Compliance19 min read

MHRA GCP Source Records: Extracting Clinical Trial Data

Dilr Mira is a class of private clinical small language models from DILR.AI that turn clinical documents into draft structured data. This guide explains where that extraction step sits under ICH E6(R3), what the MHRA's September 2026 good clinical practice updates bear on, and how human verification, scaled to criticality, decides what reaches a case report form.

On this page 12 sections
MHRA GCP Source Records: Extracting Clinical Trial Data DILR MIRA · PHARMA MHRA GCP Source Records: Extracting Clinical Trial Data 01 Define the source 02 Extract with metadata 03 Verify by criticality 04 Reconcile and retain dilr.ai/blog

Every clinical trial runs on a chain that starts with a source record and ends with a data point in a case report form. The chain used to be a research nurse with a paper chart on one side and an electronic data capture screen on the other. It is now, more often, a scanned discharge letter, an external laboratory PDF and an electronic health record export, with a growing temptation to put a language model in the middle. Since 28 April 2026 that chain has been judged in the UK against ICH E6(R3), the good clinical practice guideline the ICH adopted on 06 January 2025, which the MHRA says the UK implemented on 28 April 2026 alongside the amended clinical trials regulations, according to its UK-specific annotations to ICH E6(R3).

This guide is written for the Chief Data and AI Officer at a trial sponsor or contract research organisation, the person who answers for data integrity when an MHRA inspector asks how a value in the trial database got there. In September 2026 the MHRA changed two things on its good clinical practice inspections page: on 16 September it attached a document of position statements and guidance on electronic health records, and on 22 September it introduced a specific inspection dossier for organisations providing computerised systems in support of clinical trials. Neither mentions extraction models, but both bear on the documents and systems an extraction step depends on. If you are mapping this kind of change across a whole function, our approach to placing AI starts with exactly this sort of data flow.

The scope here is deliberately narrow: where document extraction sits between a source record and a data acquisition tool, what ICH E6(R3) expects of that step, and what the September updates ask an organisation to show. It cedes adverse event case intake to our pharmacovigilance extraction guide, manufacturing AI validation under draft EU GMP Annex 22 to our Annex 22 guide, because Annex 22 is a manufacturing text and not a trials one, and the pharmacovigilance inspection evidence pack to our GPvP inspection guide. How extraction models are measured in general is covered in our clinical document extraction guide.

This guide is shipped by the team behind Dilr Mira, private clinical small language models that turn clinical documents into structured, source-grounded JSON on hardware you control. Or see DATS, the five-stage consulting system that places AI inside regulated operations.

What counts as a source record under ICH E6(R3) in the UK?

Under ICH E6(R3), which the UK implemented on 28 April 2026, a source record is the original document or data, including relevant metadata, or a certified copy of it, whatever the medium. Medical records, participant-entered data, pharmacy and laboratory records all qualify. The investigator defines what counts as source, how it is captured and where it lives before the trial starts, and the sponsor carries its own record-keeping duties under UK law.

The glossary of the guideline is precise about the medium: paper, an electronic health record entry, an electronic patient-reported outcome and data from an automated instrument are all capable of being source. Annex 1 section 2.12.2 then sets the quality bar. It asks the investigator or institution to keep source records that are attributable, legible, contemporaneous, original, accurate and complete, to make changes traceable without obscuring the original entry, and to define the source before the trial starts, updating that definition when needed. The duty in that section binds the investigator or institution, not the sponsor and not any technology supplier.

The sponsor's side of the trail sits in UK law. Under regulation 31A of the Medicines for Human Use (Clinical Trials) Regulations 2004, the sponsor must make any alteration to a trial master file document traceable, and since 28 April 2026 the sponsor and chief investigator must retain trial master file documents, including those held in electronic form, for at least 25 years after the trial concludes. The medical files of trial participants carry the same 25-year floor. The 2026 amendments came through the Medicines for Human Use (Clinical Trials) (Amendment) Regulations 2025. A model output that lands in that file inherits the same retention clock.

Two practical consequences follow for anyone planning an extraction step. First, the source definition is set before the trial and updated only when needed, so an extraction step is best named in the data flow at the start rather than added part way through. Second, the source is the document the model reads, never the JSON it writes. Teams that already run an AI operating model for regulated data tend to find the second point obvious; teams that start from a tool demo tend to find it late.

Where does document extraction sit in the trial data flow?

Document extraction sits between the source record and the data acquisition tool, usually the electronic case report form. ICH E6(R3) defines a data acquisition tool as collecting data from an originator that can be a computer system, naming extraction from an electronic health record as an example. A model reading a clinical document is therefore a computerised system step in the data flow, not a source in its own right.

