Rare Disease Data Already Exists. It Just Can't Move.
Key facts
- 300 million people worldwide live with a rare disease.
- 95% of those diseases still have no approved treatment.
- Nearly two-thirds of FDA Breakthrough Therapy designations have gone to orphan drugs.
- The Cancer AI Alliance (CAIA) has built a federated learning network across four cancer centers and data from more than one million patients.
- The PANGEA consortium used federated analytics to study roughly 74,000 pediatric IBD patients across international sites—without ever centralizing their records—and overturned a clinical assumption that had stood for a decade.
- Federated analytics is not a narrow toolkit. Counts, means and percentiles, chi-square and t-tests, correlation coefficients, odds and risk ratios, prevalence and incidence, Kaplan-Meier survival curves and Cox proportional hazards models all run across sites without records leaving any of them.
Three hundred million people worldwide live with a rare disease. Ninety-five percent of those diseases still have no approved treatment. When I sit with that gap, the frustrating part isn't that we lack the data to close it. We don't. The data that could lead to new discoveries already exists within hospitals, registries, and research centers on every continent. One of the challenges is how to draw upon existing data to elicit a meaningful impact.
That barrier has big implications for all of us.
Rare disease research has a habit of quietly driving medicine forward for everyone. The mRNA platforms behind COVID-19 vaccines were first tested in rare conditions. Gene therapy, precision oncology, and the cholesterol treatments millions of people now rely on all trace back to scientists working on problems that affected only a handful of patients. Nearly two-thirds of FDA Breakthrough Therapy designations have gone to orphan drugs. Rare disease is where medicine does its hardest work, and what gets solved there tends to reshape everything downstream.
Why can't rare disease data be centralized?
The same should be true of data infrastructure. Solve it where datasets are the most fragmented, the most legally constrained, and the most geographically scattered, and you build something that benefits all of medicine. Yet rare disease has often been underserved by successive waves of data innovation — from relational databases and FHIR, to cloud data lakes and today's AI platforms. Many of these approaches work best when data can be pooled or accessed centrally. For rare disease, that assumption falls apart. Patient data is regulated, sensitive, and legally tied to the institution that holds it. It doesn't move — and it shouldn't have to move for progress to be possible.
How federated computing is perfect for rare disease
Working with Rhino, a federated intelligence network focused on rare disease comes together in a defined sequence. A sponsor, whether a pharmaceutical company, a foundation, or an academic consortium, sets the scope: the indications in view, the minimum dataset, the endpoints that matter, and which institutions take part. Those decisions become a common data model that each site maps its own records to, with the mapping validated by professionals on-site who understand that site's data intimately. A lightweight software client is deployed inside each site's existing environment, behind its own firewall, and connected to its data sources.
Each site's data can then be registered as a versioned dataset that never leaves the institution, and a shared catalog lets participants see each other's schemas, provenance, and locally computed summary statistics, enough to judge cohort size and data quality across the entire network before a single analysis runs. Critically, however, all of the research and downstream modeling using the network’s data can take advantage of configurable differential privacy. Noise is introduced in a way that protects against reconstructing an individual patient from the shape of a curve, which is a genuine risk when a cohort is small enough that one step function belongs to one person. Rare disease is precisely the setting where that risk is not theoretical, and being able to turn the differential privacy dial up for a fifteen-patient cohort and down for one with fifteen-thousand-patients is what makes the same technological infrastructure usable across both.
Here is what it looks like for practitioners. A researcher specifies the statistic they want, the variable it applies to, how results should be grouped, and which subset of patients qualifies. That specification is automatically dispatched to every participating site at once. Each site runs it against its own records, inside its own environment, and returns only the intermediate quantities the statistic requires—sums, counts, contingency tables, gradient terms. Those partial results are combined centrally into a single answer. The researcher sees one aggregate result, or a set of per-site results if the governance agreement permits that level of visibility. Row-level data never appears in either case.
The cohort definition travels with the analysis, which matters more in rare disease than almost anywhere else. Rare disease cohorts are defined by narrow, compound criteria: an age band, a genotype, a treatment history, a diagnostic code that only a handful of clinicians use correctly. Because filtering and grouping are part of the analysis specification rather than something each site implements locally, every institution applies exactly the same inclusion logic. That consistency is the difference between a real federated cohort and five sites answering five slightly different questions.
The range of what runs this way is broader than most people expect. Descriptive work—counts, means, standard deviations, medians and percentiles—is the easy part. Beyond those outputs sit the statistical tests that carry an actual clinical argument: chi-square for categorical association; t-tests and one-way ANOVA for group comparisons; Pearson and Spearman correlations; intraclass correlation for inter-rater reliability across sites reading the same kind of image or scoring the same scale. Epidemiology is covered directly, with two-by-two tables, odds ratios, risk ratios and risk differences, and prevalence and incidence computed over a defined time window. And survival analysis—the workhorse of rare disease natural history—runs as Kaplan-Meier curves and as Cox proportional hazards models with covariates.
The Cox case is the one I point to when someone asks whether this is truly federated or merely a polite form of meta-analysis. A federated Cox model is fit iteratively: the algorithm takes a step, each site computes its contribution using its own patients, those contributions are combined, the estimate is updated, and the cycle repeats until convergence. The result is a single set of coefficients and standard errors fit across the entire distributed population—not a weighted average of five separately fitted models. That distinction is critical in rare disease, where no individual center may have enough events to produce a stable model on its own. The same approach extends beyond survival analysis to logistic, linear, Poisson, quantile, and generalized linear models with feature selection—making a broad statistical toolkit available across the collective population, while patient-level data remains securely within each institution.
Most rare disease networks start retrospectively, reconstructing how a disease has progressed across records the sites already hold. Then those same workflows can run against newly accruing data: new encounters, new enrollees, new follow-up assessments. That continuous feed turns a one-time study into a living registry, where cohorts and endpoints update as patients are seen rather than requiring a new study each time a question worth exploring arises. Adding another site, another country, or another indication compounds the unified understanding of the disease while reusing connections, the shared data model, and the established governance framework that already exist.
For a field defined by small, scattered patient populations, this complete set of capabilities is the difference between a promising idea and a fully functional network. A rare disease community that has always been too dispersed to reach statistical significance on its own can now utilize evidence across borders in real time, and every new institution strengthens what the whole network can learn. Perhaps most importantly, geography no longer limits the ability of a rare disease community to build collective intelligence.
Real-world stories: CAIA, FAITE, and PANGEA
Over the past few years, many federated networks have proven the model in health research and provided templates for success and scale that rare disease communities can copy.
The Cancer AI Alliance (CAIA) is applying this logic in oncology. Four leading cancer centers have built a federated network spanning data from more than one million electronic health records. Initial pilot projects are focused on treatment-response prediction, biomarker discovery, and rare cancer trends. The parallel to rare disease is direct: once cancer is divided into molecular subtypes and treatment contexts, relevant patient populations can become too small and dispersed for any one institution to study adequately. Research of this kind leans on exactly the survival machinery described above—time-to-event curves stratified by subtype, hazard models with covariates for stage and prior therapy—run across four institutions that will never share clinical data directly. CAIA’s model provides the blueprint for how rare disease institutions can build collective intelligence without centralizing their patient data.
The Federated AI for Therapeutic Engineering (FAITE) Consortium shows a similar dynamic in biopharma. Multiple pharmaceutical companies are collaborating to train models for antibody developability, an area where data is generally less abundant, less standardized, and less well characterized than for small-molecule prediction. For any one company, assembling enough diverse, high-quality data would require prohibitive time and investment. A collective approach through Rhino makes the problem more tractable: each federated participant contributes to a larger body of intelligence without pooling its proprietary data, enabling progress that may not be feasible for any organization acting alone. FAITE sits on the model-training end of the spectrum rather than the analytics end, but the underlying arrangement is the same one: the work moves, the data stays.
The PANGEA Consortium shows what this looks like in practice today. It used federated analytics to empower researchers to study roughly 74,000 pediatric IBD patients across international sites without ever centralizing their records. The access to broader data overturned a clinical assumption that had stood for a decade that high gut microbiome alpha diversity is universally beneficial, identifying it instead as a potential risk factor. Such discoveries are what become possible when fragmented data can finally contribute to a shared body of evidence.
What this means for the next decade of rare disease research
Every prior wave of health data innovation left rare disease patients waiting longer and benefiting less. As AI reshapes how we diagnose and treat disease, we have an opportunity — and an obligation — not to repeat that pattern. The infrastructure now exists to include rare disease communities from the start and at global scale, without requiring institutions or patients to surrender privacy or control.
Federated computing promises to enable rare disease researchers to find patients for studies at a scale not possible before, to understand disease progression and patient journeys with greater clarity and ease, to predict treatment response with more depth and accuracy, and to identify existing and novel drugs with less effort and expense.
None of that requires a research team to become distributed systems engineers. The analyses above are configured and run from an environment (e.g., Python) that a biostatistician already knows how to work in, and thanks to our secure Model Context Protocol (MCP) server, even operate with natural language. The federated part—dispatch, execution, aggregation, privacy controls, audit—is handled by the platform underneath. Infrastructure that only a specialist can operate does not scale to a field with seven thousand diseases and no time. Rhino abstracts away all of the most challenging technical work so that the scientists and drug developers can focus their attention on winning the battle against rare diseases.
For 300 million people who have waited long enough, that is progress we cannot afford to postpone.
FAQ
What is federated computing? Federated computing is an approach where computation — analysis, model training, AI inference, and agents — runs locally inside each institution's own environment, and only the results are shared. Patient data itself never has to leave the institution that holds it.
Why can't rare disease patient data simply be pooled into one database? Rare disease patient data is regulated, sensitive, and legally tied to the institution that collected it. Centralizing it would require moving or exposing protected records, which many institutions and jurisdictions cannot or will not permit.
What kinds of analysis can actually run federated? More than most people assume. Descriptive statistics including counts, means, standard deviations and percentiles; statistical tests including chi-square, t-tests and ANOVA; correlation measures including Pearson, Spearman and intraclass correlation; epidemiological measures including two-by-two tables, odds ratios, risk ratios, prevalence and incidence; and survival analysis including Kaplan-Meier curves and Cox proportional hazards models. Broader regression families and custom machine learning run as containerized federated training jobs on the same infrastructure.
Is a federated survival model just a meta-analysis of site-level results? No. A federated Cox model is fit iteratively across all participating sites — each site computes its contribution to each step against its own patients, contributions are combined, and the estimate updates until it converges. The result is a single model fit across the full population, with coefficients and standard errors to match, rather than an average of separately-fit site models.
Does privacy protection compromise the statistics? There is a trade-off, and it is configurable rather than fixed. Survival analyses support differential privacy at off, low, medium or high settings, with medium as the default. Small cohorts — where an individual patient could otherwise be inferred from the shape of a curve — warrant a stronger setting; larger ones do not. The choice sits with the study team and the participating institutions.
Can data from different sources be linked without moving it? Yes. Federated joins allow one dataset to be filtered or enriched by another across institutional boundaries, which is how registry data gets connected to claims, imaging or genomic data without any of it being copied to a shared location.
What is the Cancer AI Alliance (CAIA)? CAIA is a federated learning network built by four leading NCI-designated cancer centers, spanning data from more than one million de-identified electronic health records. CAIA's pilot projects focus on treatment-response prediction, biomarker discovery, and rare cancer trends — without centralizing patient records.
What is FAITE? FAITE (Federated AI for Therapeutic Engineering) is a consortium of pharmaceutical companies that collaborate to train models for antibody developability, contributing to a shared body of intelligence without pooling proprietary data.
What did the PANGEA study demonstrate? PANGEA used federated analytics to study roughly 74,000 pediatric IBD patients across international sites without centralizing their records, and the analysis overturned a clinical assumption that had stood for a decade — showing federated computing can produce clinically meaningful results at scale.
