Clinical trial disclosure and redaction, explained.

Clinical trial redaction explained: why a company blacks out parts of its own study report before a regulator posts it

One study report's route to EMA's public website, from invitation to posting

1. The agency sends an invitationAround Day 121 of the review, EMA emails a draft list of the documents it plans to publish. You have two weeks to agree it.4
2. The team sets its methodWhich details count as identifiers, how each will be handled, and what risk ceiling applies.
3. Every page gets markedPatient details, staff names and any confidential passages are marked, labeled and justified. This is where a missed copy of a number happens.
4. The agency reviews the proposalEMA checks the package is complete, then reads the justifications and the method. Its conclusion letter arrives on Day 47 of this review, a clock of its own.1
5. The final copy goes publicThe company applies the agreed redactions within 7 calendar days, and EMA posts the result with its own watermark.1

Step 3 is the subject of this guide. EU trial documents and Health Canada packages follow their own routes, covered in section 02.

Why the patient listings may stay private, and three more redaction surprises

CSR, redaction, anonymization and re-identification risk

Fourteen terms are defined in the one-page summary.

EMA Policy 0070, CTIS and Health Canada PRCI: three public programs for clinical documents, and what the US and UK do instead

In shortThree programs publish full clinical documents: EMA's clinical data website under its Policy 0070, the EU's trial database (CTIS) under its Clinical Trials Regulation (CTR), and Health Canada's Public Release of Clinical Information (PRCI). Each covers different files on a different clock. The US and UK publish registry results and, in the US, the protocol. The first job on any package is to work out which program applies, and to which pages.

The committee in the EMA row is the CHMP, EMA's medicines committee; its opinion is the recommendation on whether to authorize a medicine. CTIS sorts trials into categories: Category 1 is mainly Phase 1 and bioequivalence work, Category 2 is Phase 2 and 3, and Category 3 is Phase 4 and low-intervention trials.11

We found no standing FDA program that publishes full CSRs. FDA ran a 2018 pilot to post parts of CSRs for up to nine approved applications, with confidential and personal details redacted.16 Since 2025 it has published complete response letters.17

Policy 0070's scope grew from new active substances to most new applications in 2025

EMA suspended publication in December 2018 and restarted it for new active substances with an opinion from September 2023.7 In November 2024 it set out Step 2 for procedures with a planned opinion from the second quarter of 2025: all new applications including negative opinions and withdrawn ones, line extensions (a new strength, form or route for an authorized medicine) and major clinical variations such as new indications. Biosimilars, hybrids and generics stay out.18,4

EMA's overview page still words this as a plan "from April 2025",21 and its May 2025 guidance carries the new submission windows.1 If your internal procedure predates 2025, check its deadlines. The window for a new application used to close 10 days after the opinion and now closes at 30.4

Disclosure rules and their status: EMA's guidance version 1.5, CTIS after deferral, and Canada's 2019 regulations

In shortThe biggest recent changes: CTIS dropped deferral for applications from June 18, 2024, so sponsors now protect personal data and confidential information there by redacting. EMA widened Policy 0070 in 2025 and in April 2026 published the terms of a voluntary work-share with Health Canada. The European Data Protection Board (EDPB) published draft anonymization guidelines in July 2026, and they are still a draft.

The EU's court and data protection board are reshaping what counts as anonymous

In September 2025 the EU's top court ruled in EDPS v SRB that pseudonymized data (names swapped for codes, with the key held elsewhere) is not automatically personal data in the hands of every recipient.27 The EDPB's draft guidelines of July 2026 follow that contextual reading, and say the same dataset can be personal for one organization and anonymous for another.23

The European Commission's Digital Omnibus proposed writing a recipient-based definition into the GDPR, and the EU's data protection bodies opposed it in February 2026.28 The Council's negotiating mandate of October 21, 2026 strikes that change and keeps today's definition. It also adds a separate article saying pseudonymized data is not personal data for a third party unable to identify the person, and talks with Parliament are still ahead.28

None of this eases a public posting much, because the recipient of a public file is everybody. CTIS guidance states the point directly: pseudonymization "is not an anonymisation technique".2

What sits inside a clinical study report package, and which sections EMA publishes, removes or never receives

In shortScope comes first, and two weeks is all you get to argue about it. EMA publishes all of the CSR body, sections 1 to 15, and only three of the Appendix 16.1 documents. Patient listings in Appendix 16.2 and case report forms are out. Pages agreed as out of scope are replaced by a one-page notice in the published file.

