Translating regulatory documents: FDA, EMA and other compliance challenges

A pharma client sends over a file and asks how fast it can be in German and Spanish. It is a package leaflet with the labelling annex attached. From the outside this looks like ordinary technical work, and the word count agrees. Regulatory document translation is different in one specific way: a sentence here can be fluent, accurate and still come back with a comment, because somebody has already decided what that sentence says in that language. The wording is not a choice the translator gets to make. It is a lookup.

We have watched good teams lose a week to this. Not because the linguists were weak, but because the process treated an authority's wording as a style preference.

What makes regulatory document translation a different job

In most translation work the translator has room. There is a brief, a glossary, a register to hit, and inside that the target wording belongs to them. Regulated content takes a piece of that room away. Some of the text has an official equivalent sitting somewhere already: in a template, in a standardised term list, in the version of the same document an authority approved three years ago. On those spans the task is retrieval. Write something better and you have produced a finding.

The rest of the file behaves completely normally. A clinical overview, a non-clinical summary, the body of an investigator's brochure: that is technical prose, and a translator with the right domain background handles it the way they handle any technical prose. The awkward part is that both kinds of text live in the same document, sometimes in the same paragraph, and nothing in the file marks where one ends.

The second difference is who reviews. In commercial work the client reviews, and a disagreement costs an email. Here the reviewer can be a national competent authority, the comment arrives in a language nobody on the project team reads, and the loop runs against a statutory clock you cannot negotiate with.

None of this applies to internal documents. An SOP for your own staff, an internal audit report, a training deck for the quality team: those are regulated-industry documents, not regulatory submissions, and treating them like submission content burns budget for no reduction in risk. The distinction worth drawing is not pharma versus other, it is whether a regulator will read the target text.

FDA submissions: English is the deliverable and the original travels with it

For US filings the language question is mostly settled before it is asked. FDA works in English. Where the regulation gets specific is what happens to foreign-language source material: 21 CFR 314.50(g)(1) requires that reports of foreign clinical studies and literature submitted in an NDA come with a complete and accurate English translation, and with a copy of the original. The same pattern appears for INDs under 21 CFR 312.23.

So the deliverable is a pair rather than a file, and that changes the formatting job. The translation has to stay navigable against the original: table numbering preserved, figure captions in the same order, headings that let a reviewer put the two documents side by side without hunting. We once had a set of translated study reports bounced back with no linguistic objection at all. The translator had merged two short tables into one, renumbered everything after it, and broken every cross-reference in the English text that pointed at a table number from the original.

Sponsors often ask the translation provider for a signed statement of accuracy. On its own it checks nothing. What it does is move the risk somewhere visible, and in our experience people work more slowly and read their own output more carefully when their name goes on that page.

Worth separating from all of this: Spanish-language patient material for the US market. That is an access and labelling decision rather than a submission-language rule, and the quality bar is different, because the reader is a patient instead of a reviewer. Teams that run both through the same workflow tend to over-engineer one and under-serve the other.

EMA and the EU: translation is a regulated step with its own clock

In the EU centralised procedure the product information has to exist in every official EU language before the Commission decision. Translations of the summary of product characteristics, the labelling and the package leaflet are produced after the CHMP opinion, inside a short window, and then go through a linguistic review where Member States comment on their own language version. The precise timings differ by procedure and get revised, so check current EMA guidance rather than any blog post, this one included.

The operational consequence is the part people underestimate. The translation step cannot start when the opinion arrives. Two dozen languages in a handful of days is not a scheduling problem you solve with capacity. Teams that handle it well have the terminology, the templates and the per-language reviewers lined up months in advance, and treat the opinion as a trigger for a prepared process rather than the start of a project.

Two sources do most of the work. EMA publishes product information templates per language with fixed headings and standard statements, so a translator producing a leaflet is filling a frame that already exists in the target language. And the EDQM Standard Terms database gives official per-language terms for pharmaceutical dose forms, routes of administration and container types. "Powder and solvent for solution for injection" has one correct German rendering. A fluent alternative is simply wrong.

We saw the cost of missing that on an EN to DE leaflet where the dose form had been translated freely. The German was good German. It was consistent across the document, which is the detail that makes this hard to spot in review, since consistency is what reviewers usually check for. The comment came back from the Member State, and the fix had to happen inside a window that was already short. The second recurring one is adverse reaction terminology: MedDRA has official translations, and where the target text will be read against a safety database, matching those terms matters more than finding a more readable synonym.

Devices and diagnostics: the language question belongs to the Member State

Under EU MDR, Regulation (EU) 2017/745, and under IVDR for diagnostics, the language of instructions for use and labelling is set by the Member State where the device is made available. There is no single EU list to translate into. The list comes out of your market map, and it grows every time sales opens a country.

