Medical writing quality control that traces every number to its table cell, so reviewers see only the mismatches
Every number traced to its table cell; reviewers see only mismatches
See one case, screen by screen ↓Software that checks a results document — a CSR section, synopsis, abstract, manuscript, lay summary or IB update — against the statistical outputs it quotes. Medical Writing Quality Control uses five agents to trace every number to its TLF cell, check it against your house rules and the other drafts, and raise only the numbers that disagree, for the QC lead to decide and sign.
How to choose AI solutions for life sciences →
Updated
Every CSR section, synopsis, abstract, manuscript, lay summary and IB update restates results from the TLFs. Before it goes out, a QC reviewer ticks each number against its table cell by hand — and checks that it comes from the analysis set the sentence names, uses the right denominator, is the primary analysis and is rounded the house way. One efficacy section can quote more than 200 numbers, and most of them already agree.
The dangerous case is the late re-run. A data query before lock changes one cell, the statistician re-runs the output, and the old value lives on in documents nobody re-checks: the synopsis now contradicts Section 14, the congress abstract quotes dry run 2, the plain-language summary rounds a number that no longer exists. Nobody can say, quickly, which drafts still quote the old cell.
Estimated split for one CSR section, by hand and with the solution. The totals — 4.5 working days by hand, 1.7 with the agents — are the medians the working solution's dashboard shows for a CSR section.
Agents find every number, match it to its table cell, check house rules and the other documents, and trace every quote when a table is re-run; the QC lead decides each finding and signs.
VLT-302 — Veltrimab 150 mg every two weeks against placebo in plaque psoriasis, 642 patients randomised. The final run is loaded on Oct 2, and one sPGA table has changed since dry run 2. Here is what happens next, screen by screen, in the working solution.
Lena's home is the QC desk: 8 drafts on the desk, 1,150 numbers traced to their table cells, 18 to decide. Each draft is a page — CSR §11 Efficacy with 214 numbers and 9 open, the congress abstract due Oct 9 with 3 — and each number is a dot coloured by verdict: agrees with its cell, derived and recalculated, rounding or format, mismatch, no source, decided.
“Each dot is one number · hover a large dot to see it”
Above the pages runs the results package: VLT-302 final run, Oct 2, 412 outputs. Table 14.2.2.1 is marked in red: data query DQ-1187 at site 114 re-scored one Week 16 sPGA reading before lock, so Veltrimab responders went from 285 (66.6) to 284 (66.4) and the difference from 61.4 to 61.2. “Show the ripple” draws a line from the table to each quote of the changed cells: 9 quotes in 6 documents.
“These drafts still quote dry run 2.”
The table matcher compares the two runs — 2 of 9 cells changed — finds every quote of the changed cells and re-checks each against the final run. The CSR Synopsis and IB §5.3 quotes are already settled, their corrections sent; CSR §11, the congress abstract, the manuscript and the plain-language summary still quote dry run 2. The writers — Dr. Maya Chen, Priya Raman and Dana Okafor — are notified.
Each step is shown as it runs, with what it found.
Lena opens 66.6% → 66.4% in CSR §11. The draft sits on the left, the TLF on the right, and a thread runs from the sentence to the cell — Table 14.2.2.1, Veltrimab column, 284 (66.4). Between them the rule checker explains the finding, the suggested fix and how sure the three agents are: 0.98.
Rule checker: “The draft quotes dry run 2 (Sep 12). In the final run (Oct 2) one Week 16 sPGA reading at site 114 was corrected after data query DQ-1187, so Veltrimab responders fell from 285 to 284.”
One toggle switches the table to dry run 2 of Sep 12, stamped superseded, with 285 (66.6) in the cell the draft quotes and the footnote “DRY RUN — NOT FOR REPORTING. Data before database lock.” The verdict is always taken against the final run; the superseded run only shows where a stale number came from.
The table matcher never matches against a superseded run without flagging it.
