disease, the path to diagnosis is rarely short. The average patient waits more than five years. One in four consults eight or more specialists before receiving a confirmed answer. Every year of delay means missed treatment windows, irreversible disease progression, and families suspended in diagnostic limbo.

A study published in Nature in February 2026 suggests that AI may finally be shortening that wait – and in some cases, outperforming the specialists who carry it.

What Is DeepRare – and Why Does It Matter?

DeepRare is a multi-agent AI system developed by a joint team at Shanghai Jiao Tong University's School of Artificial Intelligence and Xinhua Hospital. It was designed specifically for rare disease diagnosis, a domain where traditional AI has consistently struggled.

The core challenge is structural. Over 7,000 known rare diseases exist, each with highly variable symptom profiles that frequently overlap with common conditions. Individual diseases have so few cases that supervised learning models can't be trained effectively. Hundreds of new rare genetic diseases are identified every year. And clinical deployment demands transparent reasoning -- clinicians need to understand why an AI system has reached a conclusion, not just what it concluded.

DeepRare addresses all of this through an architecture built around explainability.

Rather than returning a black-box prediction, the system integrates more than 40 specialist digital tools and connects to medical literature databases, rare disease repositories, and real-world case archives in real time. When a patient's information is entered – whether as free-text clinical notes, standardised phenotype terms, or raw genomic sequencing data – DeepRare generates a ranked list of diagnostic candidates, each accompanied by a transparent reasoning chain linking every inference step to a verifiable medical source.

A self-reflection loop re-checks the system's own hypotheses before producing a final output, reducing errors and hallucinations. Ten rare disease physicians reviewed 180 of DeepRare's reasoning chains and agreed with 95.4% of the evidence citations.

The Head-to-Head Results

The benchmark test at Xinhua Hospital was direct. DeepRare was compared against five experienced rare disease physicians, each with more than a decade of clinical practice, using 163 real patient cases. Both the AI and the physicians received identical clinical information. The physicians had access to search engines and reference materials.

DeepRare identified the correct diagnosis as its top prediction in 64.4% of cases. The physicians achieved 54.6%.

That ten-percentage-point gap matters because the physicians were not generalists – they were specialists. The test was designed to put DeepRare in the hardest possible conditions, and it still outperformed.

Across a broader evaluation of more than 6,400 cases spanning nine datasets, 14 medical specialties, and 2,919 rare diseases – drawn from clinical centres in Germany, the United States, and China -- DeepRare consistently outperformed every comparison group.

When genomic sequencing data was added alongside clinical phenotype information, accuracy improved further. In the Xinhua Hospital dataset, correct top-1 identification reached 69.1%, compared to 55.9% for the leading bioinformatics tool. Across phenotype-only inputs, DeepRare achieved a Recall@1 score of 57.18% – a 23.79 percentage-point improvement over previous state-of-the-art models.

The system has already moved beyond benchmarking. Since July 2025, DeepRare has been deployed as a web-based diagnostic support tool, with more than 600 medical institutions worldwide registered on the platform.

The Diagnostic Odyssey, by the Numbers

The scale of the problem DeepRare addresses is significant.

Rare diseases collectively affect more than 300 million people worldwide. Around 80% have a genetic origin. The average diagnostic journey lasts five to seven years. One large survey found that 25% of patients reported eight or more consultations before receiving a confirmed diagnosis. Each year of delay carries real clinical cost: missed treatment windows, unnecessary investigations, inappropriate therapies, and families navigating the healthcare system without a map.

The U.S. Advanced Research Projects Agency for Health (ARPA-H) recognised this gap explicitly when it launched its RAPID program – Rare Disease AI/ML for Precision Integrated Diagnostics. RAPID's stated goal is to collapse the diagnostic timeline from years to months or days. The program is building a large curated dataset of longitudinal rare disease patient data designed to train and validate AI diagnostic tools at population scale, including direct-to-patient tools that can work outside specialist centres.

"By leveraging AI, we can expand access to rare disease expertise and greatly reduce time to diagnosis, from years to months or even days," said RAPID Program Manager Scott Gorman. "AI-enabled support tools allow us to sift through the 'haystack' of patient data more efficiently and pinpoint the 'needles' of rare diseases."

When Faster Diagnosis Changes Everything: The Story of Elly Krueger

Faster, more accurate AI diagnosis matters most when there is something to do with the answer. The story of Elly Krueger shows what becomes possible when AI-enabled diagnosis, coordinated research, and accelerated gene therapy development converge at the speed patients actually need.

In February 2024, Elly was diagnosed at eight months old with NEDAMSS – Neurodevelopmental Disorder with Regression, Abnormal Movements, Loss of Speech, and Seizures – a progressive neurological condition caused by mutations in the IRF2BPL gene. Only a small number of cases have been identified worldwide. There were no approved treatments and no established path forward.

Her parents launched Elly's Team immediately, connecting with a multidisciplinary group of researchers, clinicians, and regulatory advisors to pursue a gene therapy approach. Safety studies, drug manufacturing, and FDA regulatory review ran simultaneously rather than sequentially. On 3 April 2025, Elly became the first child to receive an IRF2BPL gene replacement therapy – just 14 months after diagnosis. One month post-treatment, the therapy was safe and well tolerated.

"In the future," said Michelle Krueger, "another family will sit in the hospital and receive the same diagnosis, but their doctor will tell them there is a path to treatment."

The Gene Therapy Wave Behind Individual Stories

Elly's case sits inside a much larger momentum shift in rare disease treatment.

