The EU AI Act and Its Application Under the MDR

At EU level, medical devices (including AI medical devices) are mainly regulated by the MDR. The MDR aims to ensure a high level of health and safety for medical devices while also supporting innovation.

Manufacturers should first assess whether their software meets the criteria for qualification as a medical device under these rules. If the software is classified as a medical device, it must comply with the MDR.

1. Regulatory overlap between the EU Artificial Intelligence Act (AIA) and the MDR

(1) Definition and scope of AI systems

Under the AIA, an AI system is “a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments”.

Three core criteria – all of which must be met – distinguish AI systems from traditional simple software systems or programming approaches:

1. Inference capability: able to reason and analyse based on input information;

2. Autonomy: able to operate without continuous human intervention;

3. Adaptiveness: able to adjust its operation in response to environmental changes or data feedback.

(2) Two application scenarios for AI systems

In practice, AI systems fall into two main categories:

• Standalone use: the system can operate on its own and is not directly subject to existing product safety legislation;

• Integrated as a component: the system is used as a product component, either physically embedded in the product (embedded component) or not physically integrated but providing core functions to the product (non-embedded component).

(3) When dual regulation applies

Whether an AI system falls within medical device regulation depends essentially on its own intended medical purpose and function:

1. If the AI application meets the criteria for qualification as a medical device, it must strictly comply with the MDR;

2. If the system also meets the AIA definition of AI, it is additionally subject to the AIA – i.e. “dual regulation”.

2. How AIA risk classification and the MDR interact

(1) The principle of independent risk classification

It should be clear that risk classification under the AIA does not change the risk class determined under the MDR. Although both follow a “risk-based” regulatory logic (the higher the risk, the stricter the requirements), their classification systems and assessment dimensions are independent and cannot replace or adjust each other.

(2) The AIA risk classification system

The AIA divides AI systems into four risk categories:

In addition, general-purpose AI (GPAI) systems have specific regulatory obligations, determined by three factors: whether the model is open source, the system’s computing power and the size of the user base.

(3) Scope of high-risk systems under the AIA

Under Article 6 of the AIA, high-risk AI systems are further divided into two groups:

Standalone AI systems listed in Annex III: these systems can operate on their own and are explicitly included in the high-risk list;

AI systems covered by Annex I Section A, specifically:

a. systems intended to be used as a safety component of a product, or which are themselves products, covered by the Union harmonisation legislation listed in Annex I Section A (such as the MDR);

b. systems containing a safety component under a., or which are themselves products under a., and which are required to undergo a third-party conformity assessment under Union harmonisation legislation before being placed on the market or put into service.

(4) Typical case studies

Case 1: MDSW that is itself a product. Such medical device software (MDSW) has its own intended medical purpose and itself constitutes an AI system – for example, software for automatic CT image segmentation used for the early diagnosis of a specific type of cancer. Here the AI system is the core medical device and can be used directly for diagnosis.

Case 2: MDSW containing a safety component with a medical purpose. The AI system acts as a safety-enhancing component that drives or influences the operation of a (hardware) medical device and has a medical purpose. Take an insulin pump system with glucose monitoring software: the AI algorithm analyses continuous glucose monitoring data, predicts glucose trends and adjusts insulin delivery to prevent adverse glycaemic events (such as hypoglycaemia and hyperglycaemia), directly providing medical safety assurance.

Case 3: Software containing a safety component without a medical purpose. The AI system acts as a safety component that only ensures the safe operation of the medical device, without a direct medical purpose – for example, an AI monitoring system that supervises the hardware performance of a surgical robot. Its function is to improve the reliability and safety of the robotic system (such as monitoring robotic arm precision and predicting equipment faults), but it is not involved in medical decisions (such as surgical planning) or treatment operations (such as controlling tissue cutting).

(5) Common misconceptions about risk classification

Misconception 1: all AI medical devices require third-party assessment under the MDR. This is not true. Some Class I medical devices remain Class I under the MDR even with integrated AI functions and do not require third-party conformity assessment.

For example, some AI medical devices used only to assist with record-keeping can be exempt from third-party assessment if their risk does not exceed Class I.

Misconception 2: Class I AI medical devices will never be classified as high-risk under the AIA. There are exceptions:

If a Class I medical device (e.g. classified under MDR Rule 11) becomes higher risk due to integrated AI functions, its MDR class may be raised (e.g. to Class IIa);

Certain types of Class I medical devices (such as those with a measuring function, supplied sterile, or reusable surgical instruments) will still be classified as high-risk under the AIA if equipped with an AI system, even though they are Class I, and must meet stricter requirements.

3. Correspondence of operator roles under the AIA and the MDR

The two core operators recognised by the AIA are AI system providers and AI system deployers. This division of roles closely matches the categories of economic operators under the MDR, as follows:

4. Additional regulatory requirements for high-risk AI systems

The AIA makes clear that a single AI system may be subject to several pieces of Union harmonisation legislation at the same time. If a medical device with an AI system presents risks not fully covered by the MDR (such as algorithmic bias or data security vulnerabilities), the AIA, the MDR and other regulations must be applied in a coordinated way. To simplify the process, manufacturers of AI medical devices can integrate AIA requirements into their MDR submission documents, achieving “one submission, dual compliance”.

(1) Quality management requirements (4.1)

Providers of high-risk AI systems must establish and maintain written compliance policies, procedures and instructions to ensure that compliance is traceable and verifiable throughout the system’s life cycle.

(2) Risk management requirements (4.2)

The development and deployment of AI systems may face a range of unique risks, including bias in medical decisions caused by algorithmic bias, privacy breaches caused by data security vulnerabilities, difficulty in assigning responsibility due to insufficient system transparency, and performance fluctuations caused by faulty model updates.