Lena accepts the fix. The finding records who decided and when, the dot turns to decided, and the correction goes on Dr. Maya Chen's corrections list. The agents never edit the draft — the writer makes the change in her own document.
“Fix accepted — 66.6% → 66.4% · on the corrections list for Dr. Maya Chen”
Not every finding is a re-run. The sentence says per-protocol set, but 27.6% is the ITT result for PASI 100 in Table 14.2.3.1; the per-protocol value is 28.7% (116/404) in Table 14.2.3.2. On page 47 the confidence interval 55.4, 67.0 comes from the unstratified sensitivity analysis — the primary analysis gives 55.9, 66.5.
“Change the value — or the population and the table reference.”
The findings list shows every number that disagrees with its table, across all documents: the draft value against the table value, the type — stale table version, wrong denominator, rounding, format, wrong population, different analysis, no source, wrong column, count differs — the source output and the status. From here Lena exports the comment log and sends the corrections to the writers.
Omar's topline results memo v2 arrives and Lena checks it against the VLT-302 final run under the house rules for CSRs, memos and regulatory documents. Of 31 numbers, two are findings. “87 sites” has no output: the table matcher searched all 412 outputs, and site counts live in listing 16.1.4, which has not been loaded. Any TEAE at 52.1% divides by the randomised N instead of the safety set — 52.2%.
“Load listing 16.1.4 to check the site count, or remove the sentence.”
Lena marks the site sentence for removal and accepts the safety-set fix. With every finding decided, she signs the QC record: 31 numbers checked, 2 findings decided, against the final run of Oct 2. The meaning of her signature is “QC review complete — every number verified against the final tables”, and she re-authenticates with her password.
“The QC record — every number, both locations and each decision — is kept with the document.”
The same re-run, seen by the four people who carry it — what QC week looked like, and what it looks like now.
Agents find every number, match it to its table cell, check house rules and the other documents, and trace every quote when a table is re-run; the QC lead decides each finding and signs.
Reads the draft section by section, in-text tables included, and records every number with its statistic, endpoint, arm, visit, the analysis population named or implied by the sentence, its page and sentence. It marks derived numbers — differences, ratios, “about N times” — and the cells they are built from.
Finds the TLF output whose title, population, endpoint and visit fit the sentence, then the row and column for the arm and statistic in the current run. When an output is re-run it compares the runs and finds every quote of the changed cells. If nothing fits, it says which outputs it searched and the nearest candidates.
Compares each number with its cell: population and denominator (the safety set for safety data), analysis (primary unless the sentence names a sensitivity analysis), arm and visit, then the house rules for rounding and format. It recalculates every derived number and gives a verdict, a type, a reason a writer understands and a suggested fix.
Checks that the synopsis, abstract, manuscript and summaries agree with each other and with documents already signed.
Writes the number-by-number QC report, the comment log and the corrections memo for each writer.
Decides each finding and signs. The agents propose; a named person decides.
Ask in plain words — which numbers disagree, where a table is quoted, why a finding was raised — or tell it which house rule to add. Every answer points to the sentence and the cell.
Where is Table 14.2.2.1 quoted, and what changed?
The final run (Oct 2) changed 2 cells: Veltrimab responders 285 (66.6) → 284 (66.4) and the difference 61.4 → 61.2 (data query DQ-1187, site 114). Those cells are quoted 9 times in 6 documents with the old values — CSR §11, the abstract, the manuscript, the lay summary, the synopsis and IB §5.3. 6 are still open; the synopsis and IB corrections are already with the writers.
Why does the safety set matter?
Safety percentages use the safety set — the 427 Veltrimab patients who were treated — not the 428 randomised. Dividing by 428 gives 52.1% for any TEAE instead of 52.2%, and 12.1% for nasopharyngitis in the manuscript instead of 12.2%.
What is a number with no source?
A number no output reports. CSR §11 says 88% of responders kept their response through Week 52 — but the maintenance period is still blinded and the package has no Week 52 table. The agents searched all 412 outputs; the nearest stops at Week 16. Remove the sentence, or keep it with a reason until the addendum.