That framing matters because ICH E6(R3) treats computerised systems as a governed category. Principle 9.3 says systems used in clinical trials should be fit for purpose, for example through risk-based validation where appropriate, and section 4.3 asks that the responsibilities of sponsor, investigator and other parties for each system are clear and documented. The MHRA guidance on compliance with ICH E6 GCP in the UK, published on 27 March 2026, quotes the amended regulations, which state that the functions of the sponsor include the development and maintenance of trial specific computerised systems.

Where an extraction step sits between source and case report form
01Source recordOriginal document or certified copy, defined before the tria…02Extraction draftModel output with document reference, model version and time…03Human verificationDepth set by the criticality of each field04Data acquisition toolEntry into the eCRF with its own audit trail05Transfer reconciliationChecks that nothing was lost or altered between systems
Our reading of the ICH E6(R3) data life cycle: the document stays the source, the model writes a draft, and a person decides what reaches the data acquisition tool.

A separate expectation in the guideline should shape the whole design. Annex 1 section 2.12.2, addressed to the investigator, says unnecessary transcription steps between the source record and the data acquisition tool should be avoided. An extraction model is an additional step. It earns its place only where the alternative is a person keying values out of an unstructured document, such as a scanned referral letter or an external laboratory report that arrives as a PDF. Where the value already exists as structured data in the hospital record, a direct structured transfer is the shorter chain.

Is a model's extracted output a source record?

No. A model's extracted output is a draft derived from the source record, and ICH E6(R3) keeps the original document, or its certified copy, as the source. The draft becomes trial data only after the human verification step set for its criticality, and the trail has to show which document it came from, which system produced it and who accepted or changed each value.

Section 4.2 of the guideline turns that principle into specific expectations, and each one maps onto an extraction step without stretching. The table below is our reading of how they apply; it is not MHRA or ICH text.

ICH E6(R3) expectationSectionWhat it means for an extraction step
Acquired data from any source carries relevant metadata4.2.1(b)Record the source document reference, the model and version, and the time of extraction with every draft
Automated validation checks at capture are risk-based, controlled and documented4.2.1(c)Range and format checks on extracted values are part of the validated configuration, not ad hoc scripts
Audit trails are not disabled and timestamps are unambiguous4.2.2Keep the draft, the reviewer's changes and the accepted value; the guideline gives coordinated universal time as an example
Corrections are attributed to the person or computerised system making them4.2.4When a reviewer overrides a model value, the record shows both the model's value and the person's
Transfers between systems keep integrity, with reconciliation where appropriate4.2.5Reconcile what the model extracted against what reached the eCRF, so nothing is silently dropped

Section 4.2.1(a) deserves a careful reading. It covers data captured on paper or in an electronic health record that is manually transcribed into a computerised system, and it ties the extent of verification to the criticality of the data. A model is not manual transcription, so the honest position is that the extraction step is governed by the computerised system expectations above, and that the human verification step which follows it is where criticality-based verification applies.

What did the MHRA add to its GCP inspections page in September 2026?

The MHRA made two changes to its good clinical practice inspections page in September 2026. On 16 September it added a document of MHRA position statements and guidance on electronic health records. On 22 September it introduced a specific inspection dossier for organisations providing computerised systems in support of clinical trials, alongside updated wording on how dossiers are submitted.

The electronic health records document needs careful dating. The file itself is a version 1.0 statement headed 16 September 2015, written for NHS Trust and Health Board chief executives and research and development departments, and for commercial sponsors whose trials are hosted in the NHS. What changed in 2026 is that the MHRA attached it to the inspections page. Its retention lines date from that time: it quotes the then five-year minimum for participant medical files under regulation 31A(8), and notes 25 years as standard under the then new EU Clinical Trials Regulation 536/2014. The UK 25-year rule took effect on 28 April 2026, so treat the five-year line as superseded.

The rest of the electronic health records position statement is worth reading before pointing any model at a hospital record. When it was written, the statement listed problems inspectors had found with electronic records:

  • audit trails missing or too thin to reconstruct changes to trial data;
  • no way to show investigator oversight where a research nurse or dictation service entered everything;
  • paper print-outs given to monitors that left out information held in the electronic system;
  • bulk scanning and disposal of paper records with no quality control for certified copies;
  • scans saved as PDF files in no particular order and with missing sections, which can make reconstructing a trial impossible.

