Medical AI · Foundations · Lesson 3 of 6

Data standards: FHIR, HL7, ICD, SNOMED CT, LOINC and OMOP

Reviewed by Dr. Ram Kumar G, MBBS, MD (Internal Medicine), MPH (USA), on 2026-10-18 · Last updated 2026-10-18

Beginner
Technical focus
About 30 minutes
Includes exercise

Before this lesson: Healthcare data types: records, images, signals, notes and omics

In short

Healthcare data standards fall into two groups. Exchange standards such as HL7 v2 and FHIR define how data moves between systems; terminologies such as ICD, SNOMED CT and LOINC define what each item means. Research models such as OMOP put data from many sources into one shape. AI projects that ignore these standards spend most of their time reconciling codes and units by hand.

You will learn to

Explain the difference between exchange standards, terminologies and common data models
Recognise the purpose of FHIR, HL7 v2, DICOM, ICD, SNOMED CT, LOINC, UCUM and OMOP
Parse a FHIR Observation in Python and read its codes and units
Map local lab codes and units to a standard so data from two sites can be combined

Why standards matter for AI

Two hospitals can record the same HbA1c result with different local codes, different units and different names. A model trained on one site's codes cannot read the other's. Standards solve this at three levels: how data are packaged and sent, what the codes mean, and how data are arranged for analysis.

Layer	Question it answers	Main standards
Exchange	How is the data packaged and sent between systems?	HL7 v2 messages, HL7 FHIR, DICOM (images)
Terminology	What does each item mean?	ICD (diagnoses), SNOMED CT (clinical terms), LOINC (tests and observations), UCUM (units)
Common data model	How is data arranged for analysis across sources?	OMOP Common Data Model

Exchange: HL7 v2 and FHIR

HL7 version 2 is a messaging standard from the late 1980s that still carries much of the traffic inside hospitals, such as admissions and lab results, as pipe-delimited text messages. FHIR (Fast Healthcare Interoperability Resources) is HL7's modern standard: data are split into 'resources' such as Patient, Observation, Condition and MedicationRequest, exchanged as JSON or XML through web APIs.

As of October 2026, FHIR R4 (2019) is the version most widely implemented and the one required by many national programmes; R5 was published in 2023; and R6 is still being balloted and is not yet a published standard. National 'implementation guides' constrain FHIR for local use: for example US Core in the United States, and the FHIR implementation guide for India's Ayushman Bharat Digital Mission (ABDM), which is based on R4.

python

import json

observation_json = """
{
  "resourceType": "Observation",
  "status": "final",
  "code": {"coding": [{"system": "http://loinc.org", "code": "4548-4",
            "display": "Hemoglobin A1c/Hemoglobin.total in Blood"}]},
  "subject": {"reference": "Patient/example-001"},
  "effectiveDateTime": "2026-03-14",
  "valueQuantity": {"value": 7.8, "unit": "%",
                    "system": "http://unitsofmeasure.org", "code": "%"}
}
"""
obs = json.loads(observation_json)

coding = obs["code"]["coding"][0]
quantity = obs["valueQuantity"]
print(f"{coding['display']} (LOINC {coding['code']})")
print(f"{quantity['value']} {quantity['unit']} on {obs['effectiveDateTime']}")

# Hemoglobin A1c/Hemoglobin.total in Blood (LOINC 4548-4)
# 7.8 % on 2026-03-14

Read the test, value and units from a synthetic FHIR R4 Observation.

The meaning of the test comes from the coding system (LOINC) and code (4548-4), not from the display text, which can vary.
Units are coded with UCUM (unitsofmeasure.org), so software can tell '%' from 'mmol/mol' reliably.
Real Observations often carry several codings (a local code plus LOINC), reference ranges and interpretation flags; robust code checks the system before using a code.

Terminologies: ICD, SNOMED CT and LOINC

Terminologies give each concept a stable code. They serve different purposes and are often used together: SNOMED CT for detailed clinical recording, ICD for statistics and billing, LOINC for naming tests and measurements.