Add a rule: “about N times” must be within 5% of the table ratio
Done. New house rule 7: “about N times” must be within 5% of the ratio of the table values. The rule checker re-checked the derived ratios on the desk: CSR §11 says “about 10 times”, and 72.7 ÷ 7.0 = 10.4 — 3.8% away, so it stays derived · verified. The rule is live in Settings and on the audit trail.
13 screens from the working solution, on its sample data. Pick one to see it large.
What counts as a match, how numbers are rounded and who decides are settings, alongside the house rules the rule checker applies.
| Setting | Default | Choose from |
|---|---|---|
| Results package | VLT-302 final run · Oct 2 | VLT-302 final run · Oct 2 · VLT-302 dry run 2 · Sep 12 |
| Flag numbers that match a superseded run | On | On · Off — shows where a stale number came from |
| Rounding | Half away from zero (SAS ROUND) | Half away from zero (SAS ROUND) · Half to even |
| Percentages in the body | One decimal (72.7%) | One decimal (72.7%) · Whole numbers (73%) |
| Small p-values | p<0.001 | p<0.001 · p<.001 |
| Treat format-only differences as notes | Off | On · Off — when on, rounding and format items no longer block sign-off |
| QC sign-off | Lena Ortiz | Lena Ortiz · Dana Okafor |
| Statistician queries go to | Omar Haddad | Omar Haddad · Sam Patel |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample study data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (manual QC of 2–5 days per document; LTIMindtree, Role of Intelligent Automation in Transforming the CSR QC Process, 2020; ICH E3 on where results sit in a clinical study report). People, companies, studies and products named on this page — Northwind Bio, Veltrimab, VLT-302 — are fictional: characters and sample data in the working solution.
Software that checks a results document — a CSR section, synopsis, abstract, manuscript, lay summary or IB update — against the statistical outputs it quotes. Medical Writing Quality Control uses five agents to trace every number to its TLF cell, check it against your house rules and the other drafts, and raise only the numbers that disagree, for the QC lead to decide and sign.
The number finder records each number with its statistic, endpoint, arm, visit and population. The table matcher finds the output, row and column in the current run, and the rule checker compares value, population, denominator, analysis and rounding, recalculating derived numbers such as differences and “about N times”. Each finding cites the draft page and the table cell.
The table matcher compares the new run with the earlier one, finds every quote of the changed cells across all drafts and re-checks them against the final run. In the working solution one re-run sPGA table showed 9 quotes in 6 documents still carrying dry-run values; 3 were already settled and 6 went to the writers.
Yes. Safety percentages must use the safety set, efficacy percentages the analysis set named in the sentence, and confidence intervals the same output as the estimate. Values are rounded once from the table, half away from zero by default, with one decimal in CSRs and whole numbers allowed in abstracts and lay summaries.
No. The agents never edit a draft or a signed document. Accepted fixes go to the document's author as corrections, and the writer makes the change in their own document.
Yes. Only a person can resolve a finding — accept the fix, ask the statistician, remove the sentence or keep it with a reason — and the agents never choose which of two conflicting numbers is right. The QC lead signs the QC record with an electronic signature and its meaning, and cannot sign while a finding is still open.
Yes. The results package, rounding method, decimal places, p-value format, whether format-only differences count as findings, the house rules themselves, the QC signer and the statistician for queries are all settings. New house rules can be added in plain words, and the rule checker applies them straight away.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your house style, a results package and a sample of past drafts — and it goes live once every quality gate has passed. We will walk you through it on your own documents first.
A checked first-draft protocol in days, in your approved language.
1-min demoeTMF Quality CheckAn inspection-ready trial master file every day, each site scored on what is due today.
1-min demoClinical Study Report AuthoringResults drafted within days of database lock, linked to tables.
1-min demoWe'll run Medical Writing Quality Control on a past CSR section and its TLFs of your own.