Medical writing quality control, explained.

Every results document repeats numbers from a table, and someone has to check each one

From statistical tables to a signed QC record, in five steps

1. Statistics delivers the results tablesVLT-302's final run lands on October 2: 412 outputs under the statistical analysis plan, the document that fixes the analyses in advance.
2. Writers restate them in documentsReport sections, synopsis, abstract, manuscript, patient summary, brochure update.
3. QC checks every number against its cellValue, patients, analysis, table version and format, recorded with both locations.
4. Writers correct, statisticians answerAccepted fixes go to the author; "which number is right?" goes to the biostatistician.
5. The QC lead signs the recordA named person states what was checked, against which run, with each decision kept.

Step 3 is the subject of this guide. The study, people and figures in this example are fictional.

Ticking 214 numbers, invisible errors and two more QC surprises

TLF, final run, stale quote and denominator defined

Fourteen terms are defined in the one-page summary.

What a CSR QC reviewer tests on each number: value, patients, analysis, run and format

In shortQC of a results document asks five questions of every number. Only the last concerns style. A proofreading pass reads the document against itself, and these checks need the tables open beside it, which is why they take days and why spelling software cannot do them.

The European Medical Writers Association's journal defines QC as ensuring "a document's content, style, and format are of high quality", and adds that "a good QC review takes time".4 One experienced specialist describes checking "against source data" and "for consistency", and warns that proofreading software "will not identify issues with analytical statements or incorrect data".2

A published walkthrough from a contract research organization, or CRO, a company hired to run trial work, gives a working order: format, then table numbers against their sources, then numbers in the text, then conclusions and the synopsis against the body, then language and references. Each number error goes back to the author as a comment with the corrected value, and a signed checklist closes the job.5

Per-protocol means the patients who followed the protocol closely enough. Intent-to-treat means everyone randomized, or as close as possible. Section 05 explains both.

Proofreading, QC, quality assurance and audit: four jobs people confuse

ICH E6(R3) gives quality assurance and quality control separate subsections.3

Ask a QC reviewer to "do the QA" and you have asked for the wrong thing. Show an inspector an assurance procedure when they ask for QC records and you have answered a different question.

Derived numbers need their own step. ICH E3 says "estimated or derived values, if used, should be identified in a conspicuous fashion" in tables.6 In text, "about 10 times more likely" or a difference between two groups usually has no cell to tick, so the reviewer rebuilds it from the cells it came from. Section 06 works one through.

No ICH guideline says "check 100% of CSR numbers": what E3, E6(R3), E9 and Part 11 do ask

In shortRegulators hold the sponsor responsible for accurate results and documented QC of the outputs. The exact depth of a text-to-table check is your own procedure, and it must be documented. E6(R3) is the guideline that changed most recently, and it makes outputs traceable to their programs and runs.

First, the names. ICH, the International Council for Harmonisation, writes guidelines that regulators such as the US Food and Drug Administration (FDA) and the European Medicines Agency (EMA) then publish as their own: E3 on the structure of a study report, E6 on good clinical practice and E9 on statistical principles. 21 CFR Part 11 is FDA's rule on electronic records and signatures. The sponsor is the company responsible for the trial.

From CSR section to patient summary: the documents that quote one results table

In shortThe CSR is the largest document that quotes the tables, and rarely the first one out the door. A synopsis, abstract, manuscript, lay summary, brochure and registry entry repeat the same results, under different rules for rounding and patient groups, on different deadlines.

ICH E3 lays out the report at different levels of detail: overall summary tables "may be placed in the text to illustrate important points", other tables go in section 14, and patient-level listings in appendix 16.2.6 Every number in the text is therefore a restatement, and most of the report's pages are outputs. Across 78 industry reports studied by Doshi and Jefferson, a non-random sample, the efficacy text had a median of 13.5 pages while the tables behind the report ran to 337.16

Safety analysis set or randomized patients: how one denominator changes a CSR percentage

In shortA percentage is a claim about a group of patients. A trial usually carries at least three groups: everyone randomized, those who kept to the protocol, and those who took the drug. Divide by the wrong one and the digits still look right, which is why this error survives reading.

Why 223 patients give 52.1% in one sentence and 52.2% in another

In the fictional study, 428 patients were randomized to Veltrimab and 427 took at least one dose. In the safety table, 223 of them had a treatment-emergent adverse event, one that started or worsened during treatment. A topline memo says 52.1%. The table says 52.2%. Work it through: 223 ÷ 428 = 52.10%, and 223 ÷ 427 = 52.22%. The memo divided by the wrong group.