Codes shown are for illustration; always look concepts up in the official browsers for the edition you use.
Terminology	Maintained by	Used for	Type 2 diabetes / HbA1c example
ICD-10 / ICD-11	World Health Organization	Mortality and morbidity statistics, billing, reporting. ICD-11 came into effect on 1 January 2022; many countries still use ICD-10 or national versions	ICD-10 E11; ICD-11 5A11 (type 2 diabetes mellitus)
SNOMED CT	SNOMED International (national release centres in member countries)	Detailed clinical recording in EHRs: findings, disorders, procedures	44054006 (diabetes mellitus type 2)
LOINC	Regenstrief Institute	Identifying lab tests and clinical observations	4548-4 (HbA1c); 2345-7 (glucose in serum or plasma)
UCUM	Regenstrief Institute	Unambiguous units of measure	%, mmol/mol, mg/dL

Licensing

LOINC and UCUM are free to use under their licence terms. SNOMED CT is free in member countries through the national release centre, but needs a licence elsewhere. ICD is published by WHO under its own terms. Check the licence before building a product on any terminology.

Harmonising codes and units across sites

In practice, each lab system uses local codes, and the same test may be reported in different units. Before pooling data, map local codes to LOINC and convert values to one unit. The example below does this for two synthetic hospitals.

python

LOCAL_TO_LOINC = {
    ("hospital_a", "GLU"): "2345-7",      # Glucose [Mass/volume] in Serum or Plasma
    ("hospital_b", "GLUC_P"): "2345-7",
    ("hospital_a", "A1C"): "4548-4",      # Hemoglobin A1c/Hemoglobin.total in Blood
    ("hospital_b", "HBA1C_IFCC"): "4548-4",
}

def harmonise(source, local_code, value, unit):
    loinc = LOCAL_TO_LOINC[(source, local_code)]
    if loinc == "2345-7" and unit == "mmol/L":
        value, unit = value * 18.0, "mg/dL"          # glucose: 1 mmol/L ≈ 18 mg/dL
    if loinc == "4548-4" and unit == "mmol/mol":
        value, unit = value / 10.929 + 2.15, "%"     # IFCC → NGSP master equation
    return loinc, round(value, 1), unit

results = [
    ("hospital_a", "GLU", 126, "mg/dL"),
    ("hospital_b", "GLUC_P", 7.0, "mmol/L"),
    ("hospital_a", "A1C", 7.0, "%"),
    ("hospital_b", "HBA1C_IFCC", 53, "mmol/mol"),
]
for r in results:
    print(r, "->", harmonise(*r))

# ('hospital_a', 'GLU', 126, 'mg/dL') -> ('2345-7', 126, 'mg/dL')
# ('hospital_b', 'GLUC_P', 7.0, 'mmol/L') -> ('2345-7', 126.0, 'mg/dL')
# ('hospital_a', 'A1C', 7.0, '%') -> ('4548-4', 7.0, '%')
# ('hospital_b', 'HBA1C_IFCC', 53, 'mmol/mol') -> ('4548-4', 7.0, '%')

Map local codes to LOINC and convert units so both sites' results line up.

After mapping, the four results collapse into two tests with consistent units: 126 mg/dL glucose and 7.0% HbA1c at both sites.
Without this step, a model would treat 7.0 mmol/L and 126 mg/dL as different values, or 53 as an impossibly high HbA1c.
Real mappings involve thousands of local codes and need clinical and laboratory review; tools from the OHDSI community help, but do not remove the need for checking.

OMOP: one shape for research data

The OMOP Common Data Model, maintained by the OHDSI community, defines standard tables (person, condition occurrence, drug exposure, measurement and others) and standard vocabularies. Once data from different hospitals or countries are converted to OMOP, the same analysis code runs on all of them. As of 2026, version 5.4 is the release OHDSI recommends for new work.

