Pathway map
The medical-device evidence chain: from intended purpose to post-market evidence
How do the parts of a medical device’s technical documentation connect — and what has to be in place before an AI function may answer from it?
A pathway map reworks a whole work into the logic of a pathway card: lanes for its domains, stations with before, after and maturity, and one route per reader role. Where do you start?
The short answer
A medical device file is an evidence chain: the intended purpose anchors requirements, risks and their controls, verification, clinical evidence, claims, labeling and IFU and the post-market data that feeds back into them. Structured content alone doesn’t make that chain usable for people or AI; what does is controlled relationships between these objects, each with an id, a version and a released state.
For: Technical writers, technical marketers, project managers and documentation systems developers in medical device companies
Key points
- The intended purpose is the anchor: requirements, risks, claims, the clinical evaluation and the IFU all refer to the same controlled text.
- A warning in the IFU must not only be found; it must be linked to its risk, its control measure, the evidence that users understand it, the product version, the language and the released document state.
- AI alone does not make a device high-risk under the AI Act: it takes a device or safety component that needs a notified body, and then one single set of technical documentation covers both (Art. 11(2)).
- Model version, software version and device version are three things; a planned model change is described in advance and checked against that plan.
- An AI function may answer from the file only within released states, permissions and named sources — the released document stays the authority.
Where do you start?
Pick the sentence closest to where your organization stands; it leads to the first station that matters for you.
- We are writing the IFU for a new deviceStation: Intended purpose as the anchor
- A warning changed and nobody knows what else it touchesStation: Risk-to-IFU traceability
- Our device gets an AI functionStation: The AI Act check
- We want an assistant to answer from the technical fileStation: Controlled retrieval from the file
- A complaint trend has come inStation: Post-market signals into the file
Starting points, five chapters and short tasks with solutions — in a window you can move and resize while the map stays open.
The map
Lanes run left to right; each route is numbered and drawn with its own line pattern. Stations of the shared trunk are framed.
Lanes
Regulatory frameworks
Documentation systems
AI-enabled devices
Information architecture
- Regulatory frameworks
- MDR, IVDR, FDA, AI Act, ISO 13485, ISO 14971, IEC 62304, IEC 62366-1 — the rules each part of the file answers to.
- Documentation systems
- Technical documentation, design controls, traceability, clinical evidence, labeling and IFU, PMS, vigilance and CAPA.
- AI-enabled devices
- Software as a medical device, AI-enabled functions, model change, data provenance, oversight and performance monitoring.
- Information architecture
- Controlled terminology, evidence relationships, versions and variants, retrieval permissions and source authority.
Routes
One route per reader role. Each stop says what the role does there; every route runs through the shared trunk.
Route 1
Technical writers
Technical writers and information developers
You write the IFU, the labels and the safety messages — and you are the first to notice when they no longer match the risk file.
- Intended purpose as the anchorShared trunk
Quote the controlled intended purpose in the IFU instead of rewording it; flag every place where a draft says more.
- Risk management file
Ask for the ids of the risk control measures that end as information for safety.
- Risk-to-IFU traceabilityShared trunk
Give each safety-related statement the id of its control measure and keep the link through every language version.
- Oversight and limitations in the IFU
Write the performance, limitations and “when not to rely on it” sections for the people who actually use the output.
- Controlled terminology
Use the term base for alarms, warnings and messages; report terms the software uses differently.
- Variants, reuse and their boundaries
Reuse a module only where its risk control and evidence apply to the variant.
- Release with product and document versionShared trunk
Check that the release record names the product and software version, the languages and the evidence revisions.
- Post-market signals into the file
When a CAPA names an IFU change, update it, record the reason and keep the earlier state.
Route 2
Technical marketers
Technical marketers and product managers
You make the device understandable to buyers and users — with claims that stay inside the intended purpose and the evidence.
- Intended purpose as the anchorShared trunk
Start every claim from the controlled intended purpose; what it doesn’t cover stays out of the brochure.
- Claims as controlled information
Enter each claim in the claims matrix with its evidence and the markets where it may be used; reuse only approved wording.
- Clinical and performance evidence
Check which clinical benefit or performance figure a claim rests on, and that it is the current one.
- Is the software a device?
Don’t call a device function a “wellness feature”: the qualification decides what may be said.
- FDA pathway
Keep US and EU claims apart where the authorizations differ.