Staff names follow a different rule from patient details

EMA's anonymization techniques apply "only to trial participants". Personal data of investigators and staff is redacted, with an exception: the sponsor and coordinating investigator who sign the CSR, and the principal or coordinating investigators who ran the trial and their sites, may stay if their roles are clear. Their contact details and handwritten signatures still come out, and the same names are redacted elsewhere in the report when they could point to a patient's location.1 CTIS is similar, and asks that a document's author be removed from file metadata too.2

Direct and indirect identifiers: how subject numbers, dates, ages and sites get redacted, recoded, offset or pooled

In shortDirect identifiers point straight at a person and normally come out. Indirect ones, like age, sex, site and dates, only identify someone in combination, and they are often what readers need. The aim is to change them as little as possible while keeping the risk under your ceiling, and to record what you did.

Three tests decide what counts as an identifier

Health Canada leans the same way as EMA on keeping information: it "encourages the generalization of indirectly-identifying variables", and says transformation should favor methods that retain analytical value over redaction.3

Leaving a variable off the list makes the score look better without making anyone safer. The 2014 opinion on anonymization from the EU's data protection regulators calls reducing the set of quasi-identifiers, its term for indirect identifiers, to raise group sizes "a common mistake".30

Re-identification risk worked through: why a group of 11 clears 0.09, and how pooling one site moves a patient under it

In shortCount how many patients share the details still readable about one person. Risk is 1 divided by that number. Public release looks at the worst single patient, so one small group sets the score for the whole package. EMA calls 0.09 advisable and lets a sponsor justify something else; Health Canada encourages the same figure.

EMA's guidance gives that rule in those terms, and says that for public release "it is advisable to use the maximum risk", the highest probability across all records.1 An average would let 400 comfortable patients carry one exposed one. On the threshold, it believes it is advisable "to set the threshold to a conservative level of 0.09", and leaves the sponsor free to choose another with a justification.1 Health Canada encourages the same 0.09 and says it "equates to a target cell size of 11 patients".3 A cell is one group of patients who look the same.

A made-up patient, three ways of describing her

Take a woman aged 67 in a study of 400. Suppose the report shows her exact age and site, and she is the only 67-year-old woman there: a group of 1, and a risk of 1.0. Band the age to 60 to 69, and suppose three women at that site fall in the band: a risk of 0.33. Pool the site to its country, and suppose 14 women in their 60s appear across that country: a risk of 0.07, under the line. Nothing about her symptom scores changed.

What re-identification tests found: six weak matches on one report, 99.98% in one model

The authors of the first study, who include staff of a drug company and of an anonymization vendor, concluded that the agencies' guidance "can provide adequate privacy protection".31 That test measured a carefully prepared report, so it says nothing about a number missed on one page. We found no public report of a patient identified from a document published under these programs; that absence is no guarantee.

Two ways to show a report is anonymized: measure the risk, or meet three criteria

A report can show anonymization in two ways: by measuring and managing risk, or by demonstrating three criteria from the 2014 opinion of EU data protection regulators: nobody can be singled out, records cannot be linked to one person, and nothing can be inferred about someone.1,30 EMA says of those criteria that for clinical reports "it is unlikely that ... all three criteria can be fulfilled by any anonymisation solution", because readers need to link one patient's records, and it steers sponsors toward a risk assessment.1

Commercially confidential information in a CSR: EMA's definition, its five rejection codes and how few claims survive

In shortProposing a confidential redaction is optional, and most proposals fail. Information must be non-public, and disclosure must undermine a legitimate economic interest, with a specific reason for each passage. Check what is already public first. Most rejections say either that the information was public already or that the reason did not hold.