The responsibilities in that statement fall on Trusts and Health Boards acting as sponsors or as hosts responsible for source records, and it is also addressed to commercial sponsors whose trials are hosted in the NHS; it places nothing directly on an extraction supplier. The practical point for a data leader is narrower. Any extraction model pointed at hospital records reads the kind of scanned PDFs the statement worried about, and it cannot recover a page that was never scanned or a colour code lost to black and white scanning. Scan quality and certified copy control sit upstream of any model, and an extraction programme that does not audit them is measuring the model against inputs nobody has checked.

What does the computerised systems dossier ask an extraction supplier to show?

The MHRA computerised systems dossier, introduced on 22 September 2026, asks organisations that provide computerised systems for UK trials to list each system and its trial functionality, its version history over the greater of the period since the last inspection or three years, its current validation status, the reasons for significant changes, their quality procedures, supporting tools and delegated third parties.

Organisations notified of an inspection must supply the dossier within 30 days, in two sections: six items in the first and four in the second, with the electronic copy limited to 100 pages, according to the MHRA computerised systems dossier template. The MHRA inspections page already lists organisations who develop and maintain trial specific computerised systems among those that have to comply with good clinical practice. The template's own examples of trial systems include electronic data capture, interactive response technology, digital health technology, electronic trial master files and electronic consent.

The template does not mention extraction models, so the mapping below is ours. A model fits the dossier's shape with uncomfortable precision:

  • System version is the model and adapter version actually used to produce each draft, which means pinning weights rather than calling whatever a hosted endpoint serves that day;
  • Significant change includes retraining, a new prompt or schema, and a change of base model, each with a stated reason;
  • Validation status is the evaluation evidence on documents that look like this trial's documents, not a vendor benchmark;
  • Legacy and retired systems includes earlier model versions whose drafts still sit in the trial record, with the arrangements for keeping and reading them;
  • Supporting software covers the tools that manage model versions, test sets and releases, which the template lists under source code management, test management and release management.

The template also carries the only sentence about AI that we found across the MHRA pages reviewed for this guide. It is about preparing the dossier itself, not about generating trial data. In the MHRA's words: "The MHRA recognises that organisations may use artificial intelligence (AI) tools to assist in preparing inspection-related documentation. Such use is acceptable provided appropriate human oversight is maintained." The template adds that the organisation remains responsible for the information being accurate, complete and attributable. Read as written, that is a statement about paperwork, and it should not be stretched into a view on AI in trial data.

The same discipline underpins our AI operating model consulting, which writes down who owns each system, each change and each validation decision before the first model goes live.

How much human verification does extracted trial data need?

ICH E6(R3) does not set a single verification rate for extracted trial data. It asks sponsors to focus quality control on data of higher criticality and to scale verification of transcribed data to that criticality. In practice that means full source checks on fields that decide eligibility, safety or a primary endpoint, and a lighter, documented check for low-criticality fields.

The tiering below is our view of a defensible starting point, not a regulatory table. The right split comes from the trial's own critical to quality factors, which the sponsor identifies at the design stage.

CriticalityTypical extracted fieldsVerification we would start from
HighestEligibility criteria values, primary endpoint measurements, serious adverse event datesEvery value checked against the source by a trained person before entry
MediumConcomitant medications, secondary endpoint labs, visit datesEvery draft reviewed, with field-level spot checks against the source
LowerDemographic context already held elsewhere, non-endpoint vitalsEvery draft reviewed for completeness, with documented sampling of values against the source, widened if the sample finds errors

Two failure patterns justify starting strict. The first is an error that is plausible: a model that misreads a laboratory unit produces a number that passes a range check and still corrupts an endpoint. The second is an omission: a model that skips a medication listed on a page it handled poorly produces a clean record that is simply incomplete. The MHRA grades inspection deficiencies at three levels, critical, major and other, and its critical definition includes evidence that clinical trial data are unreliable. A verification step designed only to catch obviously wrong values does not address either pattern.

The reviewer's work also has to be visible. Section 4.2.4 attributes corrections to the person or the computerised system that made them, and the electronic health records statement flagged investigator oversight that could not be demonstrated. When a person accepts a model's draft without change, the record should say so; when they change it, it should keep both values. Our AI execution office is set up to run exactly this kind of reviewed workflow alongside an internal team rather than in place of one.

Why does the deployment location matter for trial source documents?

Trial source documents carry identifiable patient information, while the case report form normally holds a participant number. Running extraction inside the site or sponsor environment keeps the identifiable document where it already lives. Sending it to a hosted extraction service adds an organisation to the data flow, and with it another system an inspector may need to reach and assess.

