Discussion

Ideally, mapping should be as deterministic as possible. That is to say, the same input will always produce the same output. Probabilistic methods, that integrate randomness and uncertainty into decision making (including many LLM-based approaches), can still be useful in handling variation or missing data but should be second-tier and clearly flagged with uncertainty scores.

At present, LLMs are not the primary tool for the mapping task itself (due to the need for deterministic, auditable outputs). However, they may become more relevant when extracting entities from unstructured text in future expansions of the program.

An ideal overall solution would have a clear delineation between any tasks in the mapping pipeline which add uncertainty. This could look like a first stage of parsing the raw data into separate entities and then categorising these entities into 'domains' with clarity on the uncertainty of the domain allocation (and any overlap of domains). A second stage would then use the entities to map to a target PaLM code using a variety of possible techniques that consider the domain information, referential mapping and syntactic structures related to the entities. Each technique applied would have its uncertainty labelled and combined with the first stage to give clarity on the quality of mapping finally achieved.

Relation to best practice and options papers

The best practice paper outlines three main mapping scenarios:

1. Single reportables.
2. Supplemental data elements in the Pathology FHIR specification.
3. Multiple SNOMED PaLM reportables for complex reports.

Most content here applies to scenario 1, covering:

• lexical vs. semantic matching
• referential tables
• entity identification
• collaborative mapping

It also supports scenario 3 (complex reports) by addressing semantic and multi-entity mapping. Future considerations around unstructured data could appear in a 'Future Functionality' section of either the best practice or options paper, ensuring any proposed solutions are compatible with more advanced text extraction approaches that emerge in the future.