In April 2025, Abeona Therapeutics received FDA approval for Zevaskyn, a cell-based gene therapy for recessive dystrophic epidermolysis bullosa (RDEB) – a severe inherited skin disorder where mutations in the COL7A1 gene cause extreme skin fragility, deep blistering, chronic wounds, and scarring. The Phase 3 VIITAL trial found that 81% of wounds treated with a single surgical application achieved 50% or more healing at six months, compared to 16% of control wounds.

Also in 2025, the FDA approved the first gene therapy for Wiskott-Aldrich syndrome – a rare X-linked primary immunodeficiency characterised by low platelet count, eczema, and recurrent infections. The approval went to Fondazione Telethon ETS, a non-profit, confirming that the path from laboratory to patient does not require a large commercial sponsor.

Pricing remains a structural challenge. Some approved gene therapies now cost several million dollars per patient – rational economics for diseases affecting tens of patients annually, but a framework that puts treatments out of reach for most health systems. Platform manufacturing approaches are being explored as a route to lower costs, though distribution of progress remains deeply uneven.

What DeepRare Cannot Yet Do

The Nature study's authors and independent reviewers are clear on this point: DeepRare's benchmarks are retrospective. All test cases came with confirmed diagnoses. Real-world clinical deployment – where information arrives incomplete, contradictory, and in motion – is a different environment.

Occasional hallucinated or irrelevant citations were identified in the reasoning chain review. The system currently performs best when high-quality phenotype data is available. In lower-resource clinical settings, where documentation is inconsistent or genomic sequencing is unavailable, accuracy will differ from benchmark results.

The researchers plan to validate DeepRare prospectively using 20,000 real-world cases. Until that validation is published, the system's performance in routine clinical practice remains to be confirmed.

Human oversight remains essential. DeepRare is designed as a diagnostic copilot – a tool that supports clinicians, not one that replaces their judgment. Data privacy considerations for systems operating across international clinical datasets also require ongoing attention.

A Proving Ground for All of Medicine

Rare diseases have always contributed disproportionately to medical innovation.

Early clinical successes in genomic sequencing, mRNA technologies, and gene therapy all emerged from rare, single-gene disorders before reaching cancer and cardiovascular disease. The tools validated today for rare disease diagnosis will eventually reach the broader healthcare system.

DeepRare's research team made this explicit in the Nature paper. The system's consistent performance across 14 medical specialties suggests its potential as a decision-support tool not just for rare disease specialists, but for general physicians who encounter these conditions without the background to recognise them.

Rare Disease Day – observed on 28 February across 106 countries – is a reminder that 300 million people with rare conditions are not a niche concern. They represent the hardest diagnostic problems medicine faces. When those problems get solved, the solutions rarely stay within their original domain.

FAQ: AI and Rare Disease Diagnosis

What is DeepRare and how does it diagnose rare diseases?

DeepRare is a multi-agent AI system developed by researchers at Shanghai Jiao Tong University and Xinhua Hospital. It integrates more than 40 specialist tools, connects to global medical literature and rare disease databases in real time, and analyses patient data -- including free-text clinical notes, standardised phenotype terms, and genomic sequencing data -- to generate a ranked list of diagnostic candidates. Each candidate is accompanied by a transparent reasoning chain linking every inference step to a verifiable medical source.

How accurate is AI at diagnosing rare diseases compared to doctors?

In a head-to-head study published in Nature in February 2026, DeepRare identified the correct diagnosis as its top prediction in 64.4% of cases, compared to 54.6% for experienced rare disease physicians with over a decade of clinical practice. Across a broader evaluation of more than 6,400 cases covering 2,919 rare diseases, DeepRare consistently outperformed all comparison groups, including 15 established diagnostic support tools.

How long does it take to get a rare disease diagnosis?

The average rare disease patient waits more than five years for a correct diagnosis. One in four patients consults eight or more specialists before receiving a confirmed answer. The U.S. ARPA-H RAPID program was specifically designed to collapse this timeline using AI, with a stated goal of reducing the diagnostic journey from years to months or days.

What are the limitations of AI in rare disease diagnosis?

Current benchmarks for systems like DeepRare are based on retrospective data – cases with confirmed diagnoses. Real-world clinical performance, where information is incomplete and arriving in real time, may differ. Occasional citation errors have been identified in AI reasoning chains. Performance depends on the quality of available patient data, and genomic sequencing data is not always available. Human oversight remains essential; these systems are designed to support clinical judgment, not replace it.

What gene therapies are now available for rare diseases?

FDA approvals for rare disease gene and cell therapies have accelerated significantly. In 2025, the FDA approved Zevaskyn (Abeona Therapeutics) for recessive dystrophic epidermolysis bullosa and the first gene therapy for Wiskott-Aldrich syndrome (Fondazione Telethon ETS). Several dozen cell and gene therapies have now received FDA approval across a range of rare disease indications.

How can AI help reduce the rare disease diagnostic odyssey?

AI systems like DeepRare can analyse large volumes of clinical and genomic data simultaneously, identify pattern matches across thousands of rare diseases, and provide ranked diagnostic hypotheses with transparent reasoning – capabilities that exceed what is feasible for any individual clinician. Tools designed for primary care settings can extend rare disease screening beyond specialist centres, reaching patients earlier in their diagnostic journey.

Stay ahead of what’s next in healthcare.

Healthy Innovations is my weekly newsletter delivering strategic analysis of emerging biotech and digital health.

No spam. Unsubscribe anytime.