CSR patient narratives, explained.

What a CSR patient narrative is: one participant's serious event, told in order

One participant's narrative, from a clinic visit to section 14.3.3

1. The site records what happensA doctor in Lyon enters a dry cough on a case report form, the study's data sheet for each participant, with a date. The form goes into the clinical database.
2. The plan says who needs a narrativeA narrative plan lists the triggers, the events that call for a narrative. In DVZ-301, 247 of 612 participants meet at least one.
3. A writer turns her records into a paragraphDoses, event, tests, treatment, outcome and opinions, timed in study days. This is where a date that moved gets caught, or carried forward.
4. A physician reviews and signsEvery fact is checked against the data. In DVZ-301, deaths take a second signature, under the sponsor's own rule.
5. It goes into the reportThe narratives are compiled, usually in section 14.3.3, and the report goes to regulators. Some of them publish it.

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

The word "brief", and three more surprises about narratives

Narrative, SDTM transfer, study day, database lock

Fourteen terms are defined in the one-page summary.

ICH E3 section 12.3.2: the fourteen things a patient narrative should cover

In shortICH E3 section 12.3.2 names eight things a narrative should generally describe and six more it should include. That makes fourteen elements. Treat them as a checklist and writing becomes a mapping job: each element needs a sentence, and each sentence needs a record.

The clause is short enough to read in full. "There should be brief narratives describing each death, each other serious adverse event, and those of the other significant adverse events that are judged to be of special interest because of clinical importance."1 The narratives "can be placed either in the text of the report or in section 14.3.3, depending on their number". One peer-reviewed account says teams use the text when there are 6 to 10 and section 14.3.3 otherwise, and that each runs to one to four pages.6

Each ICH E3 element, with a sentence from one narrative

Which deaths, serious adverse events and events of special interest trigger a narrative

In shortEvery death and every other serious adverse event, plus significant events the sponsor judges to be of special interest. A narrative plan turns that into rules a person or a program can apply. Posted results show why the count can run so high: in five large phase 3 trials we chose, 25% to 48% of participants had a serious adverse event.

The word serious has a fixed meaning. ICH E6(R3) lists an event that results in death, is life-threatening, requires inpatient hospitalization or prolongs a stay, results in persistent or significant disability, or is a congenital anomaly. An important medical event that needs intervention to prevent those outcomes should generally count too.2 Serious differs from severe. In cancer trials, severity is graded 1 to 5 on a scale called CTCAE.7 A Grade 3 rash can be severe without being serious, and a brief overnight admission can be serious without being severe.

Other significant adverse events are defined in E3 itself: marked laboratory abnormalities, and events that led to an intervention such as stopping the drug, cutting the dose or adding significant treatment.1 E3 asks for narratives only for those "judged to be of special interest because of clinical importance".

Two more clauses widen the net. A laboratory abnormality considered a serious adverse event gets a narrative under section 12.4.2.3.1 And in a US application, FDA's rule requires case report forms for each patient who died or "did not complete the study because of an adverse event, whether believed to be drug related or not".8 A plan may treat that same group as a trigger, so the writing matches the forms a reviewer can pull.

How five posted phase 3 trials show the size of the serious-event count

Those counts sit well above the 200 to 400 a vendor quotes for a large trial.3 Two writers at a contract research organization put the top of the range at "1000 or more".4 The plan decides how much of that count becomes narratives. Some categories "may be waived by prior agreement with the regulatory body", for example deaths where death is a study endpoint.14 E3 also allows clearly unrelated events to be omitted or kept very brief.1 Those decisions belong in writing before the first draft.

DVZ-301's plan: 351 matches become 247 narratives once overlaps merge

The narrative rulebook: a 1995 guideline, a 2025 rewrite of good clinical practice and FDA's new 2026 filing recommendation

In shortICH E3 still sets the content, in its 1995 text. ICH E6(R3), the good clinical practice guideline, governs how the data are handled and locked. New in 2026: FDA's conformance guide now says where subject narratives go and how to package them. The European Medicines Agency (EMA) can publish narratives from certain marketing applications, and its guidance asks for them to be anonymized and kept in the report.

