Noise to LOINC

This module converts noisy laboratory names extracted from medical faxes into validated LOINC codes for downstream EHR integration.

The production pipeline requires the local UMLS/ScispaCy asset and SapBERT model. It starts with the local LOINC release in data/, extracts source-neutral six-axis facts, retrieves universal candidates from the pinned release, applies deterministic mappings and safety validation, uses UMLS semantic evidence, reranks with SapBERT, and abstains when evidence is insufficient.

Source-laboratory mappings are optional governed evidence, not the only route to ordinary results. Start with Universal multi-lab mapping to understand how Creatinine, Urine, 24-hour urine, method, and specimen facts work without a laboratory-specific alias.

Start with the Pipeline tutorial, then use the Deployment runbook to plug the mapper into a fax service and the Maintenance runbook to refresh releases without mixing incompatible artifacts.

For the operational review loop, see Clinician-governed learning. It defines the doctor-friendly form/CSV contract, immutable registry snapshots, and the boundary between this stateless package and the main fax application.

Local development

python -m unittest discover -s tests -v

Production mapping requires all pipeline stages and therefore also needs --umls-path, an installed ScispaCy model, and locally available SapBERT weights. The mapper returns the selected code, ranked alternatives, evidence, provenance, and review reasons.