The MHRA inspections page says the complete trial master file is the basis for inspection, including electronic records and emails, and that where several electronic systems make up the file, direct access must be made available to each system. Where a hosted model's records form part of the trial master file, it is one more system in that list, with its own version history to evidence. Health data is also special category data under UK GDPR Article 9, so a new processor in the chain means new contractual and transfer questions for the privacy team, separate from the GCP questions above.

Self-hosting does not remove the obligations. It changes who carries them. A sponsor or site running its own model owns the version pinning, the validation evidence and the access control that a supplier would otherwise document. For organisations that already operate validated systems on their own infrastructure, that is familiar ground; for those that do not, the DATS methodology usually starts by deciding which party should own the system before choosing the system.

What can Dilr Mira do with trial source documents today?

Dilr Mira's current release, Mira-Q2, reads English clinical documents on hardware the customer controls and returns ten fixed JSON fields, including patient age and sex, encounter date, labs, medications and diagnoses. It is strongest on laboratory reports, medication lists and discharge summaries, every output is a draft for human review, and it is not a validated trial system.

The figures behind that summary come from DILR.AI's own published evaluation, reported on the Dilr Mira product page, with the full scorecard on Hugging Face. Across 782 evaluated documents in four test sets, the identifier leak rate was 0.0%. JSON validity was 100% on 200 held-out documents drawn from the same distribution as training, and 85.8% on 282 real physician notes across 39 specialties. The model was trained on synthetic data and struggles with document types it never saw, such as operative notes and physical examinations. Those are evaluation results on published test sets, not a measurement on any trial's documents.

Three facts shape how it could fit a trial data flow:

  • The schema is generic. The ten fields describe a clinical document, not a case report form. Mira-Q2 was trained on 2,000 schema variants so it follows a schema passed in its prompt, but mapping a trial's eCRF is a fitted deployment that starts from that trial's own documents, scoped through the DATS consulting system.
  • The output holds back identifiers by design. The patient field returns age and sex only, never a name or record number, which suits a design where the participant number is attached by the trial system rather than read from the document.
  • It runs where the documents are. Once the weights are downloaded, inference needs no network connection, so the document does not leave the environment that holds it.

Two limits matter as much. The 120 MB adapter DILR publishes is licensed under Apache-2.0, but the Qwen2.5-3B-Instruct base model it runs on is licensed for non-commercial use, and commercial users are asked to request a licence from Alibaba Cloud, so a commercial sponsor has two licences to check. And Mira-3, which aims at more languages, any schema without setup and identifiers removed before the model reads the document, is in development, not available. Validation for a specific trial remains the responsibility of the organisation that puts any model into its data flow. If you want the general picture first, what Dilr Mira is and the Mira-Q2 evaluation write-up set out the scorecard in more depth.

What is the best way to extract clinical trial source data in 2026?

The best way to extract clinical trial source data in 2026 depends on the input. Structured values already in an electronic health record are best moved by a validated structured transfer, such as an HL7 FHIR integration, with no model involved. Extraction models earn their place on unstructured inputs like scans and external PDFs, inside a governed, verified and versioned step.

We would judge any approach, ours included, on four criteria:

  1. Shortest honest chain. ICH E6(R3) asks for unnecessary transcription steps to be avoided, so the first question is whether a structured feed exists. Where it does, a direct integration over a standard such as HL7 FHIR, a specification for exchanging healthcare information electronically, beats any extraction model, and we would recommend it over Dilr Mira.
  2. Evidence on your documents. A vendor benchmark is not validation, and that includes our own published Mira-Q2 scorecard. Ask for results on a sample of the trial's own source documents, scored field by field against a human gold standard.
  3. Version control you can show an inspector. Pinned model versions, recorded changes with reasons, and the ability to say which version produced every draft, which is what the computerised systems dossier asks about any system.
  4. Location of the identifiable document. Hosted services such as Amazon Comprehend Medical, which AWS describes as a HIPAA-eligible natural language processing service, process documents on the provider's platform, which can suit an organisation with mature cloud contracts; self-hosted open-weight models keep the document in place but move the operating burden in-house.

There are scenarios where Dilr Mira is the wrong tool. A trial whose source data is overwhelmingly structured should invest in integration, and a team with no capacity to run and validate a model on its own infrastructure may be better served by an established supplier who carries that work. Where unstructured documents are the bottleneck, the AI solutions we build start with a short placement exercise to establish which of those situations applies. For the wider pharma picture, our pharma sector guide maps where each DILR.AI line fits, and the compliance category collects the related regulatory guides.

Frequently asked questions

Does the MHRA allow AI to be used on clinical trial data?