Where a narrative's facts live: the SDTM tables, from demographics to disposition

In shortA narrative's facts are scattered across a dozen tables called domains. Writers read them through listings or patient profiles, and add two sources the tables do not hold: the sponsor's medical review and any adjudication committee's minutes. Coding dictionaries and grading scales add their own versions to keep straight.

SDTM, the Study Data Tabulation Model, is the standard from CDISC that sorts trial data into tables called domains: one for demographics, one for adverse events, one for doses and so on.22 A patient profile is a program's printout of one participant's records across those domains. One peer-reviewed paper on narrative writing gives the rule for using them: "Every data point in the template should be mapped against the patient profile."6 A template field with no source is a field a writer will be tempted to guess.

How many sentences each record type feeds in one narrative

Sources outside the tables: medical review, adjudication, MedDRA and CTCAE

How to write a patient narrative's chronology: study day arithmetic, timing against dose and house style

In shortA narrative runs in time order and places each event against the doses. The arithmetic is simple and easy to get wrong: the reference date, usually the first dose, is Day 1, there is no Day 0, and the count of days since the last dose is a subtraction of study days. One corrected date changes all of them.

The SDTM rule, as a public software reference restates it (CDISC's own text sits behind a login): on or after the reference start date, study day equals the date minus the reference date, plus 1. Before it, study day is the date minus the reference date, which gives Day −1 for the day before.23 FDA asks that the matching study day variables be submitted with the dates.17 Take participant 301-1043-0217, whose reference date is her first dose, March 12, 2026.

With the old date, the same sentence read "Day 82 (June 1, 2026), 18 days". One correction changes three values in one clause. Other sentences may depend on it too: anything saying "four days after onset" needs a fresh count. A writer who forgets the plus 1 writes Day 79, and every later day in that paragraph is off by one from the listing.

One participant's course from first dose to data cut-off

House-style rules that keep 247 narratives in one voice

These examples come from DVZ-301's fictional plan. ICH E3 sets none of them. One peer-reviewed paper advises three to five sample narratives covering each kind of trigger, agreed with reviewers before the rest are written.6

Causality in a narrative: reporting who thought the drug caused it

In shortE3 names two opinions: the investigator's and, where appropriate, the sponsor's. A committee may add a third. A narrative reports those opinions and says whose each one is. A writer who adds a view of their own has invented a medical opinion, and no later edit can give it a source.

Blinding needs care too. In a blinded study, some or all of the people involved do not know who received which treatment, so the treatment arm should stay out of text a blinded reader sees. EMA's guidance also tells applicants to raise any need for blinding redactions early when a report will be published.18 A rule in the plan settles this, and changes at unblinding.

Database lock and reopened narratives: what each weekly transfer changes

In shortLarge narrative projects start before the data are clean, so every transfer can change a sentence. The changes are few and hard to find: in the fictional DVZ-301, 14 records out of 53,216 touched 9 of 247 narratives. After lock and unblinding, E6(R3) allows changes only in exceptional circumstances.

Timing depends on volume. For completed studies with more than about 150 to 200 narratives, writing often starts alongside database cleaning, before lock. Below 150 it may start alongside the report itself.6 Starting early protects the deadline and creates re-work. A 2017 survey of 257 companies put the average time from last patient's last visit to lock at 36.3 days.24 After lock, a moderately complex phase 3 report still has months to run.14

Where narrative deadlines fall on a phase 3 report's calendar

What fourteen changed records looked like in transfer 14

Fourteen of 53,216 records is 0.026% of the delivery. Nothing in the transfer says "this moved a sentence", and nothing in a finished draft says "this sentence rests on that record". Unless each sentence is tied to its record, the only sure way to find the 9 is to re-read 247.