- Risk-to-IFU traceabilityShared trunk
Make sure no marketing text weakens a warning or contradicts a limitation in the IFU.
- Release with product and document versionShared trunk
Tie each campaign to the product version and the released IFU it describes.
- Real-world performance
When real-world performance drops below a published figure, raise it before the next campaign does.
Route 3
Technical project managers
Technical project managers (incl. PLM, R&D and regulatory project leads)
You plan the work so that the evidence exists when the file is assembled — not afterwards.
- Regulation, class and route
Get the qualification, class and route decided before the documentation plan is written.
- The AI Act check
Record the AI Act check and plan the AI Act sections inside the one technical documentation, with the 2028 date in the plan.
- One quality management system
Plan with one quality system and the QMSR terms for the US; budget for both audits.
- Intended purpose as the anchorShared trunk
Put the intended purpose under change control from the first draft.
- Design controls and design history
Plan design reviews, verification and validation as milestones that leave records.
- Risk-to-IFU traceabilityShared trunk
Make the risk-to-IFU trace a deliverable with an owner, not a check at the end.
- Planned model change
Decide early whether model updates are planned (PCCP, predetermined changes) and what falls outside the plan.
- Release with product and document versionShared trunk
Release documents and product versions together; no release without the evidence revisions named.
Route 4
Documentation systems developers
Documentation systems developers
You build the information system the file lives in — the objects, links, versions and permissions — so that people and AI functions find the released evidence.
- Intended purpose as the anchorShared trunk
Model the intended purpose as one object with an id, a version and its uses.
- GSPR and requirements traceability
Configure the trace matrix so that GSPR, requirement, standard and verification are linked by id and version.
- Software lifecycle and SOUP
Connect software items, SOUP and SBOM entries to the requirements and risks they touch.
- Evidence relationships as a graph
Define the typed relationships (implements, controls, verifies, supports, states) and an impact report for changes.
- Risk-to-IFU traceabilityShared trunk
Make the link from risk control to IFU statement survive translation and reuse.
- Training data provenance
Give datasets ids and versions and link them to the model versions trained on them.
- Controlled retrieval from the file
Configure retrieval on released states only, with permissions, source and version in every answer, and a review step for generated text.
- Release with product and document versionShared trunk
Store the release record with product version, UDI-DI, languages and evidence revisions as data, not as a cover page.
Stations
Each station is a small pathway: where practice starts, where it can get to, what does not change, how mature it is as of the review date, and the evidence it leaves.
What the maturity labels mean
- Building block available
- The building block exists today — an established method, a standard practice or a common feature of the tools. You still have to set it up and run it in your organization; it is not a finished product and not proof of compliance.
- Product-specific
- Some products or platforms offer it, others don’t. Check what yours does before you rely on it.
- Draft or proposal
- Described in a draft standard, a proposal or an individual draft. It may still change or never be adopted.
- Research
- Proposed in research or shown in prototypes. Not ready for everyday operation.
Regulatory frameworks
Regulatory frameworks
Building block availableRegulation, class and route
- Before
- The team starts writing before anyone has settled whether the product falls under the MDR or the IVDR and which class it has.
- After
- Qualification, regulation, class and conformity assessment route are decided and recorded first; every document of the file refers to that decision.
- What does not change
- The class follows from the intended purpose and the classification rules, not from what the documentation would prefer.
- Evidence
- Qualification and classification rationale with the rules applied
- Medical Device Regulation (MDR)
- In Vitro Diagnostic Medical Devices Regulation (IVDR)
- Medical device (MDR)
- Device classification (MDR)
- Device classification (IVDR)
- Rule 11: software classification (MDR)
On routes:3 · Technical project managers
Regulatory frameworks
Building block availableOne quality management system
- Before
- EU and US documents live in separate structures; the US side still cites the former QSR sections.
- After
- One ISO 13485 quality management system carries both markets; procedures cite the QMSR (effective February 2, 2026) for the US, and the medical device file is the index per device type.
- What does not change
- An ISO 13485 certificate does not replace an FDA inspection or a notified body audit.
- Evidence
- Quality manual, procedures and the medical device file per device type
- ISO 13485
- Quality Management System Regulation (QMSR, 21 CFR Part 820)
- Medical device file (ISO 13485)
- Document control
On routes:3 · Technical project managers
Regulatory frameworks
Building block availableRisk management file
- Before
- Risks sit in a spreadsheet that nobody links to the requirements, the tests or the warnings.
