To translate ICD, SNOMED CT and LOINC information accurately, a translator must separate clinical language from the coded systems that represent classifications, clinical concepts and observations. Health records, research datasets, laboratory reports, electronic forms and interoperability specifications can contain readable labels beside codes whose job is to preserve meaning across systems. Translating the label is often necessary; changing the code because the wording changes is usually not.
This guide explains how to translate medical and health-data documents without corrupting ICD codes, SNOMED CT concept identifiers or LOINC codes. It solves one distinct high-intent translation problem: what does each code system represent, which human-readable terms may have official target-language forms, why mapping is different from translation, and how can multilingual teams preserve the same clinical classification, concept or observation after the language changes?
The core rule is translate the description, preserve the governed code, identify the code system and version, and verify the relationship between them. WHO describes ICD as an international classification for recording and reporting health conditions; SNOMED CT concept identifiers uniquely identify clinical ideas while descriptions can vary by language and dialect; LOINC assigns permanent identifiers to laboratory and clinical observations whose fully specified names can also be translated. This article is educational guidance for translators and health-data teams, not a substitute for clinical judgment or professional coding decisions. It connects to Master Art of Translation, the protected Vocabulary Learning Hub, and How English Works without creating a competing medical or translation hub.
A one-minute orientation: three systems, three different jobs
ICD, SNOMED CT and LOINC are often mentioned together because modern health information systems can use all three. They do not solve the same problem. ICD organizes health conditions and related information into a classification used for reporting and statistical purposes. SNOMED CT represents detailed clinical concepts. LOINC identifies observations, measurements, tests and documents.
A translator should therefore begin with the question: what kind of thing is this code representing? A diagnosis category, a clinical concept and a laboratory observation can all have human-readable labels, but they belong to different semantic systems and cannot be exchanged simply because the words look similar.
1. Clinical codes are governed data, not ordinary vocabulary
A code can look like a short word, a number, or a mixed string. Its meaning comes from the terminology or classification that issued it, not from how a translator interprets its visible characters. Treat code fields as protected structured data unless the project explicitly calls for authorized terminology mapping.
The label beside the code is different. It exists so humans can understand the coded concept. That label may have an official translation, a preferred term, a synonym, or a local display form. Translation happens at that human-language layer while the code system maintains the machine-readable identity.
2. Translation and clinical coding are different professional tasks
A translator renders meaning between languages. A clinical coder or terminology specialist determines which governed code appropriately represents documented clinical information under the applicable rules. Those tasks can interact, but one should not silently replace the other.
If a translated phrase seems to point toward a different ICD category or SNOMED concept, do not recode it by intuition. Flag the issue for the responsible clinical-coding or terminology team. The safe translation workflow preserves the source code while accurately rendering the source meaning unless authorized mapping work is part of the assignment.
3. ICD is a classification, not merely a list of translated disease names
WHO describes the International Classification of Diseases as a global standard for recording and reporting health and health-related conditions. ICD supports comparability and interoperability of health data across settings and countries. A category therefore has a place in a classification structure, not just a label.
Translating the title of an ICD category does not create a new category. The code and its official classification context remain the anchor. Use WHO or jurisdiction-approved terminology when available rather than inventing a literal translation that may not match the governed version.
4. ICD version matters
WHO identifies ICD-11 as the latest revision of ICD, but implementation occurs over time and health systems can continue to use earlier versions or national modifications during transition. A translator must therefore record which ICD version the source actually uses instead of assuming every current document uses one universal release.
“ICD code” is incomplete metadata. A robust project should capture classification family, version or release, jurisdictional modification where relevant, and the code itself. The same visible code pattern can mean something different across versions or national systems.
5. ICD-10 and ICD-11 are not term-for-term translations of each other
Moving data from ICD-10 to ICD-11 is a classification transition, not a language translation. Categories, structure and coding logic differ. A multilingual project should never replace an ICD-10 code with an ICD-11 code merely because a target-language title looks similar.
