Researchers at the University of North Carolina and Emory University have received up to $35 million to build the world's largest artificial intelligence dataset for rare disease diagnosis, potentially transforming how millions of patients get answers. The four-and-a-half-year project, funded by the Advanced Research Projects Agency for Health (ARPA-H), will integrate clinical records, genetic data, and patient information across approximately 2,700 rare diseases to train AI models that can help doctors diagnose conditions faster and more accurately.

For patients living with rare diseases, the stakes are enormous. More than 10,000 rare diseases affect an estimated 350 million people worldwide, including as many as one in 10 Americans. Yet the diagnostic journey is agonizing: patients wait an average of six years for an accurate diagnosis, and some wait decades. By the time they reach a specialist, disease may have progressed irreversibly, leading to inappropriate treatment, unnecessary medical costs, and lost opportunities for early intervention.

Why Is Rare Disease Diagnosis So Difficult?

The core problem is fragmentation. Information about rare disease patients is scattered across different hospitals, clinics, insurance companies, and research centers. Because each rare disease affects relatively few people, no single institution has enough data to spot patterns or train effective AI systems. Genomic testing has improved diagnosis in some cases, but many genetic test results remain inconclusive without additional clinical context, while other rare diseases have no known genetic cause at all.

Geographic distance, insurance barriers, and limited access to specialty care compound the problem. Patients in rural areas or underserved communities may never reach a rare disease expert, leaving them without answers. This data scarcity also slows drug development: with only about 5% of rare diseases having an approved therapy, researchers struggle to recruit enough patients for clinical trials.

How Will This AI Initiative Work?

The UNC and Emory teams will lead efforts to acquire and integrate data from multiple sources, creating what will become the largest AI-ready dataset for rare diseases ever assembled. The resource will combine:

• Health Records: Clinical data from patient visits, diagnoses, and treatment histories across institutions
• Genetic Information: Genomic sequencing results and molecular data that can reveal disease causes
• Patient-Reported Data: Surveys and direct input from patients about symptoms, family history, and quality of life
• Medical Imaging and Video: Visual data that can be analyzed by AI to detect disease signs
• Insurance Claims: Real-world data on how diseases progress and what treatments patients receive

To protect privacy, all identifying information will be removed before data enters the system. Access will be tiered based on data sensitivity, with some datasets available publicly and others requiring data use agreements and additional safeguards.

"Our aim is to create a large-scale dataset spanning approximately 2,700 of the more prevalent and potentially treatable rare diseases to train models that can then be deployed in diagnostic settings where we don't have such rich data," explained Melissa Haendel, PhD, FACMI, Sarah Graham Kenan Distinguished Professor in the Department of Genetics at the UNC School of Medicine and lead scientist on the project.

Melissa Haendel, PhD, FACMI, Sarah Graham Kenan Distinguished Professor, UNC School of Medicine

Once trained on this rich dataset, AI models can be deployed in less sophisticated settings, such as primary care clinics or community health centers, where doctors lack access to rare disease specialists. The system could help clinicians interpret genetic test results more effectively, identify patients who need specialist referrals, and move patients toward diagnosis faster.

What Real-World Problems Will This Solve?

The initiative addresses several critical bottlenecks in rare disease care. First, it will accelerate diagnosis by helping non-specialists recognize rare disease patterns they might otherwise miss. Second, it will enable researchers to identify eligible patients for clinical trials more efficiently, removing one of the biggest barriers to drug development. Third, it will reveal how rare diseases develop and progress across populations, uncovering new insights into disease biology that could lead to new treatments.

"At UNC, industry partners approach us regularly with trials on a rare disease. The challenge is that finding eligible patients can be extremely difficult. We currently have no means to find those patients. By securely bringing together more data, we hope to develop algorithms to identify trial participants more efficiently, accelerate research, and expand access to clinical care," said Haendel.

Melissa Haendel, PhD, FACMI, Sarah Graham Kenan Distinguished Professor, UNC School of Medicine

The ARPA-H initiative is part of a broader federal investment in rare disease AI. The agency announced up to $98.5 million in total awards over 4.5 years across multiple teams, including performers focused on building secure data platforms, collecting multimodal patient data, and developing AI tools for direct-to-patient use.

Who Is Supporting This Effort?

The project draws on a robust public-private partnership spanning academic institutions, patient advocacy organizations, health data companies, and technology firms. Collaborators include Johns Hopkins University, the University of California at San Francisco, the University of Iowa, and Queen Mary University of London. Patient advocacy groups like Global Genes and the National Organization for Rare Diseases (NORD) will ensure the solutions reflect the needs of the rare disease community.

Major technology companies have pledged in-kind support, including OpenAI, Anthropic, Amazon Web Services, and Google, which will provide computing credits and infrastructure to support the AI development.

What Does This Mean for Patients?

If successful, this initiative could fundamentally change the rare disease experience. Patients could receive diagnoses in years rather than decades. Doctors in underserved areas could access the diagnostic expertise of specialists through AI-powered tools. Clinical trials could recruit participants faster, accelerating the path to new treatments. And healthcare systems could intervene earlier, reducing the irreversible damage that delayed diagnosis causes.

The project also signals a broader shift in how medicine approaches rare diseases. Rather than treating each rare condition in isolation, researchers are building infrastructure to learn across thousands of diseases simultaneously, creating a foundation for precision medicine that could benefit patients far beyond the rare disease community.