FHIR and OMOP do different jobs: FHIR moves individual records between systems in real time; OMOP organises large volumes of data for analysis. Many projects use both, with HL7's FHIR-to-OMOP guidance to convert between them.

	FHIR	OMOP CDM
Main purpose	Exchanging data between systems	Analysing data across sources
Unit of work	One resource or patient at a time, via APIs	Whole populations in a database
Typical AI use	Feeding a deployed model with live data; agent tools	Training and validating models across many sites

Standards and national programmes

Governments increasingly require standards. In the United States, certified EHRs must provide FHIR R4 APIs. In the European Union, the European Health Data Space Regulation (EU) 2025/327 entered into force on 26 March 2025, with its main provisions applying from March 2027 and secondary-use rules from 2029. India's ABDM uses FHIR R4 profiles and SNOMED CT through the National Resource Centre for EHR Standards. For AI developers, these programmes mean more data will arrive in standard forms, along with rules on who may use it and for what.

Exercise

Read synthetic FHIR R4 patient bundles from Synthea, extract HbA1c and glucose Observations by LOINC code, and harmonise their units into one analysis table.

Data: Synthea synthetic FHIR R4 bundles (Synthea is Apache 2.0; generated records are synthetic)

Load a few Synthea FHIR bundles (the notebook downloads a small sample) and count the resources of each type.
Extract every Observation with LOINC 4548-4 or 2345-7 into a pandas table with patient, date, value and unit.
Add a second, made-up 'site' whose results use mmol/L and mmol/mol, and harmonise both sites with the mapping approach above.
Look up 44054006 in a SNOMED CT browser and E11 / 5A11 in the WHO ICD browsers, and note how the concepts differ.
Sketch which OMOP tables (person, measurement, condition_occurrence) each extracted item would go into.

The runnable notebook and worked solution will be published soon. You can work through the steps with the code on this page in the meantime.

Key takeaways

Exchange standards (HL7 v2, FHIR, DICOM) move data; terminologies (ICD, SNOMED CT, LOINC, UCUM) give it meaning.
FHIR R4 is the widely implemented version as of 2026; R6 is still in ballot.
Map local codes and units to standards before pooling data from different sites.
OMOP lets the same analysis run across many datasets; FHIR serves live, record-level exchange.

Terms used: DICOM, FHIR (Fast Healthcare Interoperability Resources), HL7, ICD (International Classification of Diseases), LOINC, OMOP Common Data Model, SNOMED CT, UCUM

FAQ

What is the difference between HL7 and FHIR?

HL7 is the standards organisation and also the name of its older v2 messaging standard. FHIR is HL7's newer standard, which exchanges data as web resources through APIs.

Should I use ICD or SNOMED CT?

They do different jobs. SNOMED CT is designed for detailed clinical recording; ICD is designed for statistics and billing. Many systems record in SNOMED CT and map to ICD for reporting.

Which FHIR version should I build on?

As of October 2026, FHIR R4 is the version most widely implemented and required by many national programmes, including US certification and India's ABDM. R6 is not yet a published standard.

What is LOINC used for?

LOINC gives a universal code to each lab test and clinical observation, such as 4548-4 for HbA1c, so results from different labs can be recognised as the same test.

Do I need OMOP for a single-hospital project?

Not necessarily, but converting to OMOP makes it far easier to validate a model at other sites later and to use the OHDSI community's open-source tools.

Is SNOMED CT free to use?

It is free in SNOMED International member countries through their national release centre, and requires a licence elsewhere. Check the terms for your country and use case.

Sources

Educational content only. This material teaches methods; it is not clinical advice. Code, models and examples here are not validated medical devices and must not be used to make decisions about real patients without proper clinical validation, governance and any required regulatory approval. Examples use synthetic or openly licensed de-identified data only. Never paste identifiable patient information into public AI tools. Mentions of models, tools or vendors are for identification, not endorsement.