Soft lock, hard lock and E6(R3)'s own words on restricting edits

"Soft lock" and "hard lock" are industry words that appear in neither guideline we read. E6(R3) says "Prior to provision of the data for final analysis and, where applicable, before unblinding the trial, edit access to the data acquisition tools should be restricted". Changes after unblinding "should only occur in exceptional circumstances (e.g., data discrepancies that must be resolved for the reliability of the trial results)", authorized by the investigator, reflected in an audit trail and reported in the trial report.2

EMA's inspectors have listed "Insufficient data cleaning resulting in data changes detected only after primary analysis" as a CSR finding.25

The narrative errors reviewers find: stale dates, planned doses, unattributed causality

In shortMost narrative errors are copying and timing errors: a fact that was right in an older transfer, a dose taken from the protocol, a missing value quietly filled in. Nobody publishes an error rate, but the pattern shows in inspection findings and in FDA's letter on a report that gave planned doses in place of doses given.

Where narrative errors sit by how often and how much they hurt

What inspectors and FDA have said about reports like these

On November 27, 2024, FDA sent a warning letter over a CSR that "reported dose levels for subjects as stated in the protocol" when mislabeled supply meant "clinical sites administered 80% of the protocol-required dose to subjects". It cites 21 CFR 314.50(d)(5)(iv).26 That letter concerns the doses a report stated. The same copying habit can produce the same error in a narrative.

EMA's examples under the CSR heading include "Deficient quality control of the CSR" and "Incomplete or inaccurate reporting of protocol deviations and key data in the CSR".25 Neither names narratives. There is no public figure for how many narratives carry an error. If you want one for your own team, sample 30 signed narratives from your last study and check every fact against its record.

What a wrong patient narrative risks: reviewer questions, delay and lost trust in the report

In shortA narrative is one of the few places a reviewer reads a person's story. Errors there cost questions and time, and a sentence that cannot be traced makes a reader doubt the rest. Outside researchers read them too.

E3 section 12.3.3 asks the report to discuss the significance of these events.1 A reviewer who wants to test that discussion reads the narratives, opens the listings and, in a US application, can pull the case report forms.8 A date in the narrative that disagrees with listing 16.2.7 costs that reader's confidence along with a question.

What regulatory reviewers said about poor documents

Outside researchers read the patient-level parts of a report too

Sharma and colleagues studied clinical study reports for five antidepressants: 70 trials and 18,526 patients. Trials qualified only if they had "any patient narratives or individual patient listings of harms". They reported that "some outcomes appeared only in individual patient listings in appendices, which we had for only 32 trials", and that "discrepancies in reporting" may have led to serious under-reporting.30 Someone may one day compare your narrative with the tables, line by line, in public.

Publishing patient narratives: anonymizing section 14.3.3 under EMA Policy 0070

In shortIn Europe a clinical study report can become public. EMA says narratives should be anonymized and that it cannot accept whole-narrative redaction by default. Each narrative holds the details that can identify someone, such as age, site, dates and rare history, so writing with publication in mind saves work later.

Under the EU Clinical Trials Regulation (536/2014), for a trial intended to support a marketing authorization, the applicant submits the CSR to the EU database within 30 days of authorization, completion of the procedure or withdrawal.20 Under Policy 0070 EMA also publishes clinical study reports from certain marketing applications. Publication was suspended for all but COVID-19 medicines and resumed for new active substances with opinions from CHMP, EMA's human medicines committee, from September 2023. EMA's page says it plans to widen this from April 2025; the page does not confirm that it began.21

The guidance says narratives "should not be removed nor redacted in full regardless of their location in the clinical reports". If an entire narrative has to be redacted to achieve anonymization, "it must be clearly justified in the anonymisation report".18 Adverse event terms get "the highest priority" for retention.18 Section 14.3.3 is in scope.18 A sponsor that treats anonymization as someone else's later step meets all this at the end of the project.

What in one narrative can point to a person

This is general guidance; ask counsel for a legal view. Anonymization methods, thresholds and duties differ by regulator and by procedure. The same care applies wherever a report is shared outside the sponsor.

Narrative automation so far: templates, programmed narratives and AI drafting under regulators' eyes

In shortProgrammed templates are widely used, and AI drafting is arriving. Neither replaces medical review. Regulators have written no rule on AI for narratives, though FDA's draft and EMA's reflection paper set expectations. Vendor savings come without public datasets, and measured AI error rates on nearby work show why each fact still needs checking.

