Which Medical Device Documents Suit AI Translation With Human Review
Six medical device document types carry a legal language obligation under the Medical Device Regulation, and every one of them can move through AI translation, as long as a certified linguist signs off on the result. The document decides how much review, not whether you get to skip it. A dose in an instruction for use and a header field in a database are not the same risk, and the workflow should not treat them as one.
AD VERBUM is an EU-hosted translation company that runs every one of these documents through a client-tuned language model and certified medical linguists, human-verified under ISO 13485 and ISO 17100, with no training on your data. We cover instructions for use, labels, the summary of safety and clinical performance, implant cards, and the device records you file in EUDAMED, in the official languages each member state requires.
The Medical Device Regulation, Regulation (EU) 2017/745 (MDR), never names a translation tool. Article 10(11) makes you keep the information supplied with a device accurate and in the language or languages each member state determines, and Article 10(9) makes your quality management system responsible for it. So the question is not whether AI translation is allowed, which we cover in our explainer on whether MDR allows AI translation. The question is which documents suit it, and how much certified review each one needs before a notified body accepts the file.
The six documents, ranked by how much review they demand
These are the device documents suited to AI translation with certified human review, ordered from the highest patient-facing risk to the lowest. For each, here is the MDR language rule, what goes wrong if the model output ships unreviewed, and the certified control that keeps the file defensible.
1. Instructions for use (IFU)
The IFU is the highest-stakes document, because a clinician or patient acts on it directly. Annex I Section 23.4 sets what it must contain, and Article 10(11) requires it in the languages of each market. An unreviewed model can flip a negation in a contraindication or render a warning that reads clearly in English but scans wrong in Dutch, and a notified body rejects the technical file on that alone. The control is full human post-editing under ISO 18587 by a medical linguist, with terminology locked to your device term base, so every warning matches the approved source.
2. Device labels
A label carries symbols plus text in a tight space, and the language obligation under Annex I Section 23.2 applies to every word of it. The risk with raw machine output is quiet: a storage condition, a unit, or an expiry format that reads plausibly but means something else in the target language. Certified review pairs the linguist with a label proof check, so the translated string still fits the layout and still says what the source says. Short strings hide errors, which is exactly why a human sign-off matters more here, not less.
3. Summary of safety and clinical performance (SSCP)
For implantable and class III devices, Article 32 requires an SSCP, validated by your notified body and uploaded to EUDAMED in English plus every market language. It is treated as labelling, so clinical claims carry regulatory weight and any drift between versions is a finding. AI translation produces a consistent first version fast, then a certified reviewer checks it back against the clinical evaluation and the validated English. Two-step review is not optional on an SSCP, because the notified body reads the translations, not just the source.
4. Implant card
The implant card and the information supplied with an implantable device fall under Article 18, and the text must be in the language or languages the member state determines and written so a layperson understands it. That last part is where machine output fails: it produces technically correct but clinically dense phrasing a patient cannot use. We treat the implant card as patient-facing text, so certified review works to plain-language readability, not just a correct rendering. We set out the language scope in detail in our note on the implant card language rule under MDR Article 18.
5. UDI data and device registration
The unique device identifier under Article 27 and the device registration under Article 29 are structured data, not prose, but the descriptive fields still need consistent multilingual wording. The failure mode is drift: the same device attribute rendered two ways across languages, which a reviewer reads as a discrepancy in the record. A governed term base carries one approved equivalent per language, so the attribute stays identical everywhere it appears. This is lower narrative risk than an IFU, but the consistency demand is just as strict.
6. EUDAMED device records
EUDAMED is the EU database that ties the rest together, and its four core modules, the actor, UDI and device, notified body and certificate, and market surveillance modules, became mandatory on 28 May 2026, with legacy devices to be registered by 28 November 2026. The same registration duty runs under the in vitro diagnostics regulation, Regulation (EU) 2017/746 (IVDR). Device descriptions and clinical summaries entered here must stay consistent with the IFU, label and SSCP already approved, so a single term base and certified review keep the database aligned with the documents feeding it. Inconsistent EUDAMED wording is the cheapest error to make and the most visible to a regulator.
