EU Medical Device Reform Only Partly Fixes Rules for AI-Enabled Medical Devices
November 18, 2026


The European Commission’s December 2025 package to amend the Medical Device Regulation (MDR), the In Vitro Diagnostic Medical Devices Regulation (IVDR) and the AI Act would improve the rules for AI-enabled medical devices only in part, according to a Comment in npj Digital Medicine by Oscar Freyer, Koen Cobbaert, Anett Schönfelder, Erik Vollebregt and Stephen Gilbert.
The authors judge the changes generally well-targeted, but they say the package leaves specific challenges for innovative AI-enabled medical devices unresolved. The Comment reflects the situation as of 26 August 2026 and draws on the draft report of the European Parliament’s Committee on Public Health, which proposes further amendments.
What the MDR 2.0 proposal changes
The authors call the package the MDR 2.0 proposal. The Commission published it in December 2025 as COM(2025) 1023. Its stated aims are to eliminate double regulation, reduce documentation requirements and introduce new regulatory tools.
The reform responds to the core criticisms of the MDR. Critics describe the MDR as overly complex, with strict requirements even for low-risk devices and a heavy documentation burden. Overlap with other legislation, such as the AI Act, adds development and approval costs and lengthens review times at Notified Bodies, the independent organizations that assess conformity for devices above Class I. Manufacturers have reported cancelled or delayed launches and discontinued products. Some have stopped offering certain devices in the EU, especially orphan devices, and some say they are considering leaving the EU market. Temporary extensions of MDR deadlines have not fixed these problems.
The Commission’s own evaluation concluded that these problems limit patient access to innovative medical technologies in the EU and reduce competitiveness. A Personal View in The Lancet Regional Health Europe makes a related case for an EU medical device agency with binding powers.
How the reform affects AI-enabled medical devices
The table summarises the proposal’s main changes and the authors’ assessment, paraphrasing the Comment’s text.
| Issue | What the proposal does | Authors’ assessment |
|---|---|---|
| AI Act overlap | Moves the MDR and IVDR from Annex I Section A to Section B | AI Act Chapter III obligations would no longer apply directly to devices; the Digital Omnibus has taken a different route |
| Rule 11 classification | Makes medical device software Class I by default, with up-classification when its output informs or drives clinical management in increasingly serious situations | The Class I default is likely to stay exceptional or need interpretive guidance; the line moves to the Class I boundary |
| Clinical evidence | Counts scientific literature that may not be peer-reviewed, and allows non-clinical data, including in-silico testing | A meaningful expansion, with open questions on whether the evidence is robust enough for higher-risk devices |
| Innovative products | Expert-panel review, then priority and rolling reviews by Notified Bodies | Could become uncoordinated, because panel advice and conformity assessment are separate steps and the pathway is not linked to reimbursement |
| In-house devices | Relaxes transfer conditions; under the MDR, a ten-year transition applies if a commercial equivalent appears (the IVDR drops that requirement) | Weak gatekeeping for software, long use without Notified Body oversight, and a possible route to market for commercial software |
| Post-market change | Introduces predetermined change control plans (PCCPs) submitted with the application | A positive development, but it does not resolve the conflict between device rules and adaptive AI |
Source: Freyer et al., npj Digital Medicine (2026), key challenges and proposed amendments, paraphrased by Syenza News.
The AI Act overlap and the Digital Omnibus
At present, the MDR and IVDR sit in Section A of Annex I to the AI Act. For an AI device classed as high-risk under the AI Act, the full AI Act requirements then apply alongside the MDR’s general safety and performance requirements. These cover data governance, transparency, human oversight, accuracy, robustness and cybersecurity. The overlap affects every device that needs a Notified Body, usually Class IIa or higher, where the AI system is itself a medical device or meets the AI Act definition of a safety component. The December 2025 proposal would move the MDR and IVDR to Section B, so AI Act Chapter III obligations would no longer apply directly to medical devices.
Non-binding guidance from the AI Board and the Medical Device Coordination Group (AIB 2025-1/MDCG 2025-6) explains how AI Act requirements can be built into medical device conformity assessment. It does not remove the overlap, because meeting the integrated assessment does not remove the need to meet both frameworks.