Three generations of narrative production

We found no independent, peer-reviewed test of AI on CSR narratives as of November 2026. The rates above come from clinical notes, drafted from consultation transcripts by one pipeline, and we use them only to show that both invention and omission occur and can be measured. At 1.47% of sentences and 13 sentences to a narrative, that would be about one invented sentence in every five narratives. That is our arithmetic on a rate from different work. Measure your own tool.

What regulators expect of AI in narrative writing

How medical writing and clinical teams produce narratives: roles, hand-offs and 157 working days

In shortA lead writer sets the plan and template, statistical programmers produce profiles, writers draft, a quality reviewer checks facts, a physician reviews and signs, and a data manager handles the queries. On one estimate built from a vendor's figure, 247 narratives take about 157 working days by hand. Most of that goes to assembling and writing.

From lead writer to study physician: who touches a narrative

Six steps and their hand-offs, from a data transfer to a compiled narrative

Where 157 working days go on 247 narratives by hand

Writing in house, a contract research team, or software

Participant 301-1043-0217: a moved date accepted, 41 facts checked and a signed narrative for section 14.3.3

In shortThis is the example from the top of the page, inside the product. Agents built the timeline and wrote and checked the draft. Dr. Sam Patel, the study physician, points at a sentence to see its data, accepts the corrected date, confirms a comment from the lead writer and signs. The narrative then locks.

1

Her whole course on one time axis, the narrative beside it

Sam has 31 narratives waiting before the October 16 lock, and participant 301-1043-0217 is first. On the left is one shared axis of study days: her four doses, then the dose held and the drug stopped; her adverse events; oxygen saturation against its normal band; her medicines; her hospital stay and tests.

On the right is the narrative as it will read in section 14.3.3. It runs to 287 words in 13 sentences. Each sentence carries small numbered citations. The agents did all this before Sam opened the screen.

2

Point at a sentence and its data light up

Sam picks the sentence "On Day 84 a chest CT scan showed bilateral ground-glass opacities with no pulmonary embolism, and oxygen saturation was 89% on room air." The timeline dims. The 89% reading pulses beside the dose marked Held, and a line drops at Day 84 (June 3, 2026).

The strip above the chart reads "Sentence 5 → 2 sources". He no longer has to open two listings to check it.

3

The records behind the sentence, with the values used

Below the timeline the records open: the chest CT from the procedures data and the vital signs. The values the sentence used are highlighted: SpO₂ 89% on room air on Day 84, flagged low, and 95% on Day 92, which a later sentence uses.

"Open in the listing" goes to the full record. This is the check a medical reviewer does all day, one click from the text.

4

The moved date waits as a tracked change

The site corrected the onset on Monday, October 5, after query Q-1043-118, and transfer 14 carries it. The data checker proposed one edit in sentence 4: "Day 82 (Jun 1, 2026), 18 days" becomes "Day 80 (May 30, 2026), 16 days". The sponsor's medical review already said May 30.

Sam accepts it. Lena Ortiz, the lead medical writer, has added the committee's outcome to the causality paragraph and asks him to confirm the wording. He resolves her comment.

5

Forty-one facts and all fourteen E3 elements

The checks are done before he reads. 41 of 41 facts match transfer 14: dates, study days, doses, grades, outcomes and causality. The onset now agrees with the sponsor's medical review.

The list below ties each ICH E3 element to the sentence that carries it. Post-mortem findings show as not applicable, because she is alive.

6

He signs, and the signature says what it means

The dialog shows the three things that must be settled: facts 41/41, ICH E3 elements 14/14, open changes 0. It then states the meaning of the signature: he has reviewed the narrative against the study data and approves it for the clinical study report, section 14.3.3.

He signs with his name and password. The status turns to Signed, the narrative locks and the signed count moves from 157 to 158. A later change to a record behind it reopens it only as a tracked change for him to accept or reject.

The solution's design targets are a median of 48 seconds to a first draft, 12 minutes of medical review per narrative, and 99.4% of drafted facts matching the data, with every difference flagged before review. They are targets, to be measured on your own study. The screens leave out the study drug's dose, which E3 asks for when it varies among patients, and show the comparator only as chemotherapy. A real narrative follows your plan.