- After
- Each hazard, hazardous situation and harm has an id; every risk control measure points to the requirement that implements it and the verification that proves it.
- What does not change
- Information for safety comes last in the order of risk controls: a warning never replaces a design measure that is possible.
- Evidence
- Risk management file with ids, control measures, verification and the overall benefit-risk conclusion
- ISO 14971
- Hazard
- Hazardous situation
- Harm
- Risk control measure (ISO 14971)
- Benefit-risk determination (MDR)
On routes:1 · Technical writers
Regulatory frameworks
Draft or proposalThe AI Act check
- Before
- Any AI function is assumed to make the device high-risk, and a second, separate technical file is started.
- After
- The check asks two things: is the AI system the device or its safety component, and does the device need a notified body? Only both together make it high-risk; then the AI Act information goes into one single set of technical documentation (Art. 11(2)).
- What does not change
- For devices, the AI Act high-risk requirements apply from August 2, 2028 (as amended by Regulation (EU) 2026/1744); the MDR and IVDR apply now.
- Evidence
- Recorded high-risk check with Art. 6(1) reasoning and the planned AI Act sections of the file
Background reading: MDCG 2025-6 on the interplay between the MDR/IVDR and the AI Act; harmonized standards for the AI Act are still being drafted. Context: MDR, IVDR and AI Act: one technical documentation, Provider, deployer, manufacturer: who is who.
- AI in medical devices (MDR, IVDR and AI Act)
- High-risk AI system (AI Act)
- Single set of technical documentation (AI Act Art. 11(2))
- Provider (AI Act)
On routes:3 · Technical project managers
Regulatory frameworks
Building block availableFDA pathway
- Before
- The US submission is assembled at the end from the EU file, and nobody checked whether the software is a device under US law at all.
- After
- The pathway (510(k), De Novo or PMA) and the device status of the software are decided early; the submission reuses the same requirement, risk and test ids as the EU file.
- What does not change
- An EU certificate says nothing about US marketing authorization, and the other way around.
- Evidence
- Regulatory strategy with pathway, predicate or classification and the reused evidence
Background reading: Background: the encyclopedia article EU and FDA documentation compared.
- FDA premarket pathways: 510(k), De Novo and PMA
- Predetermined change control plan (PCCP)
- Clinical decision support software (FDA)
On routes:2 · Technical marketers
Documentation systems
Documentation systems · Shared trunk
Building block availableIntended purpose as the anchor
- Before
- The intended purpose is written five times in five documents, each a little differently: the IFU, the label, the clinical evaluation, the risk file and the website.
- After
- The intended purpose is one controlled text with an id and a version; every other document quotes or references it, and a change of it starts an impact assessment.
- What does not change
- Wording the intended purpose more broadly does not widen the clinical evidence behind it.
- Evidence
- Controlled intended purpose with version history and the list of documents that use it
Background reading: Background: the encyclopedia article Technical documentation as an evidence system. Context: The intended purpose as the semantic anchor.
On routes:1 · Technical writers2 · Technical marketers3 · Technical project managers4 · Documentation systems developers
Documentation systems
Building block availableGSPR and requirements traceability
- Before
- The GSPR checklist lists document numbers; whether those documents still say what the checklist claims is not checked.
- After
- Each applicable GSPR points to the requirements that fulfil it, the standards applied and the verification records, by id and version — a trace matrix instead of a list of file names.
- What does not change
- A complete matrix shows where the evidence is; it does not show that the evidence is sufficient.
- Evidence
- GSPR checklist and trace matrix with ids and versions
- General safety and performance requirements (GSPR)
- Traceability
- Evidence chain
- Design controls (medical devices)
On routes:4 · Documentation systems developers
Documentation systems
Building block availableDesign controls and design history
- Before
- Design inputs, outputs and reviews are scattered across project folders; the design history is rebuilt for every audit.
- After
- Design and development planning, inputs, outputs, reviews, verification, validation, transfer and changes are recorded as they happen and indexed in the design and development file.
- What does not change
- Under the QMSR, DHF, DMR and DHR are practice terms: the regulation now refers to the ISO 13485 records; the work behind them stays the same.
- Evidence
- Design and development file (DHF in US practice) with review records
Background reading: Background: the encyclopedia article Design controls and design history.
