Terminology Service — AI Integration
Status: populated Owner: TBD Last updated: 2026-04-18 Companion: Service Template · 03 platform-services · 02 DDD
1. AI Integration Status
Minimal — no direct AI calls.
The terminology-service has no direct integration with ai-gateway-service or any other AI/ML inference endpoint. Its query surface (code lookup, validation, expansion, drug interaction checking) is fully deterministic: it operates against a structured relational database of licensed terminology content and returns rule-based results.
2. Rationale for No AI Integration
| Reason | Explanation |
|---|---|
| Determinism required | Clinical coding decisions (SNOMED CT lookup, LOINC validation, drug interaction severity) must be deterministic, auditable, and reproducible. AI-generated responses are probabilistic and cannot be used as authoritative coded-value sources. |
| Licensed data boundary | Terminology content (LOINC, SNOMED, RxNorm, ICD-10) is structured, licensed, and fully machine-readable. There is no natural language ambiguity to resolve. |
| Regulatory compliance | Coded clinical values used in FHIR resources and HL7 messages must conform to published coding system standards. An AI inference layer between query and coded value would introduce non-compliant variability. |
| Performance profile | Concept lookups have a p95 SLO of ≤ 200 ms. AI inference would add unacceptable latency. |
3. Future Scope (Not Committed)
| Potential future capability | Prerequisite before enabling |
|---|---|
| Natural language → code suggestion | A secondary AI-assisted search hint (not a substitute for the canonical result) could help clinicians during free-text entry. This would use ai-gateway-service as an optional suggestion layer with explicit HITL confirmation required before a code is accepted. Must not replace the authoritative $lookup or $validate-code path. |
| Automated concept mapping review | AI-assisted review of concept map quality (e.g., flagging low-confidence cross-system translations) for use by terminology stewards. Result would be advisory only; no auto-update of concept maps. |
| Onboarding concept suggestion | During tenant custom concept creation, an AI could suggest matching global concepts to avoid unnecessary duplication. Tenant admin must confirm before creation. |
All future AI features require platform AI governance review, ai-gateway-service Tier classification, and MoPH compliance sign-off before activation.