The legislators took a different route in the Digital Omnibus on AI, Regulation (EU) 2026/1744, adopted in July 2026. It keeps the MDR and IVDR in Section A. It lets the Commission limit AI Act requirements where sectoral law gives an equivalent level of protection, and it delays the applicability of high-risk requirements to August 2028. The Commission must adopt the corresponding delegated acts by 2 August 2027. The authors note that the move of the MDR from Section A to Section B remains an option for targeted revisions.
Classification under rule 11
Under MDR rule 11, software that provides information for diagnostic or therapeutic decisions is typically at least Class IIa, which brings in a Notified Body. The proposal makes medical device software Class I by default and up-classifies it when its output informs or drives clinical management in increasingly serious situations.
The authors think little room is left for the Class I default. Under the proposed wording, informing clinical management in a serious or critical situation already results in Class IIa, and non-serious situations also fall under Class IIa. Rule 3.5, which assigns the highest applicable sub-rule, is unchanged and reinforces the effect. The line between informing and driving clinical management moves to the Class I boundary, where misclassification carries greater risk. The authors say the fault lines stay vague, which invites divergent interpretation and inconsistent decisions.
To partly address this, the proposal codifies the Helsinki procedure in MDR Article 4 and new Articles 4a, 51a and 51b. Notified Bodies and manufacturers can ask competent authorities to clarify a device’s classification in a dispute, and competent authorities can consult an expert panel.
The authors also say the rule rests on a narrow and dated view of software risk. It keeps the two dimensions of IMDRF N12 (2014): how much the software’s information matters to the healthcare decision, and how serious the healthcare situation is. The newer IMDRF N81 guidance (2025) also considers autonomy, explainability, the software’s role in the clinical workflow, the intended user and use environment, and how software changes are managed. In the authors’ reading, the intention was to achieve a substantial improvement, but the current wording and its interaction with rule 3.5 would have the opposite effect.
Clinical evidence
The proposal widens the definition of clinical data. The MDR currently draws clinical data from clinical investigations, peer-reviewed literature and post-market surveillance. The reform would also include studies published in scientific literature that may not be peer-reviewed. It is unclear whether preprints would count. The proposal also allows safety and performance to be shown solely with non-clinical data, and it encourages New Approach Methodologies such as in-silico testing.
The authors see a meaningful expansion of the evidence toolkit. For some AI devices, randomized clinical investigations may be impractical or ill-suited to measuring algorithmic performance. The trade-off is whether the resulting evidence will be robust enough to protect patients, particularly for higher-risk AI devices.
An expert-panel pathway for innovative devices
The proposal adds a pathway for AI-enabled devices that introduce highly innovative diagnostic or therapeutic methods with a substantial positive impact. Such devices either perform substantially better than existing devices or address unmet needs. After review by a designated expert panel, they would get priority and rolling reviews by Notified Bodies, plus access to scientific, technical and regulatory advice. A related Syenza News report on the expert panel advice pilot says the pilot ran under Article 61(2) of the MDR from February 2023 and closed in February 2026.
The authors’ concern is coordination. Panel advice and conformity assessment are largely separate steps, not one dialogue among manufacturer, panel and Notified Body. The European Parliament’s draft report partly addresses this by letting the Notified Body join the panel’s discussions. The pathway also stops at conformity assessment and is not linked to health technology assessment or reimbursement.
Selected Class IIb and III devices that meet the criteria in Article 7 of Regulation (EU) 2021/2282 may still go through EU joint clinical assessment. Member States keep responsibility for conclusions on added clinical value, coverage and reimbursement. A related Syenza News report covers the first EU joint clinical assessment report, which the Commission said appeared on 9 June 2026.
The authors say success depends on how the vague proposal text is implemented, on how broadly or cautiously panels read the designation criteria, and on whether Notified Bodies can review faster given existing capacity limits.
Regulatory sandboxes
The proposal adds regulatory sandboxes. These are time-limited environments in which manufacturers of innovative devices and regulators test new evidence-generation methods, regulatory requirements and enforcement practices, under a sandbox plan approved by the authority. For AI devices, sandboxes can provide a controlled environment for testing adaptive algorithms and new human-AI interaction models that do not fit existing frameworks. The AI Act already has sandboxes, but the AI Board and MDCG guidance and the Digital Omnibus contain no medical device-specific provisions.
