Mapping lab/observation names to LOINC

Ground free-text lab and observation names that OpenMed surfaces to LOINC (Logical Observation Identifiers Names and Codes), the universal standard for identifying what was measured. A LOINC code is a fully specified observation — not just an analyte but the full six-axis model: Component, Property, Time, System (specimen), Scale, Method.

LOINC is free to use. It is published by the Regenstrief Institute under the LOINC license: you accept terms-of-use (and register to download the table), but there is no fee and no per-use restriction. The standard, public path for license-clean mapping is a FHIR terminology server exposing LOINC via $lookup / $validate-code, or Regenstrief's hosted fhir.loinc.org.

When to use

• A note or report contains lab/observation names ("serum potassium", "hemoglobin A1c", "blood pressure") and you need a stable LOINC code each.
• You must disambiguate by specimen/system ("glucose in serum" vs "glucose in urine") or method ("HbA1c by HPLC").
• You need UCUM units to pair with the result value, or a US Core Observation (Laboratory Result) coded with LOINC.
• You are mapping a panel (e.g. CBC, BMP) vs its individual analytes.

For diagnoses/procedures use coding-icd10; for drugs use normalizing-rxnorm; LOINC is for observations and measurements.

Quick start (real LOINC / FHIR terminology calls)

Regenstrief hosts a public FHIR terminology endpoint at https://fhir.loinc.org (HTTP Basic auth with your free LOINC account). Many sites instead point at their own server (HAPI, Ontoserver, Snowstorm-with-LOINC). The operations are the same.

import requests
from requests.auth import HTTPBasicAuth

FHIR = "https://fhir.loinc.org"
AUTH = HTTPBasicAuth("YOUR_LOINC_USER", "YOUR_LOINC_PASSWORD")  # free account
LOINC_SYSTEM = "http://loinc.org"

def lookup(code: str) -> dict:
    """$lookup: return the fully specified name + axes for a LOINC code."""
    r = requests.get(
        f"{FHIR}/CodeSystem/$lookup",
        params={"system": LOINC_SYSTEM, "code": code},
        auth=AUTH, headers={"Accept": "application/fhir+json"}, timeout=15,
    )
    r.raise_for_status()
    return r.json()

def validate(code: str, display: str) -> bool:
    r = requests.get(
        f"{FHIR}/CodeSystem/$validate-code",
        params={"url": LOINC_SYSTEM, "code": code, "display": display},
        auth=AUTH, headers={"Accept": "application/fhir+json"}, timeout=15,
    )
    r.raise_for_status()
    params = {p["name"]: p.get("valueBoolean") for p in r.json().get("parameter", [])}
    return bool(params.get("result"))

print(lookup("2823-3"))     # Potassium [Moles/volume] in Serum or Plasma

Search candidate LOINC codes from a text name with the Regenstrief search API (https://loinc.org/search/) or a ValueSet/$expand filter on your server:

def expand_filter(text: str, count: int = 10) -> list[dict]:
    """Text-filter the LOINC code system to candidate concepts."""
    r = requests.get(
        f"{FHIR}/ValueSet/$expand",
        params={"url": "http://loinc.org/vs", "filter": text, "count": count},
        auth=AUTH, headers={"Accept": "application/fhir+json"}, timeout=20,
    )
    r.raise_for_status()
    return r.json().get("expansion", {}).get("contains", [])

Workflow

1. Extract observation/analyte mentions with OpenMed.
2. Assemble the axes you have from surrounding text: component (what), specimen/system (serum, urine, blood), method (HPLC, immunoassay), and scale (quantitative vs ordinal). More axes → a more specific, correct LOINC.
3. Search candidates via $expand?filter= (or Regenstrief search).
4. Disambiguate by matching specimen and property. "Glucose" alone is ambiguous; "glucose, serum, mass/volume" resolves to one code.
5. Validate the chosen code with $validate-code, then $lookup to pull the long common name and the canonical UCUM example unit.
6. Emit {system: "http://loinc.org", code, display} plus the UCUM unit for the result value, into a US Core Observation.

Hand-off from OpenMed

openmed.analyze_text(..., output_format="dict") returns entities, each a dict with text, label, confidence, start, end. Lab analytes often surface under Chemical/Disease models; run the relevant model and feed the spans in:

import openmed

note = "Labs: serum potassium 5.1 mmol/L, hemoglobin A1c 7.8 %."
result = openmed.analyze_text(
    note,
    model_name="chemical_detection_pubmed",   # Chemical category (analytes)
    output_format="dict",
)

for ent in result["entities"]:
    name = ent["text"]                         # e.g. "potassium"
    candidates = expand_filter(name, count=5)  # LOINC candidates
    # carry OpenMed offsets so the code is traceable to the source span
    print(name, ent["start"], ent["end"], "->",
          [(c["code"], c["display"]) for c in candidates[:3]])

Pair the matched LOINC with the value and unit you parse from the same line — LOINC names the test, UCUM names the unit, the value stays in the Observation. Keep only offsets and codes in your mapping table; never persist raw report text.

Edge cases & gotchas

• Specimen ambiguity is the #1 error. Always resolve System/specimen before choosing a code. Defaulting to "Serum or Plasma" when the note says urine produces a wrong but plausible LOINC.
• Panel vs analyte. "CBC" is an order/panel LOINC; the individual results (WBC, Hgb, Plt) are separate analyte LOINCs. Map at the granularity your data is recorded at.
• Method matters for some assays (e.g. HbA1c, troponin generations). If the method is documented, pick the method-specific code; otherwise use the method-less "any method" code rather than guessing.
• UCUM, not free text, for units. Convert "mg/dL" to the UCUM string mg/dL; reject units LOINC's example unit cannot reconcile with.
• Licensing (free, with terms). LOINC is free but Regenstrief-licensed: accept the LOINC terms-of-use and register for a (free) account to call fhir.loinc.org or download the table. Do not obtain LOINC by bundling UMLS or SNOMED — those carry separate restricted licenses and must stay user-supplied and out-of-process (see mapping-to-snomed, linking-umls-concepts).
• Local-first. OpenMed NER runs on-device; only the de-identified analyte string should reach the terminology server. No PHI over the wire.

Standards & references

• LOINC home & license: https://loinc.org/ and https://loinc.org/license/
• LOINC FHIR terminology service: https://loinc.org/fhir/
• FHIR $lookup / $validate-code: https://hl7.org/fhir/codesystem-operation-lookup.html
• UCUM units of measure: https://ucum.org/
• US Core Laboratory Result Observation: https://hl7.org/fhir/us/core/StructureDefinition-us-core-observation-lab.html