- Design controls (medical devices)
- Design history file (DHF)
- Device master record (DMR) and device history record (DHR)
- Medical device file (ISO 13485)
On routes:3 · Technical project managers
Documentation systems · Shared trunk
Product-specificRisk-to-IFU traceability
- Before
- A warning in the IFU can be found by search, but nobody can say which risk it controls, which test showed that users understand it, or for which product version and language it was released.
- After
- Each safety-related statement in the IFU carries the id of its risk control measure; it links to the risk, the usability or verification evidence, the product version, each language version and the released document state.
- What does not change
- The link makes a gap visible; it does not decide whether a warning is the right control or whether users read it.
- Evidence
- Risk-to-labeling trace with ids, summative usability evidence and released language versions
Background reading: Background: the encyclopedia article From risk management to labeling.
- Risk control measure (ISO 14971)
- Labeling and instructions for use (MDR and IVDR)
- Warning message
- Usability engineering file (IEC 62366-1)
- Hazard-related use scenario
On routes:1 · Technical writers2 · Technical marketers3 · Technical project managers4 · Documentation systems developers
Documentation systems
Building block availableClinical and performance evidence
- Before
- The clinical evaluation report is written once for the certificate and then left alone while the IFU and the claims move on.
- After
- Clinical benefits, performance and safety endpoints are named objects; the clinical evaluation, the IFU, the SSCP and the claims refer to the same ones, and PMCF updates them.
- What does not change
- Linking claims to studies does not make the data sufficient; that is a clinical judgment.
- Evidence
- Clinical evaluation plan and report, PMCF plan, SSCP where required
Background reading: Background: the encyclopedia article Clinical and performance evidence.
- Clinical evaluation (MDR)
- Clinical benefit (MDR)
- Performance evaluation (IVDR)
- Post-market clinical follow-up (PMCF)
- Summary of safety and clinical performance (SSCP)
On routes:2 · Technical marketers
Documentation systems
Building block availableClaims as controlled information
- Before
- Product claims are written by marketing in brochures and on the website; nobody checks them against the intended purpose or the evidence.
- After
- Every claim is an entry in a claims matrix with its wording, the evidence behind it, the markets where it may be used and an approval; channels reuse the approved wording.
- What does not change
- A claim that is not covered by the intended purpose stays off-label, however well it is documented.
- Evidence
- Claims matrix with evidence references and approvals
Background reading: Context: From claims to evidence.
On routes:2 · Technical marketers
Documentation systems
Building block availablePost-market signals into the file
- Before
- Complaints and field data end in a PMS report; the IFU, the risk file and the claims are not touched until the next audit.
- After
- Each post-market signal is assessed against the risk file; where it leads to a CAPA or a change, the affected labeling, IFU and claims are named and updated with the change.
- What does not change
- Reporting deadlines for serious incidents run on their own clock; a documentation update never replaces a vigilance report.
- Evidence
- PMS plan and report or PSUR, trend analyses, CAPA records with the documents changed
Background reading: Background: the encyclopedia article Post-market evidence.
- Post-market surveillance (MDR and IVDR)
- Vigilance (MDR and IVDR)
- Periodic safety update report (PSUR)
- Corrective and preventive action (CAPA)
- Post-market signal (medical devices)
On routes:1 · Technical writers
Documentation systems · Shared trunk
Building block availableRelease with product and document version
- Before
- The IFU is released as a PDF; which software version, UDI-DI and risk file revision it belongs to is written in nobody’s system.
- After
- Every released document state names the product and software version, the UDI-DI where applicable, the languages and the revisions of the evidence it rests on; the release is a recorded decision.
- What does not change
- A release record proves who released what and when; it does not prove that the content is right.
- Evidence
- Release record per document state with product version, UDI-DI, languages and evidence revisions
- Release decision
- Document control
- Unique Device Identification (UDI)
- Evidence version
- Change control
On routes:1 · Technical writers2 · Technical marketers3 · Technical project managers4 · Documentation systems developers
AI-enabled devices
AI-enabled devices
Building block availableIs the software a device?
- Before
- The software is called a “wellness app” or a “decision aid” because that sounds lighter, and the label decides the documentation.
- After
- Qualification follows the intended purpose: in the EU MDCG 2019-11 and Rule 11, in the US the device definition and the criteria for clinical decision support software; the result and its reasoning are recorded.