A reviewer who sees a tenth of a point assumes rounding, and is usually right. That is how the rare case of the wrong patient group gets through. Published errors follow the same shape: when one research team tried to reproduce numbers taken from trials into meta-analyses, studies that pool many trials, the common problems were an "erroneous number of patients, means, standard deviations, and sign for the effect estimate".19

The right number from the wrong analysis set

Population errors can also leave the digits untouched. A sentence on page 48 of the fictional report says the per-protocol set reached PASI 100, complete skin clearance, at 27.6%. That number exists: Table 14.2.3.1 shows 118 of 428 in the intent-to-treat set, 27.6%. The per-protocol figure, in Table 14.2.3.2, is 116 of 404, or 28.7%. Two fixes are possible, changing the number or changing the words, and the choice depends on what the writer meant to claim. A person with the analysis plan has to choose.

The abstract has the same problem in miniature. It says 640 patients were randomized. That is the safety set across both arms. The randomized count is 642.

Rounding a percentage twice, SAS and R ties and "about 10 times": the arithmetic of a CSR check

In shortRound once, from the table's unrounded value, by a rule written down. Rounding twice can change a digit. Software disagrees about ties, so programmers and writers must use the same rule. Differences and ratios often have no cell, so they are rebuilt by hand.

How rounding 4.449 twice turns a correct 4.4% into a wrong 4.5%

A safety table shows 19 of 427 patients with an event, printed as 4.4%. The draft says 4.5%, because someone rounded twice, in two reasonable-looking steps.

Why SAS ROUND and R round can disagree on a percentage ending in 5

SAS and R, two programming languages used for trial statistics, treat a tie, a value exactly halfway between two rounded values, differently. SAS's ROUND function uses what its blog calls the traditional method, which "rounds the midpoint between two integers away from zero", so 2.5 becomes 3; SAS also offers ROUNDE, which rounds to even.20 R's round is documented to follow IEC 60559, "go to the even digit", so 2.5 becomes 2 and 0.5 becomes 0, and the same page warns that stored values matter: round(0.15, 1) "could be either 0.1 or 0.2".21

A CRO program in one language and a sponsor check in another can therefore print different digits from the same data. Clinical programmers also report comparison differences "on numbers that appear identical", because of how computers store decimals.22 A QC reviewer needs one answer: which rule did the table use? Write it in the house rules.

Rebuilding differences, ratios and "67 out of 100" from their cells

Three kinds of number need extra care, because the figure in the sentence is more than a straight copy of one cell.

The worked formula a reviewer applies to any percentage is short. Find the count and the denominator in the cell. Divide. Round once, by the house rule. Compare with the sentence, and compare the denominator with the group the sentence names.

Re-run TLF impact on a CSR: one changed cell, nine quotes and why a sample misses them

In shortA late re-run is the dangerous case. The old value lives on in documents nobody plans to reopen, and no reading of those documents can reveal it. Sampling numbers fails here, because a stale quote sits in particular cells. Only a list of every quote of that cell answers the question.

Runs happen for good reasons. A data query corrects a reading, QC finds a programming error, a reviewer asks for a change. ICH E6(R3) wants outputs "dated and time stamped", so the run is knowable.3 The hard part is the other direction: from a changed cell to every sentence that quotes it.

The UK inspectorate described the opposite of good practice in a critical finding at a contract research organization: multiple runs of a statistical analysis, no maintained records of why they had taken place, and a p-value for the primary endpoint, the trial's main measure, that "changed throughout these runs".23

A dry run is a rehearsal before database lock, the point when trial data are frozen for analysis. In blinded trials, where the team does not know who received what, it commonly uses dummy treatment codes: one published analysis plan marks its drafts "DRAFT/DUMMY TREATMENT ASSIGNMENTS APPLIED",24 and programmers preparing data monitoring packages work on dummy treatment too.25 Numbers from such a run should never reach a sentence as results.

What one data query did to VLT-302's documents

Why a 20% QC sample almost never finds all nine stale quotes

Many procedures check a sample of numbers and accept the rest. For random typing errors, that gives an estimate of quality. For a re-run it fails, because the nine stale quotes are specific. If each number has the same chance of being sampled, the expected share found is the sample rate, and the chance of finding all nine is that rate multiplied by itself nine times. That is arithmetic, and we have not measured it on real QC.

That does not make every procedure a bad one. ICH asks for risk-based QC, and a sampled pass with a written reason can fit that.3 It means a re-run deserves its own step: after every run, list what changed and find every quote of those cells. Whether you do that by searching files, by keeping a table of quotes, or with software, the record should show it was done.

