Medical Translation
Translate clinical reports, patient material, research summaries, and device instructions with terminology and unit checks close at hand.
Why the subject matters
What changes in this kind of work
Medical material shifts between specialist language and instructions written for patients. Reviewers can use a project glossary, inspect abbreviations and measurements, and check every target segment before approval. The software does not replace clinical or regulatory review.
Medical translation serves two very different readers. A clinical report or a discharge summary is written by clinicians for clinicians, with abbreviations and shorthand. A consent form or a patient leaflet must be understood by someone with no medical background, often at a difficult moment. The same term may need a technical rendering in one document and a plain one in the next, so a project starts by deciding who will read it.
Small details carry most of the risk. A dose of 0.5 mg read as 5 mg, a left that becomes a right, or ‘twice daily’ that becomes ‘twice a week’ changes what happens to a patient. In the editor, numbers and units stay visible beside the source, and the quality checks flag segments where the numbers differ from the source or an approved glossary term is missing. The checks flag; a qualified reviewer decides. The software does not judge clinical meaning.
Medical documents often contain personal health information. The model runs on DeepReference's own servers in the EU, no outside AI company sees the text, and customer content is never used to train the model. Even so, DeepReference is not HIPAA-certified and does not sign Business Associate Agreements, so check your organisation's rules before you upload identifiable patient data.
How it works
A project from upload to delivery
- 1
Build the terminology first
Upload a reference document, or a source with its earlier translation, and run terminology extraction with the domain set to medical (Freelance and up). The domain setting guides which terms are extracted; it does not change how the model translates. Approve or dismiss each term, then add the approved ones to a glossary.
How to do it, step by step: 3. Terminology extraction
- 2
Upload the document
Upload a DOCX, PDF, PPTX or XLSX file of up to 50 MB (75 MB on Business) and choose the target language. For patient material you can choose a tone such as simplified or formal. A PDF needs selectable text; an image-only scan is refused.
How to do it, step by step: 6. Document translation
- 3
Draft with glossary and memory
Approved glossary terms found in a sentence are sent to the model with that sentence. On Translator Pro, an exact match from translation memory is reused, so a warning sentence approved for the last leaflet comes back as it was approved.
How to do it, step by step: 2.15 Use a glossary correctly
- 4
Review segment by segment
The segment editor shows source and target side by side, with glossary hits, memory matches and flags for numbers, untranslated text and missing terms. The reviewer approves each segment, and the version history keeps earlier wordings.
How to do it, step by step: 7. Translation Editor
- 5
Export in the original format
The file comes back in its own format, with tables, headings, numbering, headers and footers in place. A results table in XLSX keeps its formulas. PDFs with complex layouts need a manual check after export.
How to do it, step by step: 8. Document Editor
- 6
Update when the source changes
When instructions for use or a consent form are revised, revision comparison (Freelance and up) shows the source changes as a redline and as a DOCX with real tracked changes. In ‘With translation’ mode, only the segments the revision affected are retranslated, and nothing is charged until you confirm.
How to do it, step by step: 9. Revision comparison
Document types
What we translate
- Clinical reports and lettersWritten for colleagues and dense with abbreviations; each one must be expanded or kept the way the target readers expect.
- Discharge summariesMedication lists, doses and follow-up dates must match the source exactly, because the patient or the next clinician acts on them.
- Patient information leafletsA sentence that is correct but hard to follow does not do its job, so plain wording matters as much as accuracy.
- Informed consent formsRisks and the voluntary nature of taking part must read as clearly in the target as in the source; ethics committees often review the translated version.
- Instructions for use and labels for medical devicesWarnings, cautions and the order of steps must not change, and labels quoted from the device screen must match the device.
- Clinical study protocols and summariesEndpoints, inclusion and exclusion criteria and visit schedules repeat across documents; a glossary and translation memory help keep them identical.
- Laboratory reports and results tablesReference ranges, units and decimal values sit in tables; DOCX and XLSX tables keep their structure, and the reviewer checks every value.
- Patient questionnairesA draft can help, but a validated questionnaire usually needs a separate linguistic validation process, which DeepReference does not carry out.
- Research abstracts and articlesTerminology should follow the published literature in the target language, which a reviewer from the field knows best.
- Summary of Product Characteristics (SmPC)The headings follow the regulator’s template, and posology, contraindications and warnings must say exactly what the approved source says.
- Labels and carton textShort texts with strength, batch number, expiry date and storage conditions, where a single wrong digit matters. Text inside images or artwork is not translated.
- SOPs and work instructionsSteps, responsibilities and references to other SOPs must keep the same order and numbering, so that the translated procedure can be followed exactly like the original.
How to do it, step by step: 6. Document translation
Quality controls
Checks available during review
Domain terminology
A domain glossary or termbase can be applied to keep preferred vocabulary consistent during review.
How to do it, step by step: 3. Terminology extraction
Format and structure checks
Document structure, numbering, tables, and section conventions remain visible for reviewer checks.
