The semantic trust flywheel
The flywheel has five layers, and the sequence is not optional. Each layer is a prerequisite for the one above it.
Layer 1: Capture every signal from every source
It requires comprehensive ingestion—every clinical and administrative signal from every relevant source. Provider EHR systems, pharmacies, payers, labs, connected devices, and remote monitoring streams must be captured. The semantic trust problem begins the moment these signals arrive, each carrying its own local encoding conventions and patient identifier scheme.
Layer 2: Semantic trust gateway—where meaning is established
The heart of the architecture—it sits between raw signal ingestion and every downstream consumer. It does four things, in sequence, at the point of ingestion.
- Terminology normalization: Maps every clinical concept to its canonical representation in the appropriate standard ontology. SNOMED CT or ICD-10/11 for diagnoses, LOINC for observations, RxNorm for medications with dose-form and strength disambiguation. This normalization runs against live terminology repositories, not static snapshots. When SNOMED CT updates, the gateway updates with it.
- Ontology reasoning: Enriches the normalized record. A patient coded with a SNOMED CT concept for Type 2 Diabetes Mellitus is automatically inferred to have membership in the broader diabetes mellitus class, enabling population queries and care gap identification that would otherwise require bespoke coding for every query.
- Reconciliation: Produces a single golden record where the same clinical concept arrives from multiple source systems. The gateway uses probabilistic and deterministic master patient index matching across all sources to resolve duplicate identities and consolidates medication histories across pharmacy, EHR, and claims sources. There is one patient. There is one medication history. That is what downstream systems see.
- Provenance tracking: Annotates every normalized data element with immutable lineage metadata. Which source system it came from, when it arrived, which terminology version was used to map it, the confidence score on that mapping, and any reconciliation that was applied. This chain is the foundation of AI explainability. When a predictive model fires a care gap alert, the clinical team can trace every contributing data element back to its source transaction.
Layer 3: The gold layer (the only substrate AI is permitted to train on)
It is the curated, versioned, continuously validated clinical data store that sits downstream of the gateway. It is the single authoritative representation of every patient, medication, encounter, observation, and claims event in the enterprise—normalized, reconciled, and provenance-tagged. Automated data quality checks run on every batch. No element enters without passing the four semantic trust conditions. This is the only data that any model in the enterprise trains on or infers from. The gold layer is meant to serve as a guarantee. That’s why what AI reasons over is worth reasoning over.
Layer 4: A trustworthy foundation. Accurate AI
With a semantically trusted gold layer beneath it, AI delivers what it has been promising. Pattern detection across normalized cohorts identifies comorbidity clusters and readmission precursors without terminological noise. NLP models extracting concepts from clinical notes are grounded against the same standard terminologies that govern structured data, so structured and unstructured signals can be analyzed together. Predictive models for deterioration, medication non-adherence, and care gap likelihood train on data whose integrity is assured. And a semantic knowledge graph—connecting patients, conditions, medications, providers, encounters, observations, pharmacy fills, and claims events through typed relationships—enables the kind of multi-hop reasoning that is simply impossible in a tabular data architecture.
That last capability deserves a concrete illustration. Consider the query: identify patients with a confirmed type 2 diabetes diagnosis who have been prescribed metformin but have had no pharmacy fill in 90 days, and whose most recent HbA1c exceeds 8.0%. In a tabular model, this requires joining multiple tables across systems using identifiers that may be inconsistent. In a semantic knowledge graph built on trusted data, it is a four-hop traversal filtered by date and value at each node. The answer is available in real time. The clinical team can act on it today.
Layer 5: Agents that act and a flywheel that never stops
The apex of the architecture is agentic care: autonomous workflows that execute on behalf of clinical teams. Care coordination agents orchestrate transitions across the care continuum. Pharmacy adherence agents monitor dispensing records and alert on fill gaps correlated with deterioration signals. Prior authorization agents assemble clinical justification packages from the gold layer and submit them to payers without manual intervention.
Each of these agent actions generates new clinical signals: an auth approval, a care gap closure, a pharmacist intervention record. Those signals flow back into the semantic trust gateway, are normalized and provenance-tagged, and enter the gold layer enriched. The flywheel does not stop. Every cycle improves the data foundation for the next one.
What else is covered in the PDF
Most large health systems and payers are at Stage 2 of a five-stage maturity model. The key step is Stage 3—the semantic trust gateway—which is the prerequisite for every AI capability above it. The knowledge graph, built above the gold layer, enables the AI interpretation layer by connecting a network of typed, provenance-tagged entities with meaningful relationships.
The graph enables AI predictions to be explainable: every node is provenance-tagged, so the clinical team can review the underlying evidence. As regulatory frameworks crystallize, traceability becomes a compliance requirement. Organizations with the most trusted data, not just models, will lead. The semantic trust flywheel—capture, validate, deploy, activate, and refine—drives sustainable AI advantage. You’ll find more detail in the attached PDF.