- What does not change
- An AI function inside a device is part of that device: it doesn’t need its own qualification as a separate product to be covered by the file.
- Evidence
- Qualification and classification record for the software functions
- Software as a medical device (SaMD)
- Medical device software (MDSW)
- Rule 11: software classification (MDR)
- AI-enabled device software function (AI-DSF)
- Clinical decision support software (FDA)
On routes:2 · Technical marketers
AI-enabled devices
Building block availableSoftware lifecycle and SOUP
- Before
- The software documentation is a release note and a test report; third-party libraries are not listed.
- After
- Software development plan, requirements, architecture, verification, release and maintenance follow IEC 62304 by safety class; SOUP items and the SBOM are listed with versions and known anomalies; security activities follow IEC 81001-5-1.
- What does not change
- An SBOM lists components; it is not a vulnerability assessment.
- Evidence
- IEC 62304 deliverables per safety class, SOUP list, SBOM, security risk assessment
Background reading: Background: the encyclopedia article Software and AI lifecycle documentation.
- IEC 62304
- Software of unknown provenance (SOUP)
- Software bill of materials (SBOM)
- IEC 81001-5-1
- Cybersecurity
On routes:4 · Documentation systems developers
AI-enabled devices
Product-specificTraining data provenance
- Before
- The training data is described in a data scientist’s notebook: a folder name, a date and a sample count.
- After
- Each dataset used for training, tuning and testing is a versioned record: origin, inclusion and exclusion criteria, population and site characteristics, labeling procedure, split and the model versions that used it.
- What does not change
- A provenance record shows where the data came from; whether it represents the intended population is a separate, documented assessment.
- Evidence
- Dataset records with lineage, representativeness assessment and the model versions trained on them
Background reading: Context: Training data provenance.
On routes:4 · Documentation systems developers
AI-enabled devices
Draft or proposalPlanned model change
- Before
- A retrained model is released as a “minor update”; model version, software version and device version are used as if they were the same.
- After
- Model, software and device versions are kept apart; planned modifications are described in advance with their method and limits — in the US as a PCCP, in the EU as predetermined changes the notified body assessed — and every update is checked against that plan.
- What does not change
- A change outside the plan stays a change that needs its own assessment, whatever it is called.
- Evidence
- Modification plan or PCCP, change records per model version, impact assessment
Background reading: FDA final guidance on PCCPs for AI-enabled device software functions (December 2024); in the EU the practice for predetermined changes under the AI Act is still forming. Context: Planned model change (PCCP), Model, software and device versions.
- Predetermined change control plan (PCCP)
- Model versioning
- Change control
- AI-enabled device software function (AI-DSF)
On routes:3 · Technical project managers
AI-enabled devices
Draft or proposalOversight and limitations in the IFU
- Before
- The IFU says the software “supports the clinician”; what it cannot do, where it was not tested and when not to rely on it is missing.
- After
- The IFU names the intended users, the performance with its limits and the populations tested, the known failure modes and what the user does when the output looks wrong — written for clinicians and, where relevant, patients.
- What does not change
- Human oversight is a designed measure, not a sentence; telling users to “check the result” does not make them able to.
- Evidence
- IFU sections on performance, limitations and oversight measures, with the usability evidence behind them
Background reading: Context: Documenting model limitations.
On routes:1 · Technical writers
AI-enabled devices
Product-specificReal-world performance
- Before
- Performance is measured once, in the validation study; drift after launch is noticed through complaints.
- After
- Real-world performance is monitored against the validated figures per site and population; a drop is a post-market signal that reaches the risk file, the IFU figures and, where needed, the claims.
- What does not change
- Monitoring finds a drop; it doesn’t explain it or fix it.
- Evidence
- Monitoring plan, performance reports per model version, signal assessments
On routes:2 · Technical marketers
Information architecture
Information architecture
Building block availableControlled terminology
- Before
- “Alarm”, “alert”, “warning” and “notification” are used for the same thing in software, IFU and risk file.
- After
- One term base defines the terms of the device and its safety messages, with approved translations; symbols follow ISO 15223-1; software strings, IFU and risk file use the same terms.
- What does not change
- A term base keeps words consistent; it doesn’t decide what the device does.