How often a restated number is wrong: what journals and registries show when someone checks

In shortWe found no published error rate for clinical study reports, or for how many errors a QC pass catches. Studies of similar acts, restating a result from a source, find disagreement in a fifth to about seven in ten documents, depending on what was measured. Treat these as what happens elsewhere, and measure your own.

Why reading harder does not fix a number check

Human checking by eye is weak in a way people do not expect. In one experiment, 195 undergraduates entered data by one of three methods. Those who checked their work by eye made 2,958% more errors than those using double entry, about 31 times as many, and did no significantly better than people who did not check at all.33 The participants were students entering data sheets, so read it as a warning about the method and nothing more.

It fits what QC specialists say. "It is extremely easy to add incorrect data to tables when working with source tables that are over a hundred pages long".2

"Deficient quality control of the CSR": what EMA, MHRA and FDA have written about reports and QC

In shortRegulators rarely report a wrong percentage as such. They report weak QC, thin records and reports that say what was planned in place of what happened. The risk runs from a finding at inspection to a refusal to approve, while one reviewer survey suggests data errors are already less common than wordiness.

The European Medicines Agency's inspectors publish a yearly list. For 2024, across 67 good clinical practice (GCP) inspections requested by its medicines committee, they recorded 335 critical and major findings, 11 of them critical. The clinical study report category holds 5, with 1 critical. Examples include "Deficient quality control of the CSR" and "Incomplete or inaccurate reporting of protocol deviations and key data in the CSR".34

When an inspector cannot tell how much QC was done

The UK's Medicines and Healthcare products Regulatory Agency, the MHRA, reported on 36 GCP inspections for April 2019 to March 2020. One critical finding, at a contract research organization, concerned a response hard coded, written directly into the program in place of coming from the data, into part of a primary endpoint, with "insufficient documentation available to verify the level of QC of data which had been hard coded".23 QC that left no record could not be told apart from QC that did not happen.

When a CSR states what was planned in place of what happened

An FDA warning letter of November 27, 2024 faulted a sponsor whose report "reported dose levels for subjects as stated in the protocol", when mislabeled drug meant 80% of the protocol-required dose was given between March and June 2021.35 A table-to-text check would not catch that by itself, since the text matched the plan. It shows why the check needs data, and why a regulator treats it seriously: FDA may refuse to approve an application containing "an untrue statement of a material fact".14

What regulatory reviewers told a survey about document quality

Keep that in proportion. The same survey found length, repetition and lack of clarity reported most often, and data errors "ranked relatively low", which the authors read as a sign that "many applicants have implemented processes to catch these avoidable issues".36 QC is one such process: the checks may be working, and they still take reviewer days.

How long CSR QC takes, how much of it the industry does and what a late day is worth

In shortThe only public stage benchmark we found gives QA on a final CSR a mean of 6.6 working days, ranging from 1 to 30. No public figure exists for QC of an abstract or synopsis, or for what QC costs. Registry counts of completed late-stage trials show no rise since 2022, and AI drafting may add more text to check.

Hamilton's survey of participants at eight European medical writing conferences, 2008 to 2013, asked for typical durations on a moderately complex Phase 3 study of 200 to 400 subjects. Quality assurance on the final integrated report had a mean of 6.6 working days, with a range of 1 to 30, from 77 responses. The paper does not define that stage, so read it as the nearest public stand-in for a QC pass. It concludes there are "no industry standard durations".26

One QC specialist adds a warning: "It is not realistic for a consistency and data check to be performed in one day on a 100 page clinical study report!"2

How many CSRs and restated documents the industry produces

There is no public count of CSRs. A rough stand-in is the number of late-stage industry trials recorded as completed each year, read from ClinicalTrials.gov.39 FDA's drug center approved 55, 50 and 46 novel drugs in 2023, 2024 and 2025.40,41,42 The market for all medical writing is estimated at $3.94 billion to $5.43 billion for 2025, by two research firms whose public listings disagree and show no dataset;43,44 we found no figure for QC alone.

The trend pushing more text toward the QC desk

Generative drafting could put more text in front of reviewers. A 2026 paper from one drug company found AI-drafted sections on pharmacokinetics, how the body handles a drug, scored about 90% on reporting quality relative to manually written reports, in two blinded review sessions; every author is a current or former employee of that company.45 Whether that holds elsewhere is unknown, and we found no public figure for what it does to QC workload.

