Diagnostic support algorithms, AI scribes, recommendation LLMs, DTx… How do you obtain CE marking for a Medical Device integrating AI? Classification, compliance process and interplay with the AI Act: the complete overview.
Digital Medical Devices (SaMD, Software as a Medical Device) integrating Artificial Intelligence (AI) or LLMs (Large Language Models) are multiplying in healthcare, which is good news for the efficiency of our healthcare system and for patients.
What regulatory requirements currently apply to these AI-based SaMD?
In this article, we present concrete examples of AI-based SaMD and their classification under MDR 2017/745 or IVDR 2017/746, as well as the process to achieve compliance.
Among AI-based Medical Devices, the most common are:
Diagnostic support algorithms, notably in radiology (46.9% of AI-based devices approved by the FDA¹), oncology (25%¹), neurology (15.6%¹), dermatology or genomics.
➡️ Classification (a priori): in vitro diagnostic medical device under IVDR 2017/746, often risk class C or D.
Patient prioritisation systems in hospital departments or primary care, using risk scores to recommend care, medication dosages or vaccination reminders.
➡️ Classification (a priori): MD under MDR 2017/745, class IIb.
Insulin dosing applications
➡️ Classification (a priori): MD under MDR 2017/745, class IIb.
Prescription support software (known in France as LAP) assisting physicians in writing prescriptions, based on drug databases.
➡️ Regulatory status (a priori): it is indeed an MD according to the HAS (French National Authority for Health) under MDR 2017/745, class IIa or IIb.
Gamified patient applications (DTx), offering a personalised, evolving rehabilitation or follow-up programme based on the patient's profile and progress.
➡️ Classification (a priori): MD under MDR 2017/745, class I or IIa.
Self-diagnosis applications, notably for mental health, dermatological or other conditions.
➡️ Classification (a priori): MD under MDR 2017/745, class IIa.
Some AI software still sits in a borderline area and is not always CE marked or FDA approved. A few examples:
➡️ Regulatory status (a priori): it is not an MD as long as the tool provides general information, equivalent to that of a textbook or an official guideline, without applying it to an individual case. It becomes an MD under MDR 2017/745 (class IIa or higher, Rule 11) as soon as its answer concerns a specific patient and guides a diagnosis, a treatment or the decision to see a doctor.
➡️ Regulatory status (a priori): the software becomes an MD under MDR 2017/745 (class IIa minimum, Rule 11) as soon as it sorts, interprets or creates medical information used as a basis for therapeutic decisions (for example, clinical summaries in the patient record that will be used to make decisions in future consultations).
⚠️ European authorities do not all take the same view on this subject:
In Sweden, the Medical Products Agency (Läkemedelsverket) considers that a system which does not merely record and transcribe the exchange word for word, but which uses AI to process and compile patient-specific medical information subsequently used for care decisions, is likely to fall under medical device regulation.
In the United Kingdom, by contrast, the MHRA published on 29 July 2026 guidance on AI scribes which in practice reverses the position of NHS England, according to which tools summarising exchanges between patients and clinicians were MDs. For the MHRA, products intended solely for transcription, consultation summaries, letter drafting or suggesting clinical codes subject to clinician review are not MDs. However, a product remains an MD if it suggests diagnoses or treatments, or if it triggers an automatic action (prescription, referral) without prior review by a clinician.
➡️ Regulatory status (a priori): as soon as the application uses a specific patient's data to direct them towards a consultation or not, it provides information used for a decision and falls under MDR 2017/745 (class IIa minimum, Rule 11). Class IIb may be justified if incorrect guidance could lead to a serious deterioration of health, for example by delaying care in an emergency.
As soon as software integrating AI is a Medical Device, CE marking is mandatory.
The assessment process relies on the following steps:
1️⃣ Setting up a Quality Management System (QMS) compliant with ISO 13485 and MDR 2017/745 or IVDR 2017/746.
The QMS will also need to meet the requirements of Article 17 of the AI Act, specified by the EN 18286:2026 standard. Published in July 2026, it has not yet been cited in the Official Journal of the EU.
2️⃣ Building the Technical Documentation (TD) demonstrating the safety and performance of the API/Module(s)/Software, in line with European regulatory requirements.
3️⃣ Compliance with the AI Act (Regulation 2024/1689), with progressive alignment between the requirements of the AI Act and those of MDR 2017/745-IVDR 2017/746.
4️⃣ Ask your Notified Body (NB) whether it plans a designation under the AI Act, as none has been designated to date.
At Sparta Care, we support you with:
Contact us to find out more!
¹ Source: Benjamens, S., Dhunnoo, P. & Meskó, B. The state of artificial intelligence-based FDA-approved medical devices and algorithms. npj Digit. Med. 3, 118 (2020). https://doi.org/10.1038/s41746-020-00324-0
² https://www.gov.uk/government/publications/ambient-voice-technology-enabled-products