- Evidence
- Term base with definitions, approved translations and owners
On routes:1 · Technical writers
Information architecture
ResearchEvidence relationships as a graph
- Before
- Relationships between requirement, risk, test, claim and IFU exist only as references inside documents.
- After
- The relationships are typed links between objects with ids — implements, controls, verifies, supports, states — so that a change shows every affected object across the file.
- What does not change
- A graph shows dependencies that people recorded; links nobody recorded are missing from it too.
- Evidence
- Relationship model and an impact report for a sample change
On routes:4 · Documentation systems developers
Information architecture
Building block availableVariants, reuse and their boundaries
- Before
- A warning written for one device variant is reused in another because the text is the same — although the risk is not.
- After
- Reuse follows the configuration: a module may be reused only where its risk control, its evidence and its product variant apply; the variant model says where.
- What does not change
- Identical text is not identical evidence.
- Evidence
- Variant model and reuse rules with the evidence scope of each module
On routes:1 · Technical writers
Information architecture
Product-specificControlled retrieval from the file
- Before
- An assistant answers questions from every PDF in the shared drive — drafts, superseded versions and confidential parts included.
- After
- The assistant retrieves only released document states, within the reader’s permissions, names the source and its version in every answer, and generated text goes through review before it reaches a document.
- What does not change
- A grounded answer can still be wrong; the released document stays the authority, not the answer.
- Evidence
- Retrieval rules (released states, permissions, source authority), answer logs, review records for generated text
Background reading: Context: Retrieval permissions for technical files, From structured content to controlled evidence.
On routes:4 · Documentation systems developers
Interchanges
Where one lane depends on another.
Risk management fileRisk-to-IFU traceability
Information for safety is the last risk control; its id travels into the IFU.
Intended purpose as the anchorClaims as controlled information
Every claim is checked against the controlled intended purpose.
The AI Act checkOversight and limitations in the IFU
Where the AI system is high-risk, oversight and instructions for use get AI Act content — in the same IFU.
Planned model changeRelease with product and document version
A new model version is a new released state of the device documentation.
Post-market signals into the fileRisk management file
Post-market signals update the risk file.
Real-world performancePost-market signals into the file
A drop in real-world performance is a post-market signal.
Controlled retrieval from the fileRelease with product and document version
Only released states are retrievable.
Variants, reuse and their boundariesRisk-to-IFU traceability
A reused warning carries its risk control only where the variant shares the risk.
Training data provenanceClinical and performance evidence
The validation data belongs to the clinical evidence, not only to the data science record.
The post-market cycle
- Collect
- Assess
- Act
- Change
- Update
- Report
What each step means for the documentation
- Collect: complaints, field data, literature, real-world performance and PMCF results, per product version.
- Assess: each signal against the risk file, the clinical evaluation and the claims.
- Act: corrective or preventive action where the assessment calls for it.
- Change: the change goes through change control and, where needed, the notified body or the FDA.
- Update: the affected IFU, labels, claims and summaries are named, updated and released with the product version.
- Report: PMS report or PSUR, trend reports and vigilance reports, each on its own deadline.
Post-market surveillance as required by the MDR (Articles 83–86) and the IVDR (Articles 78–81), with corrective and preventive action under ISO 13485:2016, clause 8.5 — established practice. The six steps and their reading for documentation are the editors’.
Limits of this map
The map orders the documentation work around the evidence chain of a medical device; it is not a regulatory strategy, a checklist for a submission or legal advice, and following it does not make a device compliant. The maturity labels are the editors’ assessment as of November 10, 2026: “available” means the building block exists as established practice or a common tool feature but still has to be set up and run. Classification, high-risk status, the conformity assessment route and FDA pathways depend on the individual device and its intended purpose. Dates and guidance were checked on the review date; guidance documents explain the law but are not law, and drafts may change.
The context
Why a map
A medical device's technical documentation is often treated as a set of documents to be complete. Notified bodies and the FDA read it as an argument: does the evidence support what the manufacturer says about the device? This map draws that argument as a network of four lanes — the regulatory frameworks, the documentation systems, the AI-enabled parts of a device and the information architecture underneath — and of the stations where they meet.
The lead example: one warning
Take one warning in the instructions for use: "Do not use the sensor on broken skin." In a well-kept file it can be found. In an evidence chain it is also linked to the hazardous situation it addresses in the risk management file, to the risk control measure it implements, to the summative evaluation that showed users understand it (usability engineering file), to the product and software version, to each language version and to the released document state. When the risk changes, the chain shows which sentence in which language has to change — and when someone asks an assistant about the sensor, the answer can name the released source.
