Clinical AI evaluation tiles representing documented, held-out medical AI test data.

Healthcare · · 8 min read

EU AI Act medical devices: what the 2028 deadline means for your data

Acumen Editorial · Updated November 20, 2026

Medtech teams got a reprieve this year. The Digital Omnibus on AI entered into force on July 27, 2026, according to RPC. It moved the high-risk obligations for AI built into regulated products, including medical devices, from August 2, 2027 to August 2, 2028. Standalone Annex III systems moved to December 2, 2027.

It is tempting to read that as time to wait. We read it differently. The rules that will govern your 2028 conformity assessment are already written, and the data you train on today is the data a notified body will ask about then.

The machinery is also moving. On October 14, 2026, the Official Journal referenced EN 18286:2026, the first harmonised AI quality management standard, according to a November regulatory roundup by Beyond Horizons. On November 16, the Commission opened consultation on a draft fundamental rights impact assessment template, with comments due January 15, 2027. The deadlines slipped. The detail is arriving anyway.

When does the EU AI Act apply to medical devices?

Most AI-enabled medical devices fall under Article 6(1). That covers AI that is a product, or a safety component of a product, under EU harmonisation law that requires third-party conformity assessment. For devices, that law is the MDR or IVDR, and the third party is a notified body.

Here is the current timeline.

DateWhat appliesWho it affects
August 2, 2025General-purpose AI model obligationsFoundation model providers whose models end up in devices
December 2, 2026Machine-readable marking of AI-generated content (for systems already on the market)Generative AI features, including patient-facing ones
December 2, 2027Annex III high-risk obligationsStandalone high-risk systems, such as some triage or eligibility tools
August 2, 2028Article 6(1) high-risk obligationsAI-enabled medical devices and IVDs that need notified body review

Until then, AI-enabled devices are certified under MDR and IVDR alone. DQS noted in March 2026 that no separate AI Act conformity assessment is currently required, and existing CE-marked devices remain compliant under current rules.

What does Article 10 require of training data?

Article 10 is the data article. It applies to training, validation and testing data sets for high-risk systems. It asks for three things.

Governance practices. Data must be managed in a way "appropriate for the intended purpose." The text lists what that covers: design choices, collection processes and data origin, preparation steps such as annotation and labelling, the assumptions behind what the data is meant to represent, an assessment of whether you have enough suitable data, an examination for biases, the measures you took against them, and the gaps you found.

Quality. Data sets must be "relevant, sufficiently representative" and, as far as possible, free of errors and complete for the intended purpose. They need appropriate statistical properties for the people the system will be used on.

Context. Data must reflect the geographical, contextual, behavioural and functional setting where the device will be used, to the extent the intended purpose requires it.

Read that list as a buyer. Annotation and labelling are named. So is data origin. Every label in your training set, and every reference answer in your test set, is part of the governance record. If an outside vendor produced them, their process is now your documentation.

How do the AI Act and the MDR fit together?

The Medical Device Coordination Group and the AI Board answered this in a joint FAQ, MDCG 2025-6 / AIB 2025-1, published in June 2025. It is not legally binding, but it is the clearest signal of how reviewers will think.

Three points matter for data teams.

  • One technical file. The FAQ says a single set of technical documentation should cover both the AI Act and MDR or IVDR. Data governance evidence sits next to your clinical evidence, not in a separate binder.
  • Representative populations. Training data should represent the intended patient population. The FAQ names age, sex, ethnicity and geography, and asks for performance checks across subgroups and over time, including drift.
  • Independent testing. The FAQ defines testing data as data used for "an independent evaluation" of the system. It says the Commission's guidelines will add detail on separating data sets.

This lines up with the US. FDA's January 2025 draft guidance on AI-enabled device software asks sponsors to describe how data was collected and processed, and how test data stays independent of training data. A dataset built for one regulator should be documented for both.

What changed for bias and health data?

Bias work needs sensitive data. You cannot check whether a dermatology model underperforms on darker skin without knowing skin type. Health data is a special category under GDPR.

