Country / Region
APAC
Tags
Artificial intelligence, Data quality, Mapping
Hospitals often store both SNOMED CT and ICD-10 codes for the same clinical term, yet mismatches slip through routine checks and distort analytics, billing, and quality indicators. We evaluated a three-step quality pipeline that begins with rule-based tests driven by the official SNOMED‚ÜíICD-10 map, adds a retrieval-augmented layer that shows GPT-4 only the valid ICD-10 options for each concept, and finishes with an independent GPT-4 review.
Description
This single-centre study audited every dual-coded item in a tertiary-hospital terminology table containing the free-text term, its SNOMED CT identifier, and its ICD-10 code. We quantified three classes of mapping error, compared rule-based detection with two large-language-model strategies, and measured the impact on coder workload.
Scope
SNOMED CT is far more granular than ICD-10, expressing clinical ideas at a level of detail that ICD-10 collapses into broader buckets. That extra specificity lets us judge whether an assigned ICD-10 code is merely approximate, too broad, or flat-out wrong. In addition, SNOMED CT supplies an officially maintained map to ICD-10 complete with interpretable rule logic. Together, these features allow us to pair rapid deterministic checks with tightly bounded LLM prompts, minimising hallucination and maximising reproducibility.
How SNOMED CT will be used
1. primary identifier for each concept in the local table.
2. Source of the complex-map reference set that links every concept to one or more ICD-10 targets.
3. Retrieval key in the RAG layer
Why SNOMED CT will be used
Contact