The authors see open questions. The proposal does not say how MDR sandboxes will interact with the AI Act sandboxes, the draft General Pharmaceutical Regulation, the draft EU Biotech Act or local rules on clinical research. Sandboxes are also not instruments for large-scale authorization of new technologies, because they are time-limited and project-specific. Their value depends on whether regulators use what they learn to adapt the framework for all devices.
In-house devices and hospital-developed AI
In-house devices are manufactured and used within healthcare institutions and are not sold on an industrial scale (MDR Article 5(5)). Under current rules they face limited requirements and are exempt from some conformity assessment obligations, provided no equivalent commercial device exists and the device is not sold or transferred to another legal entity.
The proposal relaxes these conditions. Devices would be transferable to other health institutions when justified by the interests of public health, patient safety or patient health, or to prepare for or respond to a public health emergency. If a commercial equivalent appears under the MDR, institutions would get a ten-year transition before the in-house exemption ends. The IVDR drops that requirement entirely. The European Parliament’s draft report would go further by applying Annex I general safety and performance requirements proportionately.
The authors think this could lower barriers for hospitals and academic medical centers that build and share their own tools. The ten-year transition would also make it more viable to invest in local development, because an institution would have a guaranteed period of use.
They see three challenges. First, eligibility and transfer conditions are weak gatekeepers for software. Any medical device is intended to serve patient health, so the patient health ground for transfer imposes little real constraint. Software can be differentiated by adding or changing features, so the no-equivalent-device test is easy to meet. Second, the ten-year transition could keep in-house AI devices in routine clinical use for long periods without Notified Body oversight. Third, a commercial developer could supply software that a health institution places into service under its own responsibility and quality management system, reaching patients without its own conformity assessment.
The authors also object to the European Parliament’s draft removing the public declaration requirement under Article 5(5)(e), which they say reduces transparency.
Post-market changes and adaptive AI
Many AI systems are built to adapt in deployment as new data, feedback and user interactions arrive. Adaptation can offset shifts in patient populations, clinical practice and measurement conditions that gradually erode a static model’s accuracy. Under the current framework, such updates may count as a change that triggers a new conformity assessment under the MDR, or as a substantial modification under the AI Act. The documentation burden and cost often lead manufacturers to market devices with a static model.
The proposal addresses this to some extent by introducing predetermined change control plans (PCCPs), a mechanism the AI Act already allows. A PCCP is submitted with the application and covers predicted, predefined changes that can be made to an authorized device without a new authorization process. For AI devices, that permits some post-market changes, but only for changes that can be anticipated at submission, which is often not possible for innovative systems. Experience with PCCPs under other legislation, such as in the US, shows that their flexibility depends on what the regulator is willing to accept.
For a US view on the same questions, a Syenza News report covers the FDA’s discussion paper on generative AI medical devices, released on 18 August 2026 with a public docket for comment.
What the proposal leaves out
The authors say the reform does not fundamentally reconsider how the MDR handles continuous learning. PCCPs do not accommodate open-ended adaptation, such as diagnostic models that keep learning from new data, or agentic systems that query updatable databases such as PubMed. The proposal also does not address broad, orchestrating or agentic AI systems. These are increasingly common in medical research and are moving toward clinical deployment.
Still a proposal
The reforms are proposals, not law. The European Parliament and Council must still adopt the package, and amendments and revisions are probable. The authors expect the process to take at least a year, which makes adoption in 2027 or 2028 likely. Many operational details are delegated to legal instruments that do not yet exist.
In the meantime, the Digital Omnibus sets the current route. AI-enabled devices stay in Annex I Section A, and the Commission may limit equivalent AI Act obligations through delegated acts. The MDR reform may be delayed or not adopted, adopted after alignment with the Omnibus, or more substantially revised. Its effect will depend on MDCG guidance, common specifications, standards and the day-to-day practice of national competent authorities and Notified Bodies. Well-designed delegated and implementing acts applied consistently across member states could deliver the intended simplification. Inconsistent or overly cautious implementation could reproduce the fragmentation the proposal aims to correct.
Source: Freyer O, Cobbaert K, Schönfelder A, Vollebregt E, Gilbert S. The proposed reform of the Medical Device Regulation and its implications for AI-enabled medical devices in the EU. npj Digit. Med. 9, 907 (2026). Licensed under CC BY 4.0.
Let Google know we are your trusted source.
Add our editorial as a preferred source in your search results.
Value Science Weekly
The week in health technology assessment, market access and health economics — what moved, and what it means.