The original Article 10(5) gave high-risk providers a narrow legal basis to process special-category data for bias detection and correction, with strict safeguards. RPC reports that the Omnibus extended that basis beyond high-risk providers to certain providers and deployers of other AI systems and models. That helps teams that want to test for bias earlier, before a product is formally high-risk.

It does not remove the safeguards. The conditions in the original text included pseudonymisation, strict access controls, no transfer to other parties, deletion once the bias is corrected, and documentation that the processing was strictly necessary. Plan for those in your data pipeline.

Checklist: is your training data ready for 2028?

Use this before your next data purchase or notified body conversation.

  • Can you state the origin of every training, validation and test item, including who created or labelled it?
  • Are annotation and labelling instructions versioned and stored with the data?
  • Were labels and reference answers written or reviewed by clinicians matched to the specialty and intended use?
  • Does the data reflect the intended patient population by age, sex, ethnicity and geography, and the care setting where the device will run?
  • Have you documented the assumptions about what the data represents, and the gaps you found?
  • Is there a written bias examination, with the mitigation steps you took?
  • Are test sets held out, never used in training, and documented as independent?
  • Can you report performance by subgroup and track it after launch?
  • If you process special-category data for bias testing, are pseudonymisation, access limits and deletion rules in place?
  • Can your vendor's documentation drop into a single technical file for both the AI Act and MDR or IVDR?

If a vendor cannot answer the provenance and labelling questions, the gap lands in your technical file.

How Acumen helps

We build clinical AI data for teams that have to defend it. Practicing physicians, matched by specialty, write cases, reference answers, rubrics and graded reasoning. A second reviewer adjudicates. Every item carries a record of who wrote it, under which instructions, who reviewed it and what changed.

We keep evaluation sets held out and documented, so test independence is something you can show, not assert. Specialty and population coverage are scoped against your intended use from the start. Our documentation is written to sit inside a technical file for a notified body or a data description for FDA.

See how it works on our Clinical AI Training Data page. For the privacy side of the same question, read What makes medical AI training data compliant.

Request a sample dataset Schedule a conversation

FAQ

When does the EU AI Act apply to medical devices?

High-risk obligations for AI-enabled medical devices and IVDs that need notified body review apply from August 2, 2028. The Digital Omnibus on AI, in force since July 27, 2026, moved that date from August 2, 2027. Until then, these devices are certified under MDR and IVDR alone. General-purpose AI model obligations have applied since August 2025.

Is an AI-enabled medical device automatically high-risk under the AI Act?

Not automatically. Under Article 6(1), AI is high-risk when it is a product or safety component covered by EU harmonisation law, such as the MDR or IVDR, and requires third-party conformity assessment. The MDCG and AI Board FAQ treats notified body involvement as the key condition. Devices that self-certify without a notified body fall outside this route.

What does Article 10 of the AI Act require for training data?

Article 10 requires governance practices for training, validation and test data: documented design choices, data origin, preparation steps like annotation and labelling, assumptions, a bias examination and mitigation, and identified gaps. Data must be relevant, sufficiently representative, as error-free and complete as possible, and suited to the setting where the system will be used.

Do we need separate documentation for the AI Act and the MDR?

No. The joint MDCG 2025-6 / AIB 2025-1 FAQ, published in June 2025, says a single set of technical documentation should cover both the AI Act and MDR or IVDR for high-risk medical device AI. Data governance evidence should sit in the same technical file as your clinical evidence. The FAQ is guidance and is not legally binding.

Can we use health data to test a model for bias?

Yes, within limits. The AI Act gives a narrow legal basis to process special-category data, including health data, for bias detection and correction. RPC reports the Digital Omnibus extended it beyond high-risk providers to certain other AI providers and deployers. Safeguards still apply, including pseudonymisation, strict access controls, no onward transfer and deletion when the work is done.

Should medtech teams wait now that the deadline moved?

We would not. The data requirements in Article 10 did not change, and datasets built today will be reviewed in 2028. Harmonised standards are starting to land, with EN 18286 referenced in the Official Journal in October 2026. Building provenance, representative coverage and held-out test sets now is cheaper than reconstructing them later.

From the Acumen archive. Explore expert data and Healthcare & Pharma.

← Back to Insights