What a justification has to say to survive EMA's review

1. Point at the words. Limit it to the word, figure or part of a sentence. EMA discourages proposing whole pages, sub-sections or full tables.1
2. Show you looked. Keep a list of where your product's information already appears: your own sites, EMA's website, trial registries, other regulators and the literature.1
3. Say what is new. Highlight the innovative features against what is common knowledge in that field.1
4. Name the program. Say which ongoing development program the passage relates to. "Ongoing study" is disregarded if the study has finished by publication.1
5. Explain the harm. Say how disclosure would undermine your interest. The same pasted reason on every row prompts a question from the agency at validation, its completeness check.1

Canada writes this as law. Confidential business information must be non-public, protected by reasonable measures, and of economic value because it is secret.3 Its regulations also exclude from release data not used to support the proposed conditions of use, and descriptions of tests, methods or assays used only by the manufacturer.24 A passage EMA has published cannot be secret in Canada, which is why claims need coordinating.

Policy 0070 day by day, Health Canada's 60 and 120 days, and when CTIS publishes each document

In shortAlmost none of the dates are yours. In Europe, the proposal is due within 30 days after the committee's opinion, and the 84-day review then runs to a posting 120 days after the opinion. EMA's advice is to start before the opinion. Canada gives 60 calendar days from its own start signal. In CTIS, the document you uploaded is the one that goes public, at the decision.

Trials submitted before June 18, 2024 follow separate rules. Their documents stay unpublished, whether or not a deferral was used, until a later application on that trial brings documents into scope: a substantial or non-substantial modification, or an added member state for its national documents. Check those documents for personal data and trade secrets before you file.2

Proposal copy and final copy: labels, colors, black boxes and what can leak from a redacted PDF

In shortOne set of decisions has to render two opposite ways. The proposal copy keeps the marked text legible so a reviewer can read it. The final copy removes it for good, with the rest still searchable. A rectangle drawn over text is the classic failure, and research shows even proper removal can leak through character spacing.

CCI means commercially confidential information; PPD means protected personal data. Interim reports of studies still blinded can carry temporary "BLD" redactions, with the package resubmitted once the study is unblinded.1

Headers, scans, signatures and metadata are where text searches go blind

How clinical redaction fails: missed copies, invalid packages, reports that do not match the pages, and science hidden by mistake

In shortErrors run in both directions. Under-redaction exposes a person, and that exposure can outlast any correction. Over-redaction removes what outsiders came to read. EMA's first-year review also found basic package failures and anonymization reports that did not match the redactions in the documents.

What software and AI can do to a clinical study report, what regulators allow, and why recall matters more than accuracy

In shortEMA and CTIS guidance both accept software for finding identifiers, and neither sets a required accuracy, so the sponsor owns the result. In one large 2013 test on clinical notes, the best system found about nine in ten identifiers, close to human reviewers. A 2025 test of two open tools found about three in four. That makes these tools useful and also means a person has to stay in the loop. Ask for recall, the share of real identifiers a tool finds, measured on your own documents.

Why 96% accuracy can hide a recall of 74%

Recall is the share of real identifiers a tool finds. Precision is the share of what it marks that are identifiers. Accuracy is flattered by the thousands of words that were never identifiers: a 2025 test of two open tools on 200 neurosurgical documents reported 96% accuracy for each, with recall of 74% and 79%.37

For this work the errors are not equal: a false alarm costs a reviewer seconds, and a miss can expose a person. As an illustration, a recall of 91.92% on a package with 5,000 identifier mentions would leave about 400 for someone else to catch. That is why careful teams add a second line of defense that does not depend on recognizing a word in context: once one occurrence of a number is known, look for it everywhere, in every form.

The identifier list, trade-secret calls and each redaction stay with people

Who prepares a disclosure package, where the working days go, and the in-house, vendor and software routes

In shortA disclosure lead runs the package, with reviewers, a medical writer, a statistician, legal counsel and a data protection officer each deciding part of it. In our illustrative split, most of the time goes to marking and checking. Teams work in house, through a service provider or in a mix. We found no published benchmark for any of this, so your own baseline is the one that counts.

