FAQ on AI in medical devices

The use of AI (artificial intelligence), in or as a medical device, is a growing phenomenon in the healthcare sector. This is because the techniques involved in AI are capable of using large volumes of data to generate generalising algorithms. These algorithms can find hidden trends in data with potentially great benefits for both patient treatment and health economics. This could for instance be an improvement or clarification of the diagnostic process achieved by means of decision support systems based on AI. AI can also help to reduce the workload of healthcare staff.

Interest in the field has grown in recent years, and both private and public projects and partnerships have commenced. To ensure the development of safe AI-based medical devices for patients, the Danish Medicines Agency has prepared this FAQ, the intention also being to answer general questions about AI used in or as a medical device. The FAQ provides information on relevant legislation and guidance material for the development of software-based medical devices, including AI. It also lists some of the legislative requirements to pay attention to in connection with the marketing of AI-based medical devices.

When is AI-based software considered a medical device?

Medical devices are not defined by the media or material that makes up the device but by its intended purpose. So, while the software itself could be a medical device, it could also be a component of a medical device. Like other software, AI-based software is classified as a medical device if it provides an effect in connection with, for example, diagnosis, prevention, monitoring, prediction, prognosis, treatment or alleviation of diseases for an individual.
The full medical device definition is found in article 2(1) of the medical devices regulation. Since the medical device definition is very broad, the European Commission has issued guidance for classifying software as medical devices. This guidance offers examples and explains the medical devices regulation, the legislation in force from 26 May 2021.

What types of AI may be used in medical devices?

The manufacturer of a software constituting a medical device in itself or being used in a medical device is responsible for using the correct AI for the task. Whereas expert systems and decision trees can be used for making diagnoses based on behaviour entered by the patient, clustering techniques might perhaps be better suited for measured patient values such as blood pressure and infection counts. In addition, it is important that the clinical performance and efficacy of the device can be documented. You can find more information on how to document clinical performance and efficacy in the Commission's guidance thereon or in one of the other questions and answers.

Who approves medical devices so they can be marketed?

Medical devices may be marketed after obtaining a declaration of conformity, documenting that the device fulfils the applicable legislation. If the medical device belongs to the lowest risk class (class I), the manufacturers may themselves fill out the device’s declaration of conformity, provided they comply with the legislation. The manufacturer may hereafter legally affix the CE marking to the medical device. However, if the device belongs to a higher risk class (class IIa, IIb or III), a notified body must, for example, approve the medical device, the device’s technical documentation and the manufacturer’s systems for handling medical device incidents involving patients. If the above matters conform to legislation, the notified body will sign a declaration of conformity for the product so that it can be CE marked and distributed in the EU.

What is a risk class of a medical device, and what risk class does AI-based medical devices belong to?

Risk classes of medical devices are used to categorise devices based on the risk of using it and the damage it may cause if the device does not function within the specifications. In principle, the risk classes of AI-based medical devices are no different from the risk classes that similar non-AI-based devices would fall under. The risk classification of software considered to be a medical device is found in rule 11 of the medical devices regulation (Chapter III on Classification Rules) and is specified in the European Commission’s guidance on qualification and classification of software (Section 4.2.1).

How is evidence on the efficacy of the software ensured for the medical device to obtain CE marking?

The performance of a medical device is established through clinical evidence and expresses the ability of a device to achieve its intended purpose. The evidence base for the clinical performance may be achieved through clinical evaluation of the product, including through using already established evidence in the area. The evaluation of clinical performance must be based on data of a sufficient amount and quality, which is to be assessed based on the device in question. This applies in particular to machine-learning techniques, and in this connection deep-learning techniques, in which the data volume and the quality thereof are already important factors.
In addition, it must be assessed if data used for the development of software are representative of the target group of the medical device. Clinical evaluation is an ongoing process within medical devices, and the manufacturer is responsible for ensuring that there is sufficient ongoing evidence that the device can achieve its intended purpose throughout the product's lifetime/time on the market. You can find more information on clinical evaluation and the considerations to make to ensure clinical evidence in the European Commission’s guidance on clinical evaluation.

What are the challenges Involved in the development of continuous learning AI-based medical devices?

Continuous learning systems (CLS) – systems that learn continuously through their use, thus adjusting their parameters and performance, must have their modifications approved by a notified body following each and every software update if the performance and safety of the device change. A neural network that changes its weights during use will in most cases also change the performance. CLS is therefore not ruled out but does pose practical challenges. This is because the notified body having issued the CE certificate for the medical device in question will have to approve every update to the algorithm that changes.
The medical device legislation rests on the fundamental principle of always being able to allocate the responsibility for device malfunctions. To illustrate, if the device is used according to the manufacturer’s instructions and a malfunction occurs during use, the manufacturer is responsible for the device and the malfunction. If, on the other hand, the user of the medical device uses the device contrary to the manufacturer’s instructions, the manufacturer cannot, in principle, be held responsible for a potential incident. It is therefore especially important for continuous learning systems that the input data have been clearly defined since the algorithm that the device is built on may change during use.

Where can I find relevant material on the development of software-based medical devices?

The legislation and guidance documents on medical devices are available on the websites of the European Commission and the Danish Medicines Agency. Guidance on medical devices and the medical devices regulation.

Who can I contact if I have questions about AI-based medical devices?

If, after reviewing these questions and answers on this page, you still have unanswered questions about AI in medical devices, we will happily respond to these. Please submit your question by sending an e-mail with “AI” in the subject field.