ARPA-H Commits $98.5 Million to Speed Up Rare Disease Diagnosis with AI

If you or someone you love has ever chased a diagnosis for years, bouncing between specialists, running out of savings, and watching a condition worsen while doctors guessed, you already understand the problem ARPA-H is trying to solve. Rare disease patients often wait five to seven years for an accurate diagnosis. Some never get one at all. The Advanced Research Projects Agency for Health, the innovation arm of the Department of Health and Human Services, is betting that artificial intelligence and better data sharing can cut that wait dramatically.

This week ARPA-H announced four contracts worth up to $98.5 million combined over nearly five years. The program is called RAPID, short for Rare Disease AI/ML for Precision Integrated Diagnostics. Its goal is straightforward even if the science behind it is not: build the infrastructure needed to let AI tools actually help diagnose rare diseases, rather than stumbling over the scattered, incomplete data that has held the field back.

Think of it like trying to solve a jigsaw puzzle when the pieces are locked in different rooms across the country, each guarded by a different institution with its own rules about who can look inside. That is roughly the state of rare disease data today. Genetic test results sit in one system. Clinical records sit in another. Patient-reported symptoms live somewhere else entirely, often in a notebook or a support group forum rather than a hospital chart. RAPID aims to unify those pieces into a single, privacy-protected resource that AI systems can actually learn from.

Building the Infrastructure

The University of North Carolina will lead the largest of the four efforts, building what ARPA-H describes as the largest real-world rare disease dataset ever assembled. UNC will work with a wide coalition of academic institutions, health data companies, and patient advocacy groups to weave together clinical, genomic and patient-reported information while preserving privacy.

Three other organizations round out the effort, each tackling a different piece of the puzzle. Sage Bionetworks, a data science nonprofit, will build a secure platform for testing and comparing AI models against common standards, essentially a referee system that lets researchers see which tools actually work. FDNA, which specializes in AI imaging analytics, will develop tools that gather photos, videos, voice recordings and patient concerns directly from clinics and patients themselves. Probably Genetic, an AI research platform company, will generate synthetic datasets, artificial but statistically realistic stand-ins for real patient data, to fill gaps where real-world information is too scarce or too sensitive to share freely. That company also plans patient-facing tools that connect people to confirmatory testing, care navigation and clinical trials.

"Rare diseases represent one of the greatest unmet needs in medicine, and one of the most important frontiers for AI," said Scott Gorman, RAPID's program manager, in ARPA-H's announcement. He noted the program is designed to build AI-ready datasets and technologies capable of learning from data that is inherently sparse and messy, the kind of information that has traditionally been almost useless to machine learning systems trained on huge, clean datasets.

The financial stakes for patients are enormous. Lukas Lange, CEO of Probably Genetic, framed the ambition bluntly in a separate statement about his company's $10 million RAPID contract. He said the years between 2026 and 2030 could become some of the most significant in genetic disease research if the program succeeds, potentially changing outcomes for as many as 400 million people worldwide living with rare genetic conditions.

RAPID is not operating in isolation. Global Genes and the National Organization for Rare Disorders will help ensure patients themselves shape how the tools get built and used, a detail worth pausing on because so much health AI development happens without much patient input at all. The Lawrence Berkeley National Laboratory, the NIH's All of Us Center for Linkage and Acquisition of Data, and the Monarch Initiative are also lending support. Meanwhile, some of the biggest names in AI, OpenAI, Anthropic, Amazon Web Services and Google, have pledged computing credits and engineering help, though none of them hold a formal contract under the program.

Beyond the core teams, ARPA-H is opening a competition platform called Rare Challenges, supported by NASA's Center of Excellence for Collaborative Innovation, so outside researchers can test their own AI approaches against RAPID's data and benchmarks. It is a way of widening the pool of people working on the problem, rather than limiting progress to a handful of contracted teams.

The challenges here are not purely technical. Digital pathology adoption, which underpins a lot of AI-based disease detection, has lagged for years. Dr. David Klimstra, cofounder of the computational pathology company Paige, told Healthcare IT News in 2024 that progress has been slowed by the high cost of digitizing medical records and images, logistical hurdles, and plain old resistance from clinicians trained for a century on glass slides and microscopes. Software alone will not fix a system still catching up on basic infrastructure.

Why It Matters

For families navigating a rare disease, the diagnostic delay is not an abstraction. It means years of unnecessary suffering, misdirected treatments, and medical bills that pile up faster than answers arrive. It means watching a child's condition progress without knowing what you are fighting. RAPID will not eliminate that pain overnight, and building trustworthy, privacy-respecting data infrastructure at this scale is genuinely hard work that could easily stretch past its five-year timeline.

But the ambition matters. Gorman put it plainly: "For millions of patients and families, living with a rare disease still means years without an accurate diagnosis and too few paths to effective treatment." If RAPID delivers even a fraction of what its backers hope, faster diagnosis, more precise treatment matching, lower costs from earlier intervention, it could reshape how rare disease care works for a generation of patients who have spent too long searching for answers nobody could give them.