If a project requires classification mapping, use authorized mapping resources and qualified coding review. Keep that transformation separate from the language-translation step so auditors can see whether meaning changed because of language, coding system or both.
6. National ICD modifications can have their own rules
Some jurisdictions maintain clinical modifications or national adaptations of ICD for reimbursement, morbidity coding or other local uses. A code from one national implementation should not be assumed to exist unchanged in another merely because both systems are based on ICD.
Translate the document as belonging to its stated system. If the target organization needs conversion to another national coding system, treat that as a separate governed mapping task with appropriate domain review.
7. ICD labels can have official multilingual forms
WHO’s ICD-11 tools support multiple languages and releases. When the target language is officially available for the relevant release, prefer the official title or description rather than independently translating the English wording.
This protects both terminology and scope. Two everyday synonyms can sound equivalent yet differ from the official classification title. The code may remain stable while the display title follows the authorized linguistic version.
8. Inclusion and exclusion notes are part of classification meaning
ICD entries can contain definitions, inclusion terms, exclusions, notes and other guidance. Translating only the category title can remove the context needed to understand how the classification is intended to be used.
If the project publishes classification guidance, translate the full governed material required for the use case. Do not summarize exclusions into a shorter sentence that changes scope. For coding decisions, consult the official classification and qualified coding professionals.
9. ICD extension codes and postcoordination require structural fidelity
ICD-11 supports a richer digital structure, including combinations and extension information in appropriate contexts. These are not punctuation choices. Translators working with exported coded strings should protect the syntax and identifiers used by the source system.
Translate human-readable display text separately. A visually cleaner string that drops extension information can represent a less specific or different coded meaning.
10. Never derive a diagnosis code from translated prose alone
A phrase such as “chest pain,” “infection,” or “abnormal result” can appear in many clinical contexts. Translation does not provide the clinical evidence required to choose a diagnosis code. Even when a target phrase resembles an ICD title, the translator should not infer or assign a new code unless explicitly working under an authorized coding process.
Preserve the source code, translate the documented phrase, and flag any apparent mismatch. This protects patients, data quality and auditability.
11. SNOMED CT represents clinical concepts
SNOMED International defines a SNOMED CT concept as a clinical idea to which a unique concept identifier has been assigned. The identifier provides stable reference to the meaning even though several human-readable descriptions may be associated with that concept.
This is a powerful translation model. The concept remains the same while different languages and dialects can provide terms that represent it naturally. The translator’s job is therefore concept-based rather than word-for-word substitution.
12. A SNOMED CT concept identifier should not be translated
SNOMED CT concept identifiers are unique numeric identifiers. They are not quantities and should be stored as text or appropriate terminology identifiers rather than reformatted as ordinary numbers.
Do not localize their digits, add thousands separators or convert them to scientific notation. The readable term can change with language; the concept identifier is the stable reference to the clinical idea.
13. SNOMED CT descriptions are associated with concepts
SNOMED CT separates the concept from its descriptions. A description is an association between a human-readable term and a concept. This lets one concept have a fully specified name, preferred terms and synonyms in different languages or dialects.
Do not assume that every synonym should be translated into one target synonym. The correct target display depends on the target-language reference set, dialect and use context.
14. Fully specified name and preferred term serve different purposes
In SNOMED CT, the fully specified name helps disambiguate the concept, while a preferred term is a synonym selected as the preferred way of expressing the concept in a particular language or dialect. A clinical interface may display the preferred term while terminology tools expose the fully specified name.
Translation should respect that role. Do not replace every fully specified name with a casual synonym or make a preferred display term carry technical qualifiers it was not meant to show.
15. SNOMED CT translation should be concept-based
SNOMED International explicitly advises that translation should represent the meaning of the concept rather than mechanically translate source terms. Translators examine hierarchy, descriptions and relationships before choosing a meaningful target-language term.