The audit trail comes from your records, not the model
A notified body does not audit your translation engine. It audits the records that show a competent person reviewed and released each version, which is what ISO 13485 requires of the device quality management system and what ISO 17100 requires of the translation process. The model is a tool inside that system. What makes a translated IFU defensible is the documented certified review behind it, the same point we make on whether AI translation is safe for device IFUs and labelling.
This is why the same document can suit AI translation for one manufacturer and fail for another. The text is identical. The difference is whether a qualified linguist took responsibility for it under a certified workflow, and whether the record proves it. AD VERBUM keeps that trail on every job, so a device manufacturer can hand the audit a version history, not an assurance.
Why EU hosting and ISO 27001 belong in this
Device documentation often carries pre-market data, and where it is processed matters as much as who reviews it. Four things separate a defensible AI translation workflow from a convenient one:
EU-hosted processing, so pre-approval device data does not transit public-cloud endpoints outside the jurisdiction that regulates it.
No training on your data, so your device content never becomes a model's future output, a distinction we explain in what AI translation means for regulated content.
ISO 27001 information security across the whole workflow, from file intake to delivery, with access controlled per project.
GDPR data residency, so any personal data in clinical summaries or vigilance records stays within EU infrastructure end to end.
We built this stack for regulated devices, not adapted a consumer tool to it. Our certified medical linguists work under ISO 13485 and ISO 17100, on client-tuned open-weight models we host in the EU, which is the same workflow behind our AI translation services for medical device IFUs and labelling. For the wider life-sciences picture, see our guide to AI translation companies for life sciences.
Our medical device translation services
Our translation services for regulated sectors run on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure, with no reliance on public cloud tooling for core processing. Every project runs through our AI+HUMAN hybrid workflow: we ingest client Translation Memories and Term Bases first, our proprietary LLM-based LangOps System generates output constrained by client terminology on client-tuned open-weight models, and our certified subject-matter experts review for technical accuracy and regulatory compliance. Our QA is aligned to ISO 17100 and ISO 18587, with sector-specific requirements such as MDR Article 10(11) and ISO 13485 device quality management applied where relevant. We serve Life Sciences, Legal, Finance, Defense, and Manufacturing clients across 150+ languages with 3,500+ subject-matter linguists. For teams managing audit-sensitive content, contact us to discuss your security and compliance requirements directly.
FAQ
Which medical device documents can use AI translation?
Instructions for use, labels, the SSCP, implant cards, UDI and device-registration data, and EUDAMED records can all use AI translation with certified human review. Under MDR Article 10(11) each must appear in the language or languages the member state determines. The document does not change whether review is needed, only how much.
Does the MDR require human review of AI translation?
The MDR does not name any tool, but Article 10(9) requires a quality management system that keeps translation accurate, and ISO 13485 and ISO 17100 make a competent human responsible for each released version. In practice that means certified post-editing under ISO 18587, not unreviewed model output.
What is the highest-risk device document to translate?
The instructions for use, because a clinician or patient acts on a dose, contraindication or warning directly. An error there can reach a user, so it takes full certified review every time under ISO 18587, with terminology locked to the device term base.
In which languages must device documents be translated?
In the official language or languages of each member state where the device is placed on the market, under MDR Article 10(11). Member states set their own requirement, so the exact set varies by market. The implant card under Article 18 must additionally be written for a layperson.
When did EUDAMED become mandatory?
The four core EUDAMED modules became mandatory on 28 May 2026, and legacy devices must be registered by 28 November 2026. Device descriptions entered there must stay consistent with the approved IFU, label and SSCP, which a single governed term base maintains across languages.
Is AI translation of device documents secure?
It is when the workflow is EU-hosted, ISO 27001 certified, and trains no model on your data. AD VERBUM processes device documentation inside EU infrastructure with GDPR data residency, so pre-market and personal data never leaves the jurisdiction that regulates it.