The disclosure lead, reviewer, writer, statistician, counsel and privacy officer

Clinical redaction in house, by a service provider, or in a mix

The market is a mix of software and services, and we found no public market size we could trust. Whichever route you choose, EMA's message since 2023 is the same: settle it early, "in house or vendor chosen", and before the opinion.7

Study VLT-418-301: 6,214 proposed redactions overnight, one patient number still readable, and a risk score back under 0.09

In shortThis is the example from the top of the page, inside the product. Ines Marchetti, the clinical disclosure lead at Veltara Therapeutics, opens a package already read overnight. Five agents have proposed every redaction with its rule. She clicks one patient number, sees every place it appears, finds the one still readable, and hides it the same way. A person accepts every redaction.

1

Twelve days out, six packages on one board

It is Tuesday, October 13, 2026. Ines's board shows every package going public, by authority and deadline. VLT-418-301 and its extension study are due to EMA on October 25. A protocol amendment and consent form are due in CTIS on October 21. The same report goes to Health Canada on December 9, and that row already says it will reuse the EMA decisions.

2

Five agents read 2,140 pages between 2:10 and 5:48 a.m.

The personal-data finder reads text, tables and 312 scanned pages, and marks 4,380 spans about patients and 1,212 about investigators and staff. The confidential-information finder marks 622 redactions in 124 passages and drafts a reason for each. The rule applier matches all 6,214 proposals to a rule. The last line of the run reads "nothing is accepted without a person".

3

Fourteen identifiers were hidden in one place and readable in another

The consistency checker links 1,946 identifiers to every place they appear. Fourteen were handled differently: four broken across a line, three split across table cells, three written another way, two on scanned signature pages and two in page headers. By the morning Ines has accepted 13 fixes. One is open: 1004-0217, on page 1,806.

4

One click shows all 38 places the same patient appears

On page 412, a passage about protocol deviations, Ines clicks the redaction over subject 1004-0217. The panel shows the rule (P-01, recoded as Subject 112), the agent, its confidence of 0.99 and who accepted it. Below that, the consistency lens: 6 places in the report body, 3 in Appendix 16.1, 29 in Appendix 16.2. Thirty-seven are hidden, and the rail down the left marks the one that is not.

5

On page 1,806, the number broke across a line

Listing 16.2.6.1 shows symptom scores at Week 52. The patient is hidden in the table row, but footnote b carries her number in two halves, "1004-" and "0217", so it was read as two numbers. Left as it is, the footnote ties a row (female, 67, Site 1004, a site with 6 patients) back to her. The product's score for her reads 0.21 now and 0.06 once hidden, against 0.09.

6

Same rule, same code, all 38 places

Ines clicks "Redact here with P-01". The lens turns green: hidden the same way in all 38 places, recoded as Subject 112 everywhere. The last of the 14 inconsistencies closes. The line-break pattern reaches the personal-data finder for the next package only after a person approves it.

7

The patient is hidden and the results stay readable

One switch shows the page as the public would see it. Subjects read Subject 104 to Subject 121, ages read as bands such as 60–69, and visit dates read as study days. The stool, bleeding, endoscopy, physician assessment and total scores stay as written, so readers can check them.

8

Risk comes from accepted decisions, by site group

The anonymization report writer computes prosecutor risk (the risk if an attacker knows the person is in the study) only from accepted decisions and shows the highest value by site group against the threshold. Every group is at 0.06 or below. The report sections are drafted, with two left waiting: the justification table, and approval by Mei Fujita, the data protection officer. No agent may change a decision to lower the score.

9

Legal counsel signs the last six confidential passages

Six of the 124 passages are still waiting: a cell line and titer, purification steps and yield, a formulation buffer, an autoinjector for an unannounced use, assay reagents, and a potency bioassay. Rafael Brandt, legal counsel, confirms each one and signs. Submitting needs every redaction decided, every justification signed and the risk at or under the threshold.

The design targets are 0 patient identifiers readable at publication, 0.06 highest risk against a 0.09 threshold, and 2 days of review per study report against an illustrative baseline of 11. They are targets, to be measured on your own packages.

The marked-up page, the identifier table, the justification table, the rule sets and your settings

In shortSection 13 followed one patient number. These screens show the rest: a page with its colored marks, a confidential passage and its drafted reason, how each identifier is handled across the package, the table a regulator receives, the rule sets per authority, and the settings your team controls.