This is the opposite of dictionary replacement. A source term can be ambiguous in isolation. The concept identifier and its formal relationships provide context that helps the translation team choose an accurate target description.
16. Language reference sets manage preference by language and dialect
SNOMED CT can mark descriptions as preferred or acceptable in particular languages or dialects through language reference sets. A term preferred in one English variant can be different from the preferred form in another.
Do not overwrite these preferences with a project-wide “one translation fits all” rule. Locale matters, especially when clinical vocabulary differs by region while the underlying concept remains the same.
17. Inactive concepts require historical awareness
Terminologies evolve. A concept can become inactive while historical records still contain its identifier. Translation of an old clinical record should preserve the original coded data rather than silently replacing it with a newer active concept.
If a migration project requires replacement or historical association mapping, use the terminology service and authorized rules. That is a data-maintenance task distinct from translating the original record.
18. SNOMED CT extensions can add local content
Member countries and organizations can maintain SNOMED CT extensions and language resources. Local content can coexist with the international edition. A translator needs to know which edition and extension the source system uses before assuming every identifier belongs to the international core.
Record the module, edition or terminology-server context when the project requires exact reproducibility. A numeric concept identifier without system metadata can be difficult to interpret safely outside its environment.
19. LOINC identifies observations, measurements and documents
LOINC is an international standard for identifying health observations, measurements and documents. Its codes are widely used for laboratory and clinical results, enabling systems to exchange data while retaining a common observation identity.
This is different from diagnosing a condition. A LOINC code often represents the question being asked by a test or observation: what is being measured, in which system or specimen, over what time, on what scale, and sometimes by what method.
20. A LOINC code is a unique permanent identifier
LOINC states that its numeric code is unique and permanent. The code itself has no intrinsic semantic structure beyond its check digit; the meaning is stored in associated terminology fields.
That means translators should not try to interpret or translate the digits. Preserve the code and work on the human-readable term. Store it as text so hyphens and digits survive exactly.
21. The LOINC fully specified name has structured semantic parts
LOINC’s fully specified name combines major parts such as Component, Property, Time, System, Scale and, where relevant, Method. These parts distinguish observations that can look similar in ordinary language.
A translation should therefore preserve the semantic distinctions among specimen, measurement property, timing and method. A short everyday label may be readable but too vague to represent the same LOINC concept.
22. Specimen or system is clinically meaningful
Two tests for the same analyte can have different LOINC codes because one uses serum, another urine, and another a different specimen or system. Translators must preserve the specimen term accurately.
Do not shorten every target label to the analyte name alone. That can remove the distinction that makes the coded observation useful.
23. Method can distinguish observations
Where method is significant, LOINC can distinguish observations by method. A translation that drops the method from a technical display name can make two different observations look identical to a human reader even though their codes differ.
Use an approved target-language method term or retain the established international technical name when that is standard practice. Do not invent a broad synonym that removes specificity.
24. Scale matters: quantitative and qualitative are not interchangeable
A laboratory observation can report a numeric quantity, an ordinal result, a nominal category or another scale. LOINC distinguishes these properties. The target label should therefore communicate the same scale where it is part of the term.
Do not turn “presence” into “concentration” or a qualitative result into a quantitative one because the target language has a more familiar phrase. The code anchors the observation; the translation must respect its semantic dimensions.
25. Units are not the same thing as the LOINC observation code
A LOINC code identifies the observation or question; the reported result may also carry a unit such as mmol/L, mg/dL or another measurement unit. Translators must not treat the unit as part of the LOINC code or infer one merely from the analyte name.
Preserve the unit supplied by the result system and follow the project’s unit-localization policy. Changing units requires validated conversion, not linguistic translation. A readable target label with the wrong unit can be clinically misleading.
26. LOINC code and local laboratory code can coexist
Health systems often retain local test codes while mapping them to LOINC for interoperability. Both may appear in an interface or export. Do not replace the local code with the LOINC code merely because the latter is international.