Under Article 9 of the AIA, providers must “establish, implement, document and maintain” a dedicated risk management system. The industry currently refers widely to BS/AAMI 34971:2023 (Guidance on the application of ISO 14971 to artificial intelligence and machine learning), which provides a concrete path for bringing AI systems into an ISO 14971-based risk management framework.

(3) Data and data governance requirements (4.3)

Article 10 of the AIA is one of the core compliance requirements for high-risk AI systems. Its data governance measures must be implemented together with several EU data regulations and cannot be applied in isolation. Key regulations to coordinate with include:

The General Data Protection Regulation (GDPR): governs the collection and use of personal data;

The Data Act: sets out rules on data sharing and access;

The Data Governance Act: establishes a data governance framework;

The forthcoming European Health Data Space Regulation (EHDS): specific regulation of health data.

(4) Post-market monitoring system requirements (4.4)

In addition to MDR Articles 32, 61 and 84, the Medical Device Coordination Group (MDCG) requirements on post-market surveillance and vigilance and ISO/TR 20416, Article 72 of the AIA additionally requires providers to actively and systematically collect data on the performance of the AI system (such as algorithm accuracy and operational stability).

Manufacturers can integrate the additional AIA post-market monitoring (PMS) requirements into their existing MDR PMS framework, avoiding the duplication of monitoring systems.

(5) Serious incident and malfunction reporting requirements (4.5)

1. Definition of serious incident

The AIA defines a “serious incident” as “an incident or malfunctioning of an AI system that directly or indirectly leads, might have led or might lead to any of the following”:

a. the death of a person, or serious harm to a person’s health;

b. a serious and irreversible disruption of the management or operation of critical infrastructure;

c. the infringement of obligations under Union law intended to protect fundamental rights;

d. serious harm to property or the environment.

2. Reporting deadlines and procedure

Providers must report to the market surveillance authorities of the Member State where the incident occurred, with deadlines depending on severity:

General serious incidents: within 15 days of establishing, or suspecting, a causal link between the AI system and the incident;

Death: “as soon as possible” and no later than 10 days after the incident is established;

Widespread infringements or serious incidents involving critical infrastructure: “immediately” and no later than 2 days after the incident is established.

(6) Technical documentation requirements (4.6)

Under Article 11 of the AIA, providers of high-risk AI systems must prepare and maintain up-to-date technical documentation containing at least all the elements set out in Annex IV of the AIA.

Retention and provision requirements:

Retention period: at least 10 years after the system is placed on the market or put into service;

Obligation to provide: the complete documentation must be provided immediately when requested by national competent authorities.

Medical device manufacturers should supplement their existing MDR technical documentation with the information required by AIA Annex IV, so that the documentation meets both regulations.

(7) Product requirements (4.7)

The MDR’s General Safety and Performance Requirements (GSPR) do not fully address the unique challenges of AI technology (such as model drift and algorithmic black boxes). The AIA therefore adds specific requirements for AI medical devices, with the core aim of protecting people’s health, safety and fundamental rights and filling the gaps in the MDR regarding AI regulation.

(8) Automatic logging requirements (4.8)

High-risk AI systems must be designed to “automatically record events (logs)”, and the logs must support the identification of risks at national level and substantial modifications throughout the system’s life cycle.

Under Article 12 of the AIA, logs must record three types of information in particular:

a. situations in the AI system’s behaviour that may give rise to risks or harm, and substantial modifications of the system;

b. information supporting post-market monitoring to ensure the system continues to operate safely (serving post-market monitoring needs);

c. information for monitoring the system’s operation to verify that it complies with safety requirements throughout its operating life.

Note that while the MDR has clear provisions on post-market surveillance and vigilance, it does not require “automatic logging”. This function is essential for long-term monitoring of AI system behaviour (such as changes in model performance and abnormal operation) and is a key new requirement of the AIA.

(9) Declaration of conformity requirements (4.9)

Providers must draw up an EU declaration of conformity for each high-risk AI system, meeting the following requirements:

Form: machine-readable, and signed on paper or electronically;

Retention: kept at the disposal of national competent authorities for at least 10 years after the system is placed on the market or put into service;

Content: must contain all the information set out in AIA Annex V (including a statement of GDPR compliance);

Integration: for AI medical devices, the AIA allows a “single declaration of conformity” covering all applicable legislation, including the AIA and the MDR.

(10) CE marking requirements (4.10)

AI system providers must affix the CE marking to the medical device itself, its packaging or accompanying documentation, together with the identification number of the notified body responsible for the conformity assessment.

For AI medical devices, a single CE marking indicates that the product complies with both the AIA and the MDR; separate markings are not needed.

5. Conformity assessment routes

The AIA allows high-risk AI systems regulated by Union harmonisation legislation (such as the MDR) to integrate the specific AIA requirements into the existing conformity assessment procedure of that legislation, without a separate AIA conformity assessment.

(1) Principles for choosing the assessment route

The choice of conformity assessment route depends essentially on the specific Union harmonisation legislation applicable to the high-risk AI system. None of the following routes applies to AI systems not classified as high-risk.

(2) The specific route for AI medical devices

Of the four assessment routes, Route 4 is the specific compliance route for high-risk AI medical devices and must strictly follow these requirements:

Assessment in accordance with Annex I Section A;

Meeting the specific “AI testing” requirements of points 4.3, 4.4 and 4.5 of Annex VII and Article 46(5);

A notified body must be involved in the assessment to ensure the compliance and authority of the results.