In Short: Compliance Is the Architecture
Healthcare needs analytics more than almost any sector - capacity, outcomes, waiting lists, workforce - and operates under constraints that make ordinary BI practice negligent: identifiable patient data, layered consent, regulatory scrutiny, and zero tolerance for leaks. The consequence for platform design is simple to state: in health, governance is not a layer on the architecture; it is the architecture. Everything else gets built inside it.
Fabric fits that inversion unusually well, because its security and governance surfaces were built platform-wide rather than bolted per tool. Here is the shape that works.
The Zoned Estate
Health data platforms succeed as zones with explicit seams, and the medallion pattern maps to them naturally.
- Identifiable zones: source-shaped clinical and patient data, minimal access, heavy audit - the bronze and early silver layers, locked to the few who genuinely need them
- De-identified analytical zones: the working silver and gold layers where most analytics happens, built through documented de-identification steps whose lineage proves what was removed and when
- Aggregate and published zones: dashboards, board packs, public reporting - freely consumable because the sensitivity was engineered out upstream
Fabric enforces the seams: workspace boundaries per zone, OneLake access roles, row and column security where cohorts and columns differ, sensitivity labels and Purview policy over everything, audit throughout. The design work is deciding the zones; the platform's job is making them hold.
Where the Value Lands First
Counterintuitively, not clinical dashboards. The pattern we see succeed starts operational: flow and capacity (beds, theatres, clinics, discharge), waiting lists and access, workforce and finance - high-value questions on comparatively lower-sensitivity data, where the platform earns trust and the data quality disciplines get proven. Clinical and outcomes analytics follow on those rails, inheriting governance already evidenced.
Ingestion follows the sector's reality: mirroring for operational systems, pipelines for the long tail, interoperability formats where they exist - with the honest note that conforming health data is specialist work, and the pipeline discipline of validation and loud failure matters doubly when the data describes patients.
AI, Earned
The conversational layer is coming to health analytics, and the sequencing rules tighten rather than change: permissions tested under AI access, definitions agreed, oversharing assessed, and Copilot enabled first on the de-identified operational domains where a wrong answer embarrasses rather than harms. Done in that order, managers asking Teams about theatre utilisation is a genuine win; done casually, it is the incident report nobody wants to write. The readiness checklist applies with the word "must" substituted for "should".
The encouraging summary for health data leaders: the strictest-constraint sector is also where Fabric's integrated governance pays best, because the alternative - stitching compliance across five tools - is precisely what kept health analytics a decade behind. One platform, one control plane, zones that hold: that is the whole strategy.
Sources and Further Reading
Frequently asked
The platform provides the control set the sector requires - workspace isolation, row and object security, sensitivity labels, Purview policy, audit - inherited from Microsoft's compliance infrastructure. Suitability then depends on configuration and process: the controls must be designed and evidenced for your regulatory context, which is the real work.
Operational analytics, usually: capacity and flow (beds, theatres, clinics), waiting lists, workforce - high value, comparatively lower sensitivity than clinical records, and the fastest route to trust. Clinical analytics follows on the proven governance.
Layered access by design: identifiable data confined to tightly controlled zones, de-identified and aggregated layers serving most analytics, security enforced at workspace, item, row and column level, with labels and audit riding everything. The medallion pattern maps naturally to de-identification stages.
On governed, appropriately de-identified analytical data, conversational access is realistic and valuable - managers asking about flow, clinicians about service metrics. It arrives after the readiness work: permissions tested under AI access, definitions agreed, oversharing assessed. AI on health data is earned, not enabled.
The lakehouse suits it well: consented, de-identified research zones built from the same governed foundation, with lineage proving provenance and shortcuts sharing without copying. The governance seam between care and research use must be explicit - which is a feature of doing it properly, not a burden.