How to do it, step by step: 8. Document Editor
Translation memory
Approved translations can be reused when the same or similar text appears again.
How to do it, step by step: 4. Translation memory
Review records
Reviewer actions, translation settings, and approval timestamps are kept where the workflow supports them.
How to do it, step by step: 7.3 Project tools
Glossary filtering
Only glossary terms found in the current text are sent with the translation request.
How to do it, step by step: 2.15 Use a glossary correctly
Recommended settings
How to set up a job in this field
Domain for terminology extraction
Choose Medical when you extract terms from these documents. The domain is used for extraction only.
Step by step in the user guideGlossary
Add the approved terms for this client and language pair, one translation per term, and select the glossary when you translate.
Step by step in the user guideTerminology examples
How key terms are handled
Adverse event / adverse reaction
Two different concepts: an event need not be caused by the treatment, a reaction implies that it was. Put both in the glossary; the check flags a segment where the approved term is missing.
Serious adverse event (SAE)
Has a regulatory definition. Add the term and its abbreviation to the glossary in the form your sponsor or authority uses.
Contraindication
Kept apart from precautions and warnings, each with its own approved target term in the glossary.
Dose and units (mg, µg, mL, IU)
Never converted. The number check flags a segment where the numbers differ from the source; the reviewer checks units, decimal separators and frequencies.
Clinical abbreviations (b.i.d., p.r.n., SOB)
Often ambiguous out of context. A glossary entry states how each one is rendered, and the reviewer confirms the reading.
Drug names (INN and brand names)
International non-proprietary names differ slightly by language, and brand names can differ by country. Add the form approved for the target market to the glossary.
Indication
What a product is approved to treat. It must not be widened or narrowed in translation, because that changes the approved use.
Laterality (left / right)
Not checked automatically. It is on the review checklist because a swapped side is a serious error that no number check catches.
How to do it, step by step: 2.4 Add terms · 2.15 Use a glossary correctly
Who uses this
Industries served
Hospitals, clinics, CROs, medical device teams, research groups, and healthcare service providers.
Standards and references
What reviewers can check
Reviewers can check clinical terms, abbreviations, product names, doses, units, warnings, and patient-facing language.
Before delivery
What your reviewer checks
- Every dose, unit, frequency and route of administration matches the source.
- Decimal and thousands separators follow the target convention without changing the value.
- Left and right, upper and lower, and other directions are unchanged.
- Abbreviations are expanded or kept as the target readers expect, and none is ambiguous.
- Drug names use the INN or brand name approved for the target market.
- Warnings, contraindications and precautions are complete and in the source order.
- Text for patients is clear to a reader with no medical background.
- Every segment flagged by the quality checks has been looked at and resolved.
- Tables, numbering and headers are intact after export, especially in PDFs with complex layouts.
How to do it, step by step: 7. Translation Editor
Transparency
What DeepReference does not do
- DeepReference is not HIPAA-certified, does not claim HIPAA compliance and does not sign Business Associate Agreements (BAAs). Do not upload protected health information where your rules require a BAA.
- DeepReference is not a medical device. It is not intended for diagnosis, treatment or any clinical decision, and its output must be reviewed by a qualified person before use.
- There is no special medical model. The same model translates every field; the medical domain setting guides terminology extraction only.
- DeepReference does not certify translations and holds no ISO 17100, ISO 27001 or SOC 2 certification. Where a certified or sworn translation is required, a certified or sworn translator provides it.
- DeepReference does not provide translators or reviewers, and does not carry out linguistic validation of questionnaires (forward and back translation, cognitive debriefing).
- Scanned or handwritten records without a text layer cannot be translated, and text inside images is not translated. Run text recognition first, or upload the original file.
Questions
Frequently asked questions
Is DeepReference HIPAA compliant, and will you sign a BAA?
No. DeepReference is not HIPAA-certified, does not claim HIPAA compliance and does not sign Business Associate Agreements. If you are a covered entity or business associate under HIPAA, do not upload protected health information; remove the identifiers first.
Where is the text processed, and who can see it?
On DeepReference's own servers in the EU. No outside AI company sees the text. Uploads are encrypted in the browser before sending, and stored files are encrypted at rest. The servers decrypt the text in order to translate it.
Is our content used to train the model?
No. Customer content is never used to train the model. Deleting your account removes your documents, translation memories and glossaries, and backups expire after 14 days.
Is there a model specialised in medicine?
No. The same model translates every field. Consistency in a medical project comes from your glossary, your translation memory and a reviewer who knows the field.
Can we process health data under the GDPR?
The servers are in the EU, and the service is built to support GDPR obligations, including deletion on request. Health data is a special category under the GDPR, so your organisation decides whether and how it may be processed here.
Can DeepReference translate scanned medical records?
Only if the PDF already has a recognised text layer. An image-only scan is refused with ‘Upload a searchable PDF or the original document’. Handwriting and text inside images are not translated.
Review framework
Open the platform from the dashboard when you are ready to work on files.
View quality documentationBegin projects: Medical Translation
Open the platform from the dashboard when you are ready to work on files.