That produces a different shape of work. Device IFU translation is maintenance rather than a project: the source gets revised, and every language has to catch up. Version drift is the most common failure we see in this content, by a wide margin. One manufacturer shipped eleven language versions of an IFU in which a newly added warning existed in English and in the two languages that happened to be translated last. The other eight came from a copy of the source taken two weeks before the change. Nothing in any of the files recorded which source revision they were made from, so finding out took a day of comparing documents by hand.

Harmonised symbols under ISO 15223-1 carry some of the load and cut translated word count, which also cuts the number of ways to get it wrong. They stop at free-text warnings and clinical instructions, which is where the risk sits anyway.

Put the source revision in the file name

The cheapest fix we know for drift is naming discipline. If the English source is revision 7, every target file carries 7 in its name and in a footer field, and the reviewer's first check is whether those match. It takes a project manager about a minute per file and it catches the failure above before the files leave the building.

Where regulatory compliance translation accuracy actually breaks

In the findings we have traced back, the cause is almost never vocabulary. Four things account for most of it.

Numbers and symbols come first. Decimal comma against decimal point, thousands separators, comparison operators. An investigator's brochure we reviewed had a seven-page dosing table where roughly a third of the rows had moved to a decimal comma and the rest had not, and in two rows a greater-than-or-equal sign had become a plain greater-than. Every cell read plausibly on its own. The table as a whole was unusable, and nothing in a normal linguistic review looks at it.

Then there are the frozen phrases: standard statements, required warnings, hazard and precautionary wording, sentences that exist in law. Fluent paraphrase is a defect. The same logic governs safety data sheets, where whole numbered sections are reuse rather than translation, and we went through that case in detail in our piece on translating safety data sheets.

Structure and cross-references make up the third group. "See section 4.4" is only correct if the target document carries the same numbering, which it does when the translator works inside the official template and does not when somebody rebuilt the file from a Word export.

The fourth is version drift, and it is the worst of them, because it produces a technically perfect translation of text that is no longer current. Linguistic review cannot find it. Only a revision check can.

Build the reference set before anyone translates a sentence

The preparation work is dull and it decides the outcome. Before a word gets translated, we want five things in one folder: the previously approved target-language version of this product or the closest sibling product, the official template in that language, the standard terms for the dose form, route and container, the official translation of any coded terminology in use, and a do-not-translate list.

Term status matters as much as term content. A two-column word list cannot tell a translator that a rendering was correct in the 2022 approved version and must not be used now. A termbase with an approved, superseded or forbidden flag can, and the superseded rows are the ones that stop a reviewer from "fixing" new text back to old wording. How domain glossaries for this kind of content differ from general translation glossaries is something we covered in our post on medical translation.

The do-not-translate rows do most of the work

For a device client we built a glossary of around 180 entries. Forty of them were do-not-translate rows: product name, model designations, strength expressions, the legal entity name, the notified body number, standard references like ISO and IEC numbers. Those forty prevented more findings than the other hundred and forty combined. Translators localise things that look localisable, and a model designation that happens to contain a real word is exactly the kind of thing that gets helpfully adapted.

This works best when there is an approved previous version to mine. For a first filing in a language with no history, the reference set is thinner and the per-language reviewer carries more of the weight, which is worth knowing when you quote the job.

What back translation proves and what it does not

For patient-reported outcome instruments and similar measurement tools, the established process runs well past translation plus review: independent forward translations, reconciliation, back translation, harmonisation across languages, cognitive debriefing with patients. The ISPOR task force report on translation and cultural adaptation of PRO measures (Wild et al., 2005) is still the reference people cite for it.

That process earns its cost when scores from different language versions get pooled and have to mean the same thing. We have also seen it bought for documents that do not need it, which is where the budget goes without buying any safety. Back translation is a detector with a narrow range. It finds omissions, additions and reversed meaning. It does not find a near-synonym used in place of an official term, because the near-synonym back-translates perfectly into the source wording.

A terminology verification pass against the reference set finds that class. A reviewer with a regulatory background finds a third one: text that is accurate, well written, and in the wrong place, like a warning that ended up outside the section the template reserves for it. When a client has had to choose on a fixed budget for non-instrument content, terminology verification plus a regulatory reviewer has caught more for us than back translation did.

Start here

If you take on this kind of work and want the next project to go better than the last one, the order matters.

1. Before quoting, split the file into prescribed text and free text, and count them separately. The prescribed portion is cheap in translation time and expensive in retrieval time, and quoting it per word is how agencies lose money on leaflets.
2. Collect the reference set into one folder and make it the first thing a translator opens. If it does not exist for that language, say so in the quote and price the reviewer accordingly.
3. Stamp the source revision into every file name and footer, and make matching it the first review check.
4. Run a numeric and symbol pass as its own step, by somebody who is reading the tables and not the prose.
5. Decide the review model per document type rather than per client. Instruments get the full process. Leaflets get terminology verification and a native regulatory reviewer. Internal documents get normal QA.

Almost every finding we have traced back came down to the same thing. Somebody wrote a sentence where they should have looked one up.