Clinical study reports and source tables for medical writing AI

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why unpaired CSRs do not train a results-section drafter

A model learns to write Section 11 or 12 text only when it sees the exact tables that text summarizes. ICH E3 defines the CSR as an integrated document that combines clinical and statistical narrative with tables and figures, plus standardized appendices covering the protocol, sample case report forms, statistical outputs and patient data listings. Under the 1995 numbering, efficacy results sit in Section 11, safety in Section 12, narratives of deaths and other serious or significant adverse events in Section 12.3, the tables and figures referred to in the text in Section 14, and patient data listings in Section 16.2; check the numbering against the EMA-hosted text [6] and its Q&A (R1) before you encode it as a schema.

The public alternatives fall short for different reasons. EMA Policy 0070 publishes clinical reports, including CSRs, but only after commercially confidential information is redacted and participant personal data protected, and publication under the policy was paused for several years before resuming in 2023; check EMA's current guidance [7] as of October 2026. Registry-derived benchmarks such as TrialBench convert ClinicalTrials.gov records into 23 AI-ready task datasets, which is useful for trial-design prediction but contains no CSR prose or TLF-to-text alignment [1].

What buyers really need is the input-to-output pair: a TLF number such as "Table 14.3.1.2" linked to the paragraph that cites it, with the protocol and SAP as context. That pairing is sponsor work product, usually confidential, and sits in regulatory document management systems (Veeva Vault RIM or Clinical, Documentum-based eCTD archives) and statistical computing environments, not on the open web.

What a complete CSR training unit contains

A complete unit is one study, versioned, with every document a writer touched and the links between them. Ask suppliers to package per study rather than per document, because splitting the CSR from its TLFs destroys the supervision signal.

• Protocol and amendments, with the version that applied at database lock flagged. See our guide to protocol and amendment corpora if authoring is your primary task.
• Statistical analysis plan and, where available, the TLF shells or mock-up document.
• Final TLF package as RTF or PDF outputs, ideally with the SAS or R program names and the ADaM dataset (ADSL, ADAE, ADEFF-style domains) each output was generated from, plus Define-XML.
• CSR body and synopsis, final signed version, with the draft history if the sponsor kept tracked rounds. Drafts show reviewer corrections, which make strong preference or critique data.
• Patient narratives and their source listings (AE, conmed, lab, disposition) in Section 16.2 style appendices.
• Metadata: therapeutic area, phase, design, indication, study status, writer type (sponsor or CRO), and the redaction or de-identification method applied.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "study_unit_id": "STUDY-0412",
  "phase": "2",
  "therapeutic_area": "immunology",
  "documents": {
    "protocol": {"version": "Amendment 3", "format": "docx"},
    "sap": {"version": "2.0", "format": "pdf"},
    "tlf_package": {"count_outputs": 186, "format": "rtf", "adam_linked": true},
    "csr": {"version": "final", "sections_present": ["synopsis", "9", "10", "11", "12", "14", "16.2"]}
  },
  "alignments": [
    {"csr_section": "12.2.2", "paragraph_id": "p-0381", "cites_output": "Table 14.3.1.2", "claim_type": "incidence_by_soc"},
    {"csr_section": "12.3.2", "paragraph_id": "n-0027", "source_listings": ["L16.2.7.1", "L16.2.8.3"], "type": "patient_narrative"}
  ],
  "deidentification": {"method": "expert_determination", "subject_ids": "re-keyed", "dates": "offset_per_subject"},
  "ownership": "sponsor; CRO authorship disclosed"
}

The alignments array is the expensive part. If a supplier cannot produce it, budget for your own alignment pass, since automated matching of numbers in prose to RTF cells breaks on rounding, footnote references and pooled versus by-arm columns.

Patient narratives carry the hardest privacy problem

Patient narratives are the highest-value and highest-risk part of a CSR corpus, because they tell an individual's clinical story in sequence. A narrative typically holds a subject ID, age, sex, race, site, study day of onset, dosing history, concomitant medications, lab values and outcome, which together can single out a person even with names removed. Research on de-identification documents that data released as de-identified has been re-identified in practice [3].

If the supplier is a HIPAA covered entity or the data includes PHI, the de-identification route matters: Safe Harbor removes the listed identifiers, including all elements of dates except the year, while Expert Determination lets a qualified expert retain study-day structure with documented risk [2]. A limited data set keeps some dates but still requires a data use agreement and is not de-identified [2]. Most sponsor trial data is governed by informed consent and GDPR or other regimes rather than HIPAA alone, so ask what the consent form said about secondary use.

Practical checks for narratives: confirm that subject IDs are re-keyed consistently across listings and narratives (otherwise alignment breaks), that dates are shifted per subject rather than per study, that rare events and small sites are reviewed for small-cell risk, and that investigator and sponsor staff names are handled under a stated policy as well as participant data.

Ownership, CRO authorship and license scope

The sponsor normally controls a CSR even when a CRO's medical writers drafted it, but that depends on the master services agreement, so verify it per study rather than assuming. A CRO offering "its" CSR archive may hold drafting know-how but not the right to license sponsor documents, TLFs or patient-level outputs. Ask who executed the license, which studies it covers, and whether any partner or co-development agreement restricts reuse.

License hygiene problems are common in AI data generally: an audit of more than 1,800 text datasets found license information missing for over 70% and miscategorized for over 50% on popular hosting platforms [4]. For CSR data, the license should name the studies, the document types, permitted uses (training, fine-tuning, evaluation, retrieval), whether outputs may quote verbatim text, term and deletion. If you ship a generative model to users in California, AB 2013 has required posted documentation (due 1 January 2026; status as of October 2026) about training data [5], so keep provenance records per study unit.

Buyer checklist for CSR and TLF data

Use this list to screen offers before you spend time on samples.

Illustrative example: invented to show structure; it does not describe an available dataset.

Two failure modes recur in evaluation. Template leakage occurs when one sponsor's boilerplate dominates training, so the model scores well on that sponsor and poorly elsewhere; split by sponsor where you can. Numeric hallucination is the other: score generated results text by extracting every number and checking it against the cited TLF cell, not with ROUGE alone. For broader input-to-report methods, see report-generation fine-tuning data.

How CSR data fits with adjacent clinical sources

CSR pairs are distinct from safety case data and from general clinical text, and mixing them blurs what your model learns. Individual case safety report narratives follow ICH E2B and MedDRA coding conventions; see pharmacovigilance case narratives and ICSRs. For HIPAA trade-offs across clinical and administrative records, see healthcare LLM fine-tuning data, and for the full set of sector guides, the industry-specific operational data hub. SourceX's healthcare buyer page and enterprise document datasets cover related requests.

How SourceX approaches CSR data requests

SourceX sources operational datasets, including documents, finance and legal workflows, from US companies on request; nothing is held in stock and a request does not guarantee a match. Buyers describe the data they need, and SourceX looks for US businesses that hold it, with every release approved by the supplying company. Each dataset is rights-reviewed for ownership and consents, personal details are removed or replaced with the method recorded and a sample checked, and health records require HIPAA de-identification by Safe Harbor or Expert Determination. You can describe a paired CSR request on the SourceX buyer page.

Request paired CSR and TLF data for medical writing AI

SourceX manages the commercial process for operational datasets from US companies, from finding a supplier through licensing that defines records, uses, term and delivery. Nothing is contracted until a supplier agrees, and delivery runs through private, access-controlled workflows after an executed agreement. Describe the studies, documents and pairing you need at sourcex.si/buyers.