Where AI comes in
AI touches the chain twice. It can be part of the device — then the AI Act check asks whether the AI system is the device or its safety component and whether a notified body is involved; only both together make it high-risk, and the AI Act information goes into the one technical documentation. And it can read the file — then retrieval must be limited to released states and to the reader's permissions, and generated text goes through review before it becomes documentation.
How to read it
Each station is a small pathway: before, after, what does not change, how mature the building block is and what evidence it leaves. The four routes follow the readers of the AI learning path; all of them pass the trunk — intended purpose, risk-to-IFU traceability and release with version. The post-market cycle closes the loop: signals from the field come back into the risk file, the IFU and the claims.
Questions readers ask next
- Does an AI function make a medical device a high-risk AI system?
- Not by itself. Under Article 6(1) of the AI Act, the AI system must be the device or a safety component of it, and the device must need a third-party conformity assessment by a notified body under the MDR or IVDR. For devices, the high-risk requirements apply from August 2, 2028 (as amended by Regulation (EU) 2026/1744).
- Do we need a second technical file for the AI Act?
- No. For products covered by Annex I Section A, Article 11(2) of the AI Act requires a single set of technical documentation that holds the information of the AI Act and of the MDR or IVDR. The quality management aspects can be integrated into the existing system (Article 17(3)).
- Are DHF, DMR and DHR still required in the US?
- Since the QMSR took effect on February 2, 2026, 21 CFR Part 820 incorporates ISO 13485:2016 by reference and no longer defines these records itself. The terms remain common in practice; the records behind them follow ISO 13485 (design and development file, medical device file, production records).
- Is structured content (XML, DITA) enough to make the file AI-ready?
- It is a good start, not the goal. What an AI function needs — and what a reviewer needs — are the relationships: which risk a warning controls, which evidence supports a claim, which version a statement is valid for and whether it is released.
Sources
- Regulation (EU) 2017/745 on medical devices (MDR) — EUR-Lex, May 5, 2017Law in force: Binding legislation that applies as of the review date.
- Regulation (EU) 2017/746 on in vitro diagnostic medical devices (IVDR) — EUR-Lex, May 5, 2017Law in force: Binding legislation that applies as of the review date.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), as amended by Regulation (EU) 2026/1744; high-risk requirements for devices apply from 2028-08-02 — EUR-Lex, July 12, 2024Law adopted, not yet applicable: Adopted and published legislation whose obligations apply from a later date.
- MDCG endorsed documents, incl. MDCG 2025-6 (interplay between the MDR/IVDR and the AI Act) and MDCG 2019-11 (software qualification and classification) — European Commission, June 2025Final guidance: Guidance published in its final form by an authority or an official body. Guidance explains the law; it is not law itself.
- 21 CFR Part 820, Quality Management System Regulation (effective 2026-02-02) — eCFR, February 2, 2026Law in force: Binding legislation that applies as of the review date.
- Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions — U.S. Food and Drug Administration, December 2024Final guidance: Guidance published in its final form by an authority or an official body. Guidance explains the law; it is not law itself.
- ISO 13485:2016, Medical devices — Quality management systems — Requirements for regulatory purposes (cited under the MDR as EN ISO 13485:2016/A11:2021) — ISO, March 2016Harmonized standard (cited): A harmonized standard whose reference is published in the Official Journal of the EU; applying it can give a presumption of conformity for the requirements it covers.
- ISO 14971:2019, Medical devices — Application of risk management to medical devices (cited under the MDR as EN ISO 14971:2019/A11:2021) — ISO, December 2019Harmonized standard (cited): A harmonized standard whose reference is published in the Official Journal of the EU; applying it can give a presumption of conformity for the requirements it covers.
- IEC 62304:2006+A1:2015, Medical device software — Software life cycle processes — IEC, June 2015Standard: A published standard or specification. Not cited in the Official Journal as a harmonized standard, so it gives no presumption of conformity on its own.
- IEC 62366-1:2015+A1:2020, Medical devices — Part 1: Application of usability engineering to medical devices — IEC, June 2020Standard: A published standard or specification. Not cited in the Official Journal as a harmonized standard, so it gives no presumption of conformity on its own.
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.