Translate labels around the two fields so users understand which identifier belongs to the local system and which belongs to LOINC. Mapping decisions should be governed by laboratory or terminology experts, not inferred from translated names alone.
27. Mapping is not translation
Mapping connects concepts or codes between systems. Translation changes human language while aiming to preserve meaning. The two processes can happen in one project, but they require different evidence and review.
A phrase translated perfectly into another language does not prove that a new ICD, SNOMED CT or LOINC code should be assigned. Conversely, an approved terminology map can connect codes whose preferred labels are not literal translations. Keep a separate audit trail for language changes and coding-system transformations.
28. One phrase can represent different concepts in different contexts
Clinical language is context-sensitive. A short phrase such as “cold,” “positive,” “fall,” or “screen” can refer to different meanings depending on specialty, surrounding text and workflow. Controlled terminologies exist partly to remove that ambiguity at the coded level.
Translate with the code, hierarchy and surrounding record visible whenever possible. A term-only translation can choose the wrong meaning even when its dictionary equivalent is linguistically sound.
29. Negation changes clinical meaning without changing every noun
“Pneumonia” and “no evidence of pneumonia” contain the same disease word but communicate opposite propositions. Clinical translation must preserve negation, uncertainty and scope before any coder or system interprets the text.
Do not infer a positive diagnosis code from a negated sentence. If coded data already exists in the source, preserve it and flag apparent inconsistencies for qualified review rather than repairing them through translation.
30. Suspected, confirmed, history of and family history are distinct meanings
Clinical documentation uses qualifiers that change what a statement says about the patient. “Suspected,” “confirmed,” “history of,” and “family history of” are not stylistic variants. A translation that drops one can transform the clinical proposition.
Preserve these qualifiers explicitly. If the target terminology has governed descriptions for such concepts, use the authorized terms rather than compressing everything into the bare condition name.
31. Laterality and anatomical specificity matter
Right, left, bilateral, upper, lower, proximal and distal distinctions can be clinically significant and may correspond to different terminology concepts or classification detail. Translators should never drop them merely to shorten a phrase.
When a target language expresses anatomical relationships differently, restructure the sentence while preserving the exact side and location. A fluent but less specific term can weaken coded-data alignment.
32. Severity and stage are not decorative adjectives
Mild, moderate, severe, acute, chronic, recurrent, stage and grade descriptors can distinguish clinical concepts or classification categories. Translate them consistently and preserve their relationship to the condition.
Do not standardize two source terms into one target term simply for stylistic consistency if the source distinguishes them. Clinical language often uses small modifiers to carry structured meaning.
33. Human-readable labels and code-system identifiers should be stored separately
A robust multilingual data model stores the code, code system, version and display text in distinct fields. That allows the target interface to show a localized description without changing the underlying coded value.
Do not concatenate everything into one free-text translation field. Structured separation supports terminology updates, language switching, auditability and machine exchange.
34. Code-system URI or namespace belongs to the machine layer
HL7 FHIR and other interoperability formats can identify code systems with URIs or other machine identifiers. For example, LOINC uses a defined code-system URI in FHIR contexts. These strings are not user-facing language and should not be translated.
Translate documentation that explains the field, not the URI itself. Changing a code-system identifier can make the payload point to a different or nonexistent terminology.
35. FHIR resource keys and JSON properties are not translation targets
A FHIR payload can contain keys such as coding, system, code and display. The JSON keys and code-system fields form a machine contract. The display string can be localized under the application’s terminology policy.
Do not translate JSON property names inside a standards-compliant payload. If developers want a localized user interface, translate the labels generated around the data rather than the transport schema.
36. HL7 messages can carry local and standard codes together
Legacy HL7 Version 2 messages and other exchange formats can carry coded values alongside local descriptions. Translators working on specifications or testing material must distinguish message syntax from display content.
