A Dutch regional healthcare network serving 1.2M patients across 14 facilities needed an AI clinical decision support layer that could improve diagnostic consistency, reduce documentation burden, and integrate with existing EHR systems — while meeting stringent GDPR, NEN 7510, and EU AI Act requirements.
Zorgnetwerk Noord is a regional healthcare network in the Netherlands operating 14 hospitals, specialist clinics, and community health centres serving 1.2 million patients. Their challenge was clinical consistency: with 400+ clinicians across 14 facilities, diagnostic pathways, documentation quality, and treatment protocol adherence varied significantly between sites. Evidence-based guideline adoption was self-reported at 71% but independently audited at closer to 52%.The network's Medical Director had a mandate from the board: improve clinical consistency and reduce the 2.8 hours per clinician per day spent on documentation — without replacing their existing Epic EHR system.
Building AI for clinical decision-making in the Netherlands involves navigating an unusually complex regulatory environment: GDPR applies with heightened sensitivity for health data (Article 9 special categories), the NEN 7510 medical data security standard mandates specific access controls, the Dutch NZa healthcare authority has reporting requirements, and the incoming EU AI Act classifies clinical decision support as high-risk AI requiring conformity assessment. Any system that could not navigate all of these simultaneously was not deployable.
We designed a clinical augmentation system — explicitly not a diagnostic replacement — that integrates as a layer within Epic rather than alongside it.
Deep Epic FHIR R4 integration — the AI layer surfaces within the clinician's existing Epic workflow, requiring zero context switching. Clinical insights appear as embedded advisory panels within patient charts.
Evidence-based diagnostic pathway recommendations based on presenting symptoms, patient history, and lab values — citing Dutch NHG clinical guidelines with confidence intervals and supporting evidence links.
Structured note generation from clinician dictation — transforming voice notes into SOAP-formatted EHR entries, reducing documentation time by 64% in pilot.
All AI inference runs on-premise within the healthcare network's own infrastructure — no patient data leaves the network boundary. Models are locally hosted and auditable.
Full conformity assessment documentation, human oversight controls ensuring clinician always makes final decision, and complete algorithmic auditability log for every clinical recommendation.
Network-level analytics for the Medical Director: guideline adherence rates by facility, department, and individual — enabling targeted quality improvement interventions.
What Zyrova understood — and many AI vendors do not — is that clinical AI in the Netherlands must be built for a Dutch regulatory environment, not adapted from an American one. They built compliance in from the architecture, not as a layer on top. That made the difference between a deployable system and a proof of concept.
Dr. Ingrid van der Berg Medical Director · Zorgnetwerk Noord (Anonymized)