Can Machine Translation Errors in an IVDR IFU Delay CE Marking?
- 2 hours ago
- 7 min read
Yes. A machine-translation error in an IVDR instructions for use can hold up your CE marking, because for most in-vitro diagnostics a notified body now reads that leaflet before the device reaches the market, and a wrong word in the intended-purpose or warning text is precisely the kind of nonconformity an assessor writes up. Under the old In Vitro Diagnostic Directive (98/79/EC), manufacturers self-certified roughly 80% of IVDs. The In Vitro Diagnostic Regulation (Regulation (EU) 2017/746) reverses that: around 80 to 90% of IVDs now need a notified body, so your leaflet is read by a reviewer, not just printed and shipped.
That reviewer matters here. The IFU sits inside the technical documentation a notified body assesses, and under the IVDR every risk class except non-sterile Class A needs that body, with Class D checked by an EU reference laboratory on top. Language that used to be a late production detail is now assessed evidence.
What the IVDR asks of an IFU, and in which language
Article 10(10) sets the rule. The information supplied with the device goes to the user or patient in the official language or languages decided by each member state where the device is made available, and Annex I, Section 20 defines what that information covers, from intended purpose to warnings and performance claims, whether it sits on the device, the packaging, or in the leaflet. Member states can accept another language, usually English, for professional-use devices where safe use is not at risk, so a lab analyser and a home self-test can face different language demands in the same country.
Take a coagulation analyser sold in France, Germany, and Poland. Its IFU has to reach each site in French, German, and Polish unless a given market accepts English for professional users, which several do and several do not. Miss one language, or ship a version that reads differently from the approved source, and you carry a gap that surfaces at assessment rather than after launch. Our explainer on which MDR documents need translation into every EU language walks the same language-scope logic for devices under the MDR.
Why a machine-translation slip becomes a conformity gap
The risk sits in two fields. Intended purpose and warnings carry almost all of it. Change an analyte, a specimen type, or a cut-off value, or flip a single negation, and the device now claims something the manufacturer never validated. A notified body reads the translated IFU against the intended purpose and the performance evaluation in the technical file, so a mismatch there becomes a conformity finding the assessor has to record.
The next steps are procedural. The assessor raises a nonconformity, asks for corrective action, and holds the file open until a corrected IFU and its evidence arrive, all on the notified body's schedule rather than yours. Expect a CE-marking delay measured in review cycles, on a device that was otherwise ready to ship.
The transition clock makes a rejected IFU expensive
Timing is what turns a delay into a real problem. Regulation (EU) 2024/1860 extended the IVDR transition for legacy devices, but only for devices that need a notified body, and it set hard end dates by risk class:
Class D devices can stay on the market until 31 December 2027, with the notified-body application due by 26 May 2025 and a signed agreement by 26 September 2025.
Class C devices have until 31 December 2028, with the application due by 26 May 2026 and the agreement by 26 September 2026.
Class B and sterile Class A devices have until 31 December 2029, with the application due by 26 May 2027 and the agreement by 26 September 2027.
Non-sterile Class A devices get no extension and must already meet the IVDR in full.
A legacy device here means one covered by a valid certificate or declaration of conformity under Directive 98/79/EC before 26 May 2022. The dates do not move. A leaflet that bounces back from a notified body then costs more than a review cycle, because it can push a submission past the class deadline, and a device that misses its date comes off the market until it conforms. You can see how EUDAMED expects consistent multilingual device data to hold the same label content across every market version.
Where machine translation actually breaks on an IFU
Raw neural machine translation does not fail at random. On regulated text it fails in specific, repeatable ways, and these are the ones that reach a notified body:
Negation and modality flips, where a warning ends up telling the user to do the thing the source told them to avoid.
Intended-purpose drift, where a loose synonym quietly changes the analyte, the specimen type, or the target population.
Units, cut-offs, and reference ranges left in the source format or converted wrongly for the target market.
Symbol and terminology mismatches between the label and the IFU, against the harmonised terms in EN ISO 15223-1.
Silent omission, where the engine drops a clause it cannot parse and leaves no trace that anything is missing.
A spell-check catches none of these. They surface only when a qualified reviewer who reads the target language and knows the assay compares the machine output against the approved source, line by line, which is the step that separates a real speed gain from a hidden compliance risk.
What controls the risk
Governing machine translation beats banning it from regulated work. ISO 18587 sets the requirements for full post-editing of machine-translation output by a qualified human, while ISO/IEC 42001 puts that machine step inside an audited AI management system, so the process a notified body might ask to see is documented rather than improvised. The engine drafts. A qualified linguist decides what ships.
We at AD VERBUM run machine translation only inside that ISO 18587 workflow, on client-tuned open-weight models hosted on our own EU infrastructure, never on public web tools that would put your assay data outside your control. ISO 13485 ties each translated IFU to a named source version under change control, ISO 17100 puts a second qualified linguist on the revision, and ISO 27001 keeps the data on EU-hosted systems. Because the model is constrained by your term base, the analyte names, cut-offs, and warning phrasing come out consistent across every language version, and the safety-critical fields carry a back-translation record into the technical file. This is the same discipline we bring to MDR and IVDR device documentation and to work that has to meet ISO 13485 for medical device translation.
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 ISO 13485 medical device quality management and ISO/IEC 42001 AI-management governance 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
Does the IVDR really require a translated IFU for every market?
Yes, within limits set nationally. Article 10(10) of Regulation (EU) 2017/746 says the information supplied with the device must be in the official language or languages of each member state where it is made available, and Annex I, Section 20 defines that information. Some member states accept English for professional-use devices where safe use is not compromised, so the exact language set depends on where you sell and to whom.
Can I use machine translation for an IVDR IFU at all?
Yes, if it is post-edited. Nothing in the IVDR bans machine translation, but unreviewed output is a risk in the intended-purpose and warning text. ISO 18587 defines full human post-editing of machine-translation output, and running that step keeps the machine draft from reaching a notified body unchecked.
Which IVD classes need a notified body to review the IFU?
Every class except non-sterile Class A. Under the IVDR, sterile Class A, Class B, Class C, and Class D devices all go through a notified body, and Class D adds an EU reference laboratory. Non-sterile Class A devices are self-certified, so the manufacturer carries full responsibility for the IFU without an external reviewer.
What are the IVDR transition deadlines for legacy devices?
Under Regulation (EU) 2024/1860, legacy Class D devices may stay on the market until 31 December 2027, Class C until 31 December 2028, and Class B and sterile Class A until 31 December 2029, each subject to notified-body application and agreement dates. Non-sterile Class A devices get no extension. A legacy device is one covered by a valid Directive 98/79/EC certificate or declaration of conformity before 26 May 2022.
How does a translation error actually delay CE marking?
The notified body assessing your technical documentation reads the IFU against the intended purpose and performance evaluation. A mistranslated warning, cut-off, or intended-purpose statement is a conformity finding, which triggers a corrective-action request and a re-review on the body's timeline. That review cycle is the delay, and it can arrive close to a class transition deadline.
How does ISO 13485 help control IFU translation risk?
ISO 13485 is the quality-management standard for medical devices, and for translation it ties each IFU version to a named source under change control, with records a notified body can inspect. Combined with ISO 17100 revision by a second linguist and ISO 27001 data controls, it gives the technical file the documented trail that IVDR assessment expects.
