Healthcare AI regulation: 5 key points on CE marking

Explore five key points on intended purpose, classification and CE marking for healthcare AI, and learn what the mark does and does not show.

Elegant abstract microscopic colonies, with cellular textures and microbiota patterns

CE marking for a healthcare AI tool depends on its intended purpose, whether it is considered a medical device and the applicable classification. It indicates conformity with the relevant requirements, but does not by itself guarantee clinical performance for every use, according to the European Medical Device Regulation.

In summary:

  • The product’s intended purpose and risks influence its classification and the conformity assessment required under the applicable framework.
  • Regulatory compliance may involve addressing data, risks, human oversight, traceability, cybersecurity and user information, depending on the relevant rules.
  • The European AI Act and medical device regulations may apply complementarily to medical devices that incorporate high-risk AI systems; they are not equivalent.
  • Documentation on design, validation, changes, limitations and use helps explain how a clinical AI tool is developed and managed.

To guide decisions from the design stage, the first step is to define the intended use and assess which obligations apply to the product.

What does CE marking mean for a healthcare AI tool?

CE marking indicates that the manufacturer declares the product conforms to the applicable requirements and has followed the required conformity assessment procedure where applicable. It does not, by itself, establish that the tool is effective for every patient or suitable for every clinical decision (Regulation (EU) 2017/745).

To understand what CE marking means for software, check its intended purpose: what it was designed for, the context in which it is used and the function it serves. The Medical Devices Regulation links classification to the intended purpose and the risks inherent in the product. European guidance on medical device software also explains that the function of the information it provides in supporting care decisions matters (MDCG 2021-24).

Before adopting a clinical AI tool, it is therefore worth checking that its stated uses and limitations meet the needs of the healthcare organisation. For Sophia, Xenogene’s clinical AI, consult the available information about its role in supporting the interpretation of reports.

What rules may apply to AI used in healthcare?

AI used in healthcare may be subject to several frameworks at once: the European AI Act, medical device regulations if its intended purpose is medical, and data protection rules when it processes personal information. The specific obligations depend on its function and context of use (AI Act, Medical Devices Regulation and GDPR).

The AI Act does not automatically replace sector-specific rules. The European Commission explains that both frameworks apply alongside each other to medical devices incorporating high-risk AI systems (MDCG 2025-6 guidance).

When health data is processed, GDPR obligations must also be considered, including security measures appropriate to the risk of the processing. In clinical practice, healthcare professionals retain their duties to act and inform, and must keep records in accordance with Law 41/2002. Before implementing a tool, it is therefore advisable to review its intended purpose, the data it processes and who uses it together.

How are the risks of clinical AI assessed?

A risk assessment for clinical AI should identify what could go wrong, who could be affected and what controls could reduce harm. This includes reviewing the quality and representativeness of the data, validating performance, defining professional oversight and monitoring the system after it is put into service.

A practical review can be organised around four checks:

  • Data: review data quality and representativeness, check where the data came from and identify potential errors that could distort the results.
  • Use and performance: test the tool in its intended contexts and consider the consequences of an incorrect recommendation or use outside its intended purpose.
  • Integration: check that exchanges with other systems do not alter the data or make the result harder to interpret.
  • Control: define who reviews recommendations and how incidents and changes are recorded. For high-risk systems, the European AI Act requires the system to be designed to allow effective human oversight.

The review should continue after the system is put into service. If personal data is processed, security measures must take the risks of the processing into account, as required by the General Data Protection Regulation. Involving the professionals who will use the tool can help identify problems in the care workflow; they can visit the access page for healthcare professionals.

What evidence should the person responsible for a healthcare AI system retain?

The person responsible for a healthcare AI system should retain a file that makes it possible to reconstruct how the intended purpose was defined, which data and versions were used in development, how the system was validated and deployed, which limitations were identified, and which decisions were reserved for the healthcare professional. The documentation should be kept up to date throughout the system’s life cycle.

The scope of the obligations depends on whether the software is a medical device and whether the AI is subject to requirements for high-risk systems; each framework has its own criteria (Medical Devices Regulation and European AI Act). As a practical guide, the file should include:

  • Intended purpose and limits of use; users and context.
  • System architecture, data sources and versions used.
  • Validation methods and results, identified risks and controls applied.
  • Change history, relevant usage logs, incidents and post-deployment monitoring.
  • Instructions, responsibilities, training and decisions that remain with the healthcare professional.

If the system is integrated into a metagenomic workflow, see the good laboratory practices in metagenomics for further guidance on traceability and documentation.

How can a company develop AI tools without postponing compliance?

A company can develop AI tools without postponing compliance by incorporating regulatory requirements into the design: defining the intended clinical purpose, determining which frameworks may apply, assigning responsibilities and making any expansion of use conditional on verifiable evidence. Classification depends on the intended purpose and the risk involved (Medical Devices Regulation; European AI Act).

  • Define the intended use: describe who will use the system, in what context and for which decisions; do not present an application as validated if it has not been evaluated.
  • Turn requirements into tests: define acceptance criteria and check performance using data and scenarios representative of the intended use.
  • Prepare for oversight: train users so they understand the system’s limits and can intervene. For high-risk AI systems, the Regulation requires effective human oversight (European AI Act).
  • Control changes: review each update before deployment; a substantial modification may require a new assessment for high-risk systems that have already been assessed (European AI Act, consolidated text).

Frequently asked questions

Is complying with the EU AI Act the same as obtaining CE marking?

No. CE marking and compliance with the EU AI Act are separate matters, although they can be related. MDCG 2025-6 guidance explains that both frameworks apply simultaneously and complementarily to medical devices incorporating high-risk AI systems. Therefore, CE marking alone does not show that all the requirements of the AI Act have been met.

Who is responsible if an AI recommendation influences a clinical decision?

There is no automatic answer based solely on the AI having influenced the decision: responsibility depends on the circumstances and the applicable framework. Spanish Law 41/2002 sets out duties for professionals involved in healthcare. Using a tool therefore does not replace professional judgement or remove the need to review its recommendations.

Can an algorithm be modified or updated after obtaining CE marking?

Yes, but the regulatory implications depend on the change and the tool. The EU AI Act provides for a new conformity assessment when a substantial modification is made to high-risk systems that have already been assessed. Before deploying an update, it is therefore advisable to assess and document whether it changes the system’s intended purpose, operation or risks.

How should the limitations of a clinical AI tool be explained to patients?

They should be explained in clear terms, including what the tool does and that its output does not replace professional judgement. Spanish Law 41/2002 sets out information duties in healthcare, but the available sources do not specify a particular way to communicate the limitations of each AI system. The care provider should assess what information is appropriate in its healthcare context.

This information is general and does not replace an assessment by a healthcare professional.

Discover Sophia, Xenogene’s clinical AI for interpreting reports with professional support

Sources