Protect delimiters, field positions and code-system values. Translate explanatory documentation and permitted text fields only. A pipe, caret or code value can be structural rather than typographic punctuation.
37. Right-to-left interfaces need code-display testing
Arabic, Hebrew and other right-to-left interfaces can visually reorder Latin letters, digits, hyphens and punctuation. ICD strings, LOINC codes and SNOMED identifiers should remain readable in their logical order even when surrounding labels are right-to-left.
Test the real application or exported PDF rather than relying on the editor view. Bidirectional rendering bugs can make an unchanged code appear different to the user.
38. Do not localize digits inside clinical code fields
A target locale may display ordinary numbers using localized numeral glyphs. Clinical code systems and interoperability software generally expect the identifier exactly as defined by the terminology. Keep code characters unchanged.
Localize surrounding dates, quantities or prose only under the relevant project rules. Code-field preservation and locale formatting should be separate formatting paths.
39. Spreadsheets can damage long terminology identifiers
SNOMED CT identifiers are long integers and can be corrupted when spreadsheet software converts them to scientific notation or loses precision. LOINC codes can be misread as arithmetic expressions because they contain a hyphen. Store terminology identifiers as text.
Define column types before import and compare exact values after export. A linguistically untouched code can still be damaged by the tool used to transport it.
40. CSV files need an explicit schema
CSV does not carry strong data types by itself. The receiving application decides whether a value is text, number or date. For clinical terminology projects, define the schema externally and import code columns as strings.
Keep code system, code, display, version and language in separate columns when possible. This reduces the risk that a translator edits the wrong layer.
41. OCR is unsafe as the only source for clinical codes
Scanned discharge summaries, laboratory reports and old coding sheets can contain characters that OCR misreads. One symbol in an ICD code or one digit in a LOINC or SNOMED identifier can change the record.
Verify critical codes against structured source data or an authoritative terminology service when available. If a character is unclear and no verification source exists, flag the uncertainty instead of guessing.
42. PDF extraction can separate code from description
A report table may visually align code, description, result and unit, while text extraction reorders columns. Translation from extracted text alone can attach the right code to the wrong description.
Inspect the page image or structured export whenever field relationships are uncertain. Clinical data quality depends on preserving the pairing, not merely preserving each individual token.
43. Translation memory can spread a wrong code-label pair
If a translation-memory segment contains an outdated or incorrectly paired clinical term, automatic reuse can propagate the problem across thousands of records. High match percentage is not proof of terminology correctness.
Keep governed terminology resources separate from general translation memory. Where an official language release exists, let it outrank informal historical translations.
44. Official terminology translations should outrank literal translation
SNOMED CT language reference sets, WHO ICD language releases and LOINC linguistic variants exist to provide governed target-language terminology. When the relevant official or authorized target content exists, use it rather than inventing a new literal rendering.
This does not mean every user-facing sentence must copy a terminology title word for word. It means the clinical concept display should remain anchored to the governed language resource while surrounding prose can be written naturally.
45. Locale and dialect preferences can change the preferred term
Clinical language varies across countries that share a language. SNOMED CT explicitly supports language and dialect preferences, and LOINC supports linguistic variants. A term acceptable in one jurisdiction may not be the preferred form in another.
Specify target locale, not just target language. “English,” “Spanish,” or “Portuguese” alone may be insufficient for a clinical terminology project that must align with a particular healthcare system.
46. Version control is part of meaning preservation
WHO, SNOMED International and LOINC issue updated releases. New concepts can be added, descriptions can change, and content can become inactive or be replaced. The target should record which release it follows when reproducibility matters.
Do not silently refresh terminology during translation unless the project requests an update. A historical dataset may need the historical label and code relationship preserved exactly.
47. Clinical code translation should not expose patient identity
Terminology codes themselves are not patient identifiers, but the surrounding record can contain names, dates, accession numbers and other personal information. Translation teams should handle files under the project’s privacy and access controls.