Nineteen more screens, from the 247-tile narrative wall to the sign-off settings

In shortOne participant is the story. These screens show the rest: how a data point points back to sentences, a missing dose and its site query, a death narrative, the whole study on one wall, a drafting run, what a transfer changed, the plan, the progress figures and the settings a sponsor controls.

Select any screen to open it larger. The sponsor, study, drug, people, participants and figures are fictional. Dates and spellings are shown in US form. The narrative plan's section numbers and wording are the sample's own.

Trigger finder to style editor: five narrative agents, and a study physician who signs every one

In shortEach agent has one job and limits it does not cross. None signs a narrative, edits study data or gives an opinion on cause. A study physician signs every narrative; under this study's settings, deaths also take the medical lead's signature. Every draft, change, comment, query and signature goes on the audit trail with a person and a time.

Five agents between a transfer and a draft narrative

The study physician signs each narrative; the plan owner controls the rules

The system carries the audit trail, the stated meaning of each signature and the link between signature and record, controls that support your compliance with 21 CFR Part 11 and EU Annex 11 when validated in your environment. Your procedures decide who signs what. Botminds' own credentials are SOC 2 Type II audited and information security managed to ISO 27001.

Who does what on DVZ-301: physician, writer, medical lead, monitor and data manager

SDTM transfers and a narrative plan in; signed, cited narratives for section 14.3.3 out

In shortThe solution reads the data a narrative already comes from, follows your plan and leaves behind signed narratives with every sentence cited, tracked changes when data move, drafted site queries, a tracker and an audit trail. Your clinical database stays the system of record.

Clinical data, medical review and report systems a narrative rollout connects to

What a hand-written narrative costs you in writer hours, physician minutes and re-reads

In shortStart from your own costs, because few are public. A few published inputs frame the sum: what a writer's and a physician's hour cost, the one vendor figure for writing time, and what a day of delay costs a development program.

Timing ten narratives to set your own baseline

A late report, an unread trace and writers lost to copying

What faster narratives return when you count writer and physician hours against the tool's cost

In shortPut in your own numbers. The calculator starts from published median wages and asks you for the rest: how many narratives you write a year, the time they take now, the time you expect after, and what the tool and your testing cost. Every input marked illustrative is an assumption to replace.

Drafting, medical review and transfer re-work: where narrative automation saves

Narratives a year, writer hours and physician minutes

A sponsor or contract research team writing 1,000 narratives a year covers the cost in the first two cases and falls short in the third. A team writing 40 narratives a year on one small study probably does not cover it on labor alone, and should weigh that before buying anything.

How ready your medical writing team is: volume, data standards, plan discipline and sign-off

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 one narrative plan first

What to ask narrative software or a writing service about citations, transfers and causality

In shortAsk every vendor the same questions, us included. The answers show how a narrative is built: what it reads, how each sentence ties to data, what happens at the next transfer, who owns causality, and how you would measure it on your own study.

We found no peer-reviewed independent test of any narrative vendor's accuracy, as of November 2026. A planted-error test on your own study is the fairest comparison, and the same test measures your current process.

Test narrative automation on a finished study with planted errors, then decide on three numbers

In shortMeasure first, then test on a completed study's de-identified data with errors you planted, then run alongside a live study and decide with three numbers. None of these steps commits you to a vendor.

1. Time ten narratives, simple and complex, by step, and log reviewer minutes. That is your baseline.
2. Check 30 signed narratives from your last study, every fact against its record, to learn your own error rate.
3. Pick one completed study with its plan, template and de-identified SDTM transfers. Copy 20 narratives' records and plant known errors, including a moved date.
4. Run in parallel on a live study through two or three transfers, with your physicians signing every narrative.
5. Decide with three numbers: catch rate by error type, false alarms on clean records, and reviewer minutes to a signed narrative.

Catch rate, false alarms and trace accuracy for each planted error

CSR narrative questions answered: ICH E3 12.3.2, section 14.3.3, length, timing and AI

In shortThe questions medical writers, clinical teams and buyers search for most, each answered on its own.