The MHRA has not issued a general approval of AI for clinical trial data. The only AI statement on the MHRA pages reviewed for this guide, in the computerised systems dossier template, accepts AI help with preparing inspection documents where human oversight is maintained. A model in the trial data flow is assessed like any computerised system under ICH E6(R3), on fitness for purpose and evidence.

That means the questions an inspector can ask about an electronic data capture system apply equally to an extraction model: what it does, which version ran, how it was validated and who checked its output. Our about page explains how the team approaches regulated delivery, and the ambient scribe comparison explains why document extraction and clinical note generation are different problems.

How long must clinical trial source records be kept in the UK?

Since 28 April 2026, regulation 31A of the Medicines for Human Use (Clinical Trials) Regulations 2004 requires the sponsor and chief investigator to keep trial master file documents, including electronic ones, for at least 25 years after the trial ends. Participant medical files carry the same 25-year minimum, or longer where another law requires it.

There is one extension. If, when the 25 years expire, the trial data are supporting a UK marketing authorisation application, the trial master file documents must be kept for at least two years after that authorisation is granted. Any model drafts and review records that form part of the trial master file follow the same clock, which is a reason to decide early what is kept and in what format. The pharmacovigilance extraction guide covers the separate retention questions for safety data.

Want to test this on your own documents? See Dilr Mira and its published scorecard, book an AI placement diagnostic, read the clinical extraction model guide, or read about our approach to placing AI inside regulated systems.

Talk to the operators

Keep trial data traceable from scan to case report form.

30-min scoping call · No deck · Confidential. We will tell you where an extraction step belongs in your data flow, and where a structured feed is the better answer.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.

  • mhra gcp source records extraction clinical trials
  • pharma clinical trials
  • clinical trial source data extraction
  • ich e6 r3 source records ai
  • electronic health records clinical trials mhra
  • clinical trial data extraction reddit
  • best clinical document extraction model 2026
  • dilr mira

What does this article answer?

The questions this article answers, in short

  • Under ICH E6(R3), which the UK implemented on 28 April 2026, a source record is the original document or data, including relevant metadata, or a certified copy of it, whatever the medium. Medical records, participant-entered data, pharmacy and laboratory records all qualify. The investigator defines what counts as source, how it is captured and where it lives before the trial starts, and the sponsor carries its own record-keeping duties under UK law.
  • Document extraction sits between the source record and the data acquisition tool, usually the electronic case report form. ICH E6(R3) defines a data acquisition tool as collecting data from an originator that can be a computer system, naming extraction from an electronic health record as an example. A model reading a clinical document is therefore a computerised system step in the data flow, not a source in its own right.
  • No. A model's extracted output is a draft derived from the source record, and ICH E6(R3) keeps the original document, or its certified copy, as the source. The draft becomes trial data only after the human verification step set for its criticality, and the trail has to show which document it came from, which system produced it and who accepted or changed each value.
  • The MHRA made two changes to its good clinical practice inspections page in September 2026. On 16 September it added a document of MHRA position statements and guidance on electronic health records. On 22 September it introduced a specific inspection dossier for organisations providing computerised systems in support of clinical trials, alongside updated wording on how dossiers are submitted.
  • The MHRA computerised systems dossier, introduced on 22 September 2026, asks organisations that provide computerised systems for UK trials to list each system and its trial functionality, its version history over the greater of the period since the last inspection or three years, its current validation status, the reasons for significant changes, their quality procedures, supporting tools and delegated third parties.
  • ICH E6(R3) does not set a single verification rate for extracted trial data. It asks sponsors to focus quality control on data of higher criticality and to scale verification of transcribed data to that criticality. In practice that means full source checks on fields that decide eligibility, safety or a primary endpoint, and a lighter, documented check for low-criticality fields.
  • Trial source documents carry identifiable patient information, while the case report form normally holds a participant number. Running extraction inside the site or sponsor environment keeps the identifiable document where it already lives. Sending it to a hosted extraction service adds an organisation to the data flow, and with it another system an inspector may need to reach and assess.
  • Dilr Mira's current release, Mira-Q2, reads English clinical documents on hardware the customer controls and returns ten fixed JSON fields, including patient age and sex, encounter date, labs, medications and diagnoses. It is strongest on laboratory reports, medication lists and discharge summaries, every output is a draft for human review, and it is not a validated trial system.
Written by Dilr.ai EngineeringEngineering team

How do I deploy voice AI without failing an audit?

Dilr Voice applies call compliance rules for each country

Consent, do-not-call checks and permitted calling hours are built in, and the UK AI compliance changelog tracks ICO, FCA and EU AI Act changes as they land.