For examples and training, use fictional or de-identified records. Do not publish real clinical screenshots merely to demonstrate how a code appears in an interface.
48. Worked example: ICD code plus translated category title
Imagine a fictional training record containing an ICD code and an official source-language category title. The target-language release of the same ICD version provides an authorized title. The translator keeps the code and version metadata unchanged and displays the authorized target title.
The translator does not use the translated phrase to choose a different code. If the source code and source title disagree, the discrepancy is flagged for the coding team rather than repaired through linguistic intuition.
49. Worked example: SNOMED CT concept with several descriptions
Imagine a concept that has a fully specified name, a preferred term and several synonyms. A target-language reference set supplies an approved preferred term. The concept identifier remains unchanged while the interface displays the target preferred term.
A search screen may also expose acceptable synonyms to help clinicians find the concept. Translation therefore enriches human access without creating multiple concept identities.
50. Worked example: LOINC observation with result and unit
Imagine a fictional laboratory record containing a LOINC code, target-language observation name, numeric result and unit. The translator localizes the observation display using the appropriate LOINC linguistic resource, preserves the code, and leaves result conversion outside the translation task unless explicitly authorized.
The code continues to identify the same observation. The result and unit continue to describe what was measured. The language changes only the human-readable presentation.
51. Error clinic: twelve failures that can survive fluent medical translation
One: recoding from translated prose. Two: replacing ICD-10 with ICD-11 because the title looks similar. Three: ignoring a national ICD modification. Four: converting a SNOMED identifier to scientific notation. Five: translating a SNOMED term word for word without concept context. Six: dropping laterality or negation.
Seven: treating local lab code as LOINC. Eight: removing specimen or method from a LOINC display. Nine: translating a FHIR code-system URI. Ten: changing code digits to localized numerals. Eleven: using terminology from the wrong version. Twelve: repairing a source coding inconsistency without clinical review.
52. A practical release checklist
ICD: correct family, version and jurisdiction preserved. SNOMED CT: concept identifiers unchanged; approved target descriptions used. LOINC: code unchanged; Component, Property, Time, System, Scale and Method distinctions retained where relevant. Mappings: separated from translation decisions.
Data files: code fields stored as text and protected from formatting. Interfaces: code-system URIs and API keys unchanged. Clinical meaning: negation, uncertainty, laterality, severity and history preserved. Privacy: patient data protected. QA: token equality plus code-description relationship checks complete.
53. Frequently asked questions
Should I translate an ICD code? No. Use the code from the applicable classification and translate or use the official target-language description. Can I choose a new ICD code from the translated sentence? Not as an ordinary translation task; coding decisions require the applicable coding rules and qualified review.
Should I translate a SNOMED CT concept identifier? No. Translate the description using the relevant language resources. Should I translate a LOINC code? No. Localize the term, not the identifier. Are mapping and translation the same? No; mapping connects coded concepts between systems, while translation changes human language.
54. Connect this specialist guide to the wider eduKateSG translation architecture
This article owns the specialist search intent of translating ICD, SNOMED CT and LOINC content while preserving clinical classification, concept and observation identity. It does not replace medical coding manuals, terminology services, clinical governance or the broader translation architecture.
The deeper educational principle is consistent with the rest of the Translate series: a human-readable term and a machine-readable identifier can occupy the same record while following different rules. Translate language for the reader, preserve governed codes for interoperability, and escalate coding or mapping decisions to the professionals responsible for those systems.
Authoritative reference points
For ICD, consult the World Health Organization’s International Classification of Diseases resources and ICD tools. For SNOMED CT, consult SNOMED International’s translation and language-preference guidance and terminology documentation. For LOINC, use the LOINC Users’ Guide, major-parts guidance, and international translation resources. These authorities maintain the governed code systems; translators should use their current releases and language assets for production work.