Before an AI reads the chart, it must know what is true, for whom, and when.

The system landscape is a chain of clinical ownership.

Hospital labels vary, and systems often overlap. What matters in discovery is which system is authoritative for each identity, order, result, document, and state transition.

Standards are necessary. A local profile makes them real.

JSON does not establish clinical meaning, and an “HL7 feed” does not identify the fields, events, statuses, correction behavior, or acknowledgements a deployment actually supplies.

Freshness needs a recovery path, not just a notification.

A transport acknowledgement is not a clinical outcome.

For every flow, agree a contract that makes processing outcomes, revisions, and safe retries explicit.

Acknowledgements & idempotency

• Deduplicate v2 by sender/facility + MSH-10, plus a durable business/event key—never solely patient/time/value.
• Persist before acknowledging receipt where the receiver’s contract requires durable acceptance.
• Monitor transport acceptance and application processing separately.
• For FHIR writes, use conditional create/update or a business identifier and version-aware concurrency when supported.
• Make writes safely repeatable: a timeout retry must not create a duplicate order, note, or medication administration.

Forms are clinical objects, not merely fields to flatten.

Rows such as GCS_total = 9 or SBAR_summary = “…” lose the assessment instance, instrument version, hierarchy, author, signature state, encounter, narrative and whether a score was machine-derived or clinician-confirmed.

FHIR explicitly separates capture through Questionnaire / QuestionnaireResponse from standardized observations, and requires preserving questionnaire hierarchy/order. For TPN, the signed order/document and its lifecycle remain authoritative; component facts can support calculations but are not an active order without their status and identifiers.

JSON is syntax. AI needs a semantic layer with evidence.

JSON can preserve fields, but it does not supply a stable concept, unit, encounter, source, validation state, time semantics, or a decision about authority.

Raw source payload (immutable, access controlled)
  → parsed canonical event (source IDs, delivery facts)
  → clinical model (Patient / Encounter / Order / Result / Form / Device)
  → terminology mapping (local code + standard code + mapping version)
  → AI retrieval views and derived features

Every AI-visible fact should point to source system + source object/message ID + source version + original timestamp(s) + transformation/mapping version + validation state.

• Timeline retrieval: use patient/encounter scope, code/status/time filters and a clinical-time cutoff.
• Artifact retrieval: return the full versioned SBAR, assessment or report—not only extracted fields.
• Relationship traversal: Observation → DiagnosticReport → ServiceRequest → Specimen; QuestionnaireResponse → Encounter → author.
• Tool calls: named, read-only functions with scope, filters, pagination and audit records.
• Derived features: store as versioned derivatives; never overwrite source facts.

FHIR Provenance records the entities, agents and processes that created, revised or influenced a resource—useful for assessing authenticity and lifecycle state.

Start read-only, de-identified, and auditable.

• Start with de-identified or synthetic data; use read-only service accounts for production discovery.
• Apply least privilege by system, resource, patient/encounter scope, purpose, environment and time-bound credential.
• Segment integration infrastructure from bedside/device networks; use authenticated encrypted transport, secret rotation, allow-lists and monitored identities.
• Keep PHI out of application logs, prompts, support tickets, vector-store metadata and unbounded analytics exports.
• Encrypt in transit and at rest; define retention, deletion, backup and incident procedures.
• Audit reads, searches, exports, transformations, AI tool calls, writes and privilege changes. AuditEvent aligns with IHE ATNA/DICOM audit concepts.
• Confirm Kuwait law, Ministry/hospital policy, data residency and cross-border processing with the hospital’s legal/privacy owner; do not assume a U.S. HIPAA model.

Reconcile records and their clinical context.

Build a test pack across ICU states: admission, transfer, discharge, duplicate/replay, merge, cancelled order, preliminary/final/corrected result, delayed validation, a complete SBAR, GCS components + total, a TPN order, and imaging result/image link.

Common failures hide in the missing context.

1. Patient matched by name/MRN but not encounter; values land in the wrong stay.
2. ADT arrives late or moves/merges are ignored.
3. Acknowledgement is treated as clinical persistence.
4. Retries create duplicate orders or results.
5. Values arrive but OBR, NTE, specimen, status, effective time, validation or correction is missing.
6. “Latest” selects preliminary, corrected-old, or late-arriving data incorrectly.
7. Unit, method, reference range or local test code changes silently.
8. Device values are mistaken for clinician-validated assessments.
9. Form values are extracted but question/version/author/signature/narrative is lost.
10. Warehouse timestamps are treated as clinical event time.
11. Free text is embedded without document identity, access control, retention or source citation.
12. A production write is enabled before read-only reconciliation proves safety.

The 14-question discovery checklist.

These questions turn a vendor conversation into a concrete pilot contract.

Glossary & a 60-minute learning path.

ADT

Admission, discharge and transfer events.

ACK

Technical acknowledgement of an HL7 message; not inherently proof of clinical processing.

CDC

Change-data capture from database transaction logs.

FHIR

HL7 standard organizing healthcare information into interoperable resources/APIs.

HIS / EHR / EMR

Hospital/enterprise clinical record platforms; usage varies by vendor and region.

Idempotency

Repeating the same request/event has the same outcome as processing it once.

LIS / RIS / PACS / PDMS

Laboratory / radiology / image archive / patient-data-management systems.

OBR / OBX

HL7 v2 report/order context / individual observation-result segment.

Provenance

Evidence of where data came from and how it was created or transformed.

UCUM

Standardized representation of units.

Watermark

Saved high-water mark used to resume polling/CDC safely.