Considerations for AI-Assisted Maps
Last updated
The consistent view is that AI should be treated as an additional "mapper" rather than a replacement for QA processes. The same quality principles still apply, but AI introduces specific new considerations.
AI assistance is useful for speeding up the initial draft of maps.
A human-in-the-loop remains essential when quality-assuring AI-generated maps.
In SNOMED International projects like SNOMED-ICD-10 and MedDRA, QA remains fundamentally manual, and AI assistance has not yet been incorporated into these workflows.
Experience with AI-generated maps has shown mixed quality:
Issues observed include hallucinated concepts or SCTIDs that do not match their descriptions.
Where the code was plausible, it was not always as precise as it could have been.
Better prompting may improve results, but verification remains critical.
When AI is involved, the QA framework should explicitly include:
Confidence scoring and risk-based review, with stronger review requirements for low-confidence or high-impact maps.
Ensuring reproducibility, since AI-generated outputs can vary over time and across models.
Clearly distinguishing between AI-suggested maps and human-approved maps throughout the dataset.
Last updated