Select any screen to open it larger. Recoding the same way across packages is a setting: EMA asks for that choice to be made case by case on its effect on risk, and it allows redaction instead.1

Five agents find, rule, link and score; a disclosure lead, legal counsel and a data protection officer decide

In shortEach agent has one job and limits it does not cross. None of them accepts a redaction, signs a claim or changes a rule. The disclosure lead accepts each redaction, legal counsel signs each confidentiality reason, and the data protection officer approves the report. The unredacted original stays in a locked store and is never part of an output.

From personal-data finder to anonymization report writer: the five redaction agents

The lead accepts, counsel signs, the privacy officer approves

Every redaction records its rule, agent, confidence and who accepted it, and internal audit can read everything and change nothing. These are controls that support your compliance with 21 CFR Part 11 and EU Annex 11; no product is certified against them. Botminds' own credentials are SOC 2 Type II audited and information security managed to ISO 27001.

The lead, reviewer, writer, counsel and privacy officer in the screens

Study reports, rule sets and your confidential-information list in; redacted PDFs, a justification table and an anonymization report out

In shortThe solution reads the documents you would send anyway, the rules for each authority and your list of what is already public, and leaves behind the files each regulator asks for. Your document system, your publishing tool and your people stay where they are. One set of decisions renders as both the marked-up proposal and the final copy.

Systems a disclosure rollout connects to

For each, check whether it has an API, a file feed, or only screens, and who owns it.

Your method, your counsel and your submission stay yours

What a page-by-page redaction costs you in reviewer days, vendor weeks and a patient's privacy

In shortStart from your own costs, because the numbers that matter are yours. A few published figures frame them: how fast the volume is growing, how short the last clock is, and how large the legal exposure can be.

Timing one past package to find your CSR redaction baseline

A patient's privacy, downloaded copies and hidden science: redaction costs beyond the invoice

What faster page review returns when you count reviewer days against the tool's yearly cost

In shortPut in your own numbers. The calculator asks how many study report packages you handle, the days they take now and after, how many people work on each, a loaded hourly cost, and what the tool and your testing cost a year. Every default is an assumption to replace, and days saved only count if you decide what they are for.

Reading, linking and reusing: where redaction work saves

Packages a year, days per package and your hourly cost

A sponsor with a steady pipeline and several authorities can reach the first two bars. A sponsor with one or two study reports a year probably cannot on hours alone, and should admit that before buying anything. For it, a service provider may cost less, and the argument for software, if there is one, is control over a single package that cannot go wrong.

How ready your transparency team is: package volume, a written method, source access and named signers

In shortPick the answer closest to your team today. Eight questions score 0 to 2; the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

A low score means writing down your anonymization method first

What to ask redaction software or a service provider about recall, consistency, what stays in the file and who decides

In shortAsk every vendor the same questions, us included. The answers show how a redaction is made: what is read, how each copy of a number is found, what is left in the file, who decides, and how you would measure all of it on your own documents.

We found no published independent test of any CSR redaction vendor's accuracy, as of November 2026. The July 2026 report on EMA's tool selection, covered in section 10, does not rank vendors for sponsors.34

Test redaction on a closed package with planted identifiers, then decide on recall, leaks and reviewer days

In shortMeasure first, then test on a package you have already finished, with identifiers you planted, then run alongside a live one. None of these steps commits you to a vendor.

1. Search one final package yourself: take 20 subject numbers and look for every form of each. You will learn something from it whichever way you go.
2. Time your last package: days per person, vendor cost, agency comments.
3. Pick a closed package and get approval to use it. Plant known identifiers: numbers split across lines and cells, other formats, a name in a header, a signature on a scan.
4. Run alongside a live package for one cycle, with your lead accepting and your counsel signing as usual.
5. Decide with three numbers: recall by identifier class, readable copies left, and reviewer days to a clean proposal.

Recall by identifier class, page-level leaks and reviewer minutes per 100 pages

Clinical trial redaction questions answered: Policy 0070, CTIS, Health Canada, 0.09 and AI tools

In shortQuestions disclosure teams and buyers search for, each answered on its own.