On the data side, the Analysis Results Standard from CDISC, the clinical data standards body (v1.0, April 2024), aims to deliver results as data that a computer can read,46 which would make matching a number to its result far simpler once tables stop being PDF and RTF files.

AI in medical writing QC: what FDA, EMA and the EU say in 2026, and what must stay with a person

In shortWe found no rule that forbids AI in checking a document, and none that excuses the sponsor from what it submits. FDA's guidance is still a draft and leaves out some writing uses. EMA expects close human supervision of generated text. Finding and comparing can be automated; deciding which number is right and signing should stay with a person.

FDA's Computer Software Assurance guidance, final February 2026, covers software in medical device production and quality systems,56 so it does not apply to this work either.

Which QC tasks suit automation, and which need a named person

The limits of automated CSR number checking

A checker's confidence is a score, and a sponsor still needs evidence. Published evidence on automating this kind of work is thin: a preprint review of statistical programming found quantitative effectiveness data in only 12 of 527 validation papers, or 2.3%.1 Three limits apply to any tool, rule-based or AI.

How medical writing QC runs today: reviewer, writer, statistician and the hand-offs between them

In shortA writer finishes a draft and sends it to a reviewer who did not write it, usually near the end. The reviewer works through outputs and text, comments go back, and a statistician answers questions about numbers. Teams staff this in house, through a CRO, with scripted checks or with software, and by our estimate about half the time goes to finding and ticking.

QC lead to publications manager: who handles a results document

Six hand-offs from a draft to a signed QC record

Where the days go when one CSR section is checked by hand

One sum shows why teams feel this. At an illustrative three minutes a number, 214 numbers take about 10.7 hours of ticking. That is close to the 1.5 days above, and it holds only if nobody is interrupted.

In-house reviewers, outsourced QC, scripted checks or QC software

The VLT-302 re-run: one changed results table, nine old quotes found, and a fix on the writer's list

In shortThis is the example from the top of the page, inside the product. AI agents, each a program with one job, trace every number to its table cell. Lena Ortiz, the QC lead, opens a draft with the agreements already settled. When a table changes, one click draws a line to every quote of the changed cells, and she decides each finding. The agents never edit a draft.

1

The QC desk: every draft a page, every number a dot

Lena's home screen shows the work waiting. Eight drafts sit on the desk as page thumbnails. Each dot is one number, colored by verdict: agrees with its cell, derived and recalculated, rounding or format, mismatch, no source, or decided.

The results package runs across the top: VLT-302's final run of October 2, with 412 outputs. Table 14.2.2.1 stands out in red. Beside each draft sit its due date and the count to decide: 9 on the efficacy section, 3 on the abstract due October 9.

2

Show the ripple: from two changed cells to nine quotes

Data query DQ-1187 at site 114 re-scored one Week 16 reading before lock. Responders went from 285 (66.6) to 284 (66.4), and the difference from 61.4 to 61.2.

Lena clicks "Show the ripple". Lines fan out from the table to the exact dots in six pages. Red lines reach quotes still open, gray lines reach those already settled, and drafts it does not touch fade back. This is the question nobody could answer quickly by hand.

3

Re-check the nine quotes: six open, three settled

She clicks "Re-check the 9 quotes" and each step shows as it runs. The two runs are compared: 2 of the table's 9 cells changed. Every quote of those cells is found, including the abstract's rounded 67% and the patient summary's "67". Each is re-checked against the final run.

Six still disagree. Three, in the synopsis and brochure update, are already settled, with corrections sent. The writers are notified.

4

Page 47: the sentence, the reasoning and the cell, side by side

Lena opens 66.6% in Section 11. The draft sits on the left and the table on the right, with a line joining the sentence to the cell: Table 14.2.2.1, Veltrimab column, 284 (66.4).

Between them the rule checker explains itself. Eight small checks pass or fail: output found, endpoint and visit, analysis, current run, patients and N, arm and column, value, rounding and format. Two fail here: current run and value. It suggests 66.4% and shows how sure the agents are: 0.98.

5

Where the old number came from, and where it goes

One toggle shows the superseded run, stamped as such, with 285 (66.6) in the cell the draft quotes. The verdict is always taken against the final run; the old one only explains the mistake.

Lena accepts the fix. The finding records her name and the time, the dot turns to decided, and the correction goes on the list of the lead medical writer, Dr. Maya Chen. She changes the sentence in her own document.

What the sample's nine documents add up to

Wrong analysis set, safety denominator, rounding twice: the sample's other CSR QC findings

In shortMost findings have nothing to do with the re-run. Here each error type from Part A appears on the screen with its reasoning. Every finding cites the draft's page and the table cell, and a person chooses what to do with it. The agents never choose between two conflicting numbers.

Which kinds of error the sample's week of QC turned up

Documents, the results package, a table's quotes, the QC dashboard and your rounding settings

In shortThe story used the desk and the check screen. These are the rest: all documents with their verdict bars, all findings with a comment log, every output and how often it is quoted, a dashboard of what went wrong, and the settings that hold your house rules.

Five QC agents trace every number, and a named QC lead decides each finding and signs the record

In shortEach agent has one job and limits it does not cross. They find, match, check and report. A person decides each finding: accept the fix, ask the statistician, remove the sentence or keep it with a reason. The QC lead cannot sign with a finding open. The solution reads the drafts and tables you already produce and leaves a QC report, a comment log, corrections per writer and a signed record.

Number finder to QC report writer: the five medical writing QC agents

A QC lead, a writer, a statistician and a publications manager in the VLT-302 screens

Signing the QC record: a topline memo with a site count no table holds

A topline memo from Omar Haddad, the lead biostatistician, arrives, and Lena checks it against the final run. Of 31 numbers, two are findings. "87 sites" has no output: the table matcher searched all 412, and site counts live in listing 16.1.4, which has not been loaded. The other, any treatment-emergent event at 52.1%, divides by those randomized.

She marks the site sentence for removal and accepts the fix. With every finding decided, she signs. The stated meaning is "QC review complete — every number verified against the final tables", and she enters her password again.

Controls that support your Part 11 and GCP obligations

These are controls that support your compliance. Validation and the decision to rely on them stay with you.13 Botminds' security credentials are SOC 2 Type II audited and information security managed to ISO 27001.

From draft documents and TLFs to a QC report, comment log and signed record

Systems a medical writing QC rollout connects to

The two design targets for medical writing QC, and how to read them

The design covers roles and separation of duties, maker-checker sign-off, an audit trail you can export, connections to the systems above, and scale across studies and CROs.

What ticking numbers and late corrections cost your medical writing team

In shortStart from your own records. We found no published figure for what a QC pass costs, how many errors it catches or what a correction costs, so your QC logs and timesheets hold the numbers that matter. A correction's labor is usually small. The re-checking, the records and any lost day are what add up.

Pulling a QC baseline from your last ten results documents

The QC costs that no timesheet shows

What automating number checks returns once you count tracing, findings and re-runs

In shortPut in your own numbers. The calculator values two things you can measure: minutes per number, and hours per re-run. Labor alone pays only above a certain volume. A day off a filing can be worth far more, so it sits on its own line and stays out of the total.

Ticking, re-run checks and statistician time: where QC hours come back

Work out your CSR QC return from documents a year, minutes per number and re-runs

Change the inputs and the answer changes with them. A team running twenty documents a year on a small portfolio should expect hours alone to fall well short of any software cost. Such a team might do better with a written re-run step and a rule on rounding, which cost only staff time. A larger group, or a CRO running QC as a service, reaches the numbers faster.

How ready your medical writing group is for automated QC: run control, house rules and the sign-off rule

In shortPick the answer closest to your team today. Eight questions score 0 to 2, and 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 usually means writing down your QC rules first

What to ask a CSR QC software or medical writing QC vendor about cells, runs and sign-off

In shortAsk every vendor the same questions, us included. The answers show whether the tool reads tables or only words, what it does with a re-run, who decides a disagreement, and how you would measure it on your own study.

Plant errors in a finished CSR section: a six-step test for any QC process or tool

In shortPull your baseline, plant errors in a copy of a finished document, then replay a past re-run and run live documents in parallel. Decide with three numbers. None of these steps commits you to a vendor, and the same test measures your current process.

1. Time and type ten past documents: hours, days, numbers quoted, and findings by type.
2. Choose a finished study that had at least one re-run, and freeze each delivery as it left statistics.
3. Plant known errors in a copy: a wrong denominator, a swapped column, a stale value, a sensitivity result quoted as primary, a number rounded twice, a sentence with no source. Keep clean copies too.
4. Replay the re-run. Can the process list every quote, including rounded and worded ones like "67 out of 100"?
5. Run live documents in parallel, with your people deciding and signing everything.
6. Decide with three numbers: catch rate by error type, false alarms on clean copies, and hours to a signed record.

Which medical writing QC documents to automate first

Catch rate, false alarms and reviewer hours for each planted error

Medical writing QC questions answered: CSR checks, denominators, re-runs, rounding and AI

In shortCommon questions from writers, reviewers and buyers, each answered on its own.