AI-based software has become increasingly prevalent in medical applications, however, it is not always simple to determine if such software falls under the definition of a medical device. We've had clients who had developed an AI product and, after doing some market research, realized that it could be considered a medical device. For example, a cancer diagnosis system for MRI imaging is a clear-cut case, but what if the software processes and summarizes medical data without a clear treatment recommendation?
In this article, we will discuss the legal challenges associated with AI-based software that may not be immediately recognizable as a medical device. We will explore the qualification guidance from the MDR and from the Medical Device Coordination Group (MDCG) to provide an understanding of the qualification of AI software as a medical device. We will provide concrete examples of software that do and don't qualify as medical devices and discuss practical advice on how to determine if your AI product is a medical device and what to do if it is.
Let's cut to the chase: The document that determines whether your AI product is a medical device or not is the Medical Device Regulation (MDR) 2017/745. In particular, see Chapter I, Article 2: the definition of the term "medical device".
(1) ‘medical device’ means any instrument, apparatus, appliance, software, implant, reagent, material or other article intended by the manufacturer to be used, alone or in combination, for human beings for one or more of the following specific medical purposes:
and which does not achieve its principal intended action by pharmacological, immunological or metabolic means, in or on the human body, but which may be assisted in its function by such means.
The following products shall also be deemed to be medical devices
Obviously, this definition covers a lot of cases that do not apply to AI-based SaMD. Here are the parts that interest us:
(1) ‘medical device’ means software intended by the manufacturer to be used for human beings for one or more of the following specific medical purposes:
The vast majority of AI-based medical devices fall under the terms of the definition I've marked in bold. This makes sense, as medical imaging is currently the primary application of AI-based medical software, and most use cases clearly fit this definition.
However, there is still some vagueness in the MDR's definition. For example, AI products are usually focused on processing information and displaying it to physicians. From the definition in the MDR, it sounds like displaying any kind of diagnosis-relevant information may automatically render my product a medical device - even if it only involves trivial information processing.
Another unclarity remains: what if a decision support system supports a clinician in forming a diagnosis without actually issuing one? Or, conversely, if an AI is used to process data in a population, and the insights are published in a medical paper, do these still classify as medical devices? In both cases, the product does not directly provide a diagnosis but may have medical implications.
Thankfully, there is more guidance available.
MDCG 2019-11 is a guidance issued by the Medical Device Coordination Group (MDCG). It provides more details on the qualification and classification of software as a medical device in the MDR and IVDR. The guidance covers the categories of software, the criteria used to classify software as a medical device, and the steps necessary to meet the requirements for qualification and classification. As opposed to the MDR, this document isn't legally binding. However, it serves as an official reference point for how to comply with the MDR.
Let's see if it offers anything useful for AI products - here are some snippets from chapter 3.1: Introduction to qualification criteria:
For example, “Simple search”, which refers to the retrieval of records by matching record metadata against record search criteria or to the retrieval of information does not qualify as medical device software (e.g. library functions).
Software which is intended to process, analyse, create or modify medical information may be qualified as a medical device software if the creation or modification of that information is governed by a medical intended purpose. For example, the software which alters the representation of data for a medical purpose would qualify as a medical device software. (e.g. “searching image for findings that support a clinical hypothesis as to the diagnosis or evolution of therapy” or “software which locally amplifies the contrast of the finding on an image display so that it serves as a decision support or suggests an action to be taken by the user”).
However, altering the representation of data for embellishment/cosmetic or compatibility purposes does not readily qualify the software as medical device software.
This document provides two considerations:
Unfortunately, there is no clear-cut point that distinguishes e.g. between embellishing and modifying medical information, leaving a large area of uncertainty. Nevertheless, this text provides an initial idea of what types of software are classified as medical devices. The next chapter in the document goes into a bit more detail.
The most useful guidance regarding AI qualification is found in chapter 3.3 of MDCG 2019-11. It comes in form of a decision tree, which we've embellished:
In particular, decision steps 3 and 4 are relevant to AI software.
Decision step 3 further clarifies the difference between data processing techniques. It includes storage, archiving, and communication in the list of activities that do not necessarily make software a medical device. This provides a definitive solution to the previous confusion: The definition in the MDR suggests that displaying any kind of diagnosis-relevant information may automatically make my product a medical device. However, this is not the case - trivial information processing such as storage, archiving, communication, simple search, or embellishment that delivers diagnosis-relevant information does not automatically make AI a medical device.
What about the other question we had earlier? What about cases in which the product does not directly provide a diagnosis, but may have medical implications, such as a population-based study using machine learning?
Let's look at decision step 4. It is described in more detail:
Decision step 4: is the action for the benefit of individual patients?
Examples of software which are not considered as being for the benefit of individual patients are those
which are intended only to aggregate population data, provide generic diagnostic or treatment
pathways (not directed to individual patients), scientific literature, medical atlases, models and
templates as well as software intended only for epidