From FHIR Repositories to Learning Health Systems: Inside the Modern Health Data Platform
— by Vivax
Healthcare data is scattered across departments, devices, and decades of legacy systems — and none of it is useful to AI until it is unified, standardized,…
Every hospital is, in effect, a data company that never set out to be one. Patient information lives in the EHR, the lab system, imaging archives, monitoring devices, and a long tail of departmental tools — each with its own format, identifiers, and quirks. None of that data is useful to analytics or AI until it has been pulled together, standardized to a common model, de-duplicated, and made trustworthy. The category that has grown up to do this is the "health data platform": a governed layer that ingests messy source data, normalizes it — most commonly to HL7 FHIR R4 — and exposes it as a clean, queryable, AI-ready foundation. Below are four reference points that show how different the approaches can be, from commercial repositories to cloud lakes to the methodology of continuous learning itself.
Smile Digital Health is one of the most established players. Its platform, "Smile Omni," is a FHIR-native clinical data repository and interoperability layer built so that data, quality measurement, and compliance all live on a single shared foundation rather than in disconnected silos that need constant manual reconciliation. The pitch is computable quality and intelligent automation on top of standardized data — the idea that once your records are genuinely normalized and governed, downstream work like quality reporting, care-gap detection, and analytics stops being a series of brittle one-off integrations. It is a good illustration of why "FHIR-native" matters: when the standard is the storage model and not just the export format, far more becomes possible without bespoke plumbing.
The major clouds have entered the same space with managed services. AWS HealthLake is a HIPAA-eligible service that ingests health data from disparate sources, transforms it into the FHIR R4 standard, and stores it in a queryable data lake with built-in medical natural-language processing — turning unstructured notes into structured, searchable, ML-ready data. Microsoft's Azure Health Data Services takes a broader, multi-modal stance: a managed FHIR service for clinical data, a DICOM service for medical imaging, and a MedTech service that converts high-frequency device and IoT data into FHIR — all in one compliant workspace. Both reflect the same conviction as Smile's: standardize first, then build. The difference is largely one of scope and operating model — a focused interoperability platform versus a broad managed cloud estate.
A platform, however, is only the substrate. The discipline that gives it purpose is the Learning Health System (LHS) — and the George Washington University LHS Toolkit is a clear, practical guide to it. The LHS, as the U.S. National Academy of Medicine framed it, is a system in which "science, informatics, incentives, and culture are aligned for continuous improvement and innovation" — where data generated during everyday care is continuously aggregated and analyzed, the resulting knowledge is fed back into practice, and the cycle repeats. The GWU toolkit organizes this into workable components — strategy, complexity, technology, and a learning community — making the point that becoming a learning health system is not a product you buy but a capability you build. A standardized data platform is what makes that loop technically possible; the LHS is what makes it matter.
The Vivax open-source data platform is our answer to the same problem these systems address — but built so the learning loop can close without a single patient record ever leaving the hospital. Like a FHIR-native repository, the Vivax Standardized Data Platform first grades every record for quality and reliability, then cleans, validates, and harmonizes it into unified, structured data forms — and, unlike a proprietary lake, publishes those as open-source data archetypes and a new device-and-software standard that minimizes the layers between hardware, records, and model. Where a cloud lake centralizes data to learn from it, Vivax inverts that: in-place federated learning, differential privacy, and secure aggregation let a shared clinical world model improve across many hospitals while the raw data stays behind each hospital's own walls. That is the Learning Health System loop — data to knowledge to better care — run on an open foundation, with the clinical world model open-source at its core and only the implementation layer commercial on top. The result is the AI-readiness these platforms promise, without the ownership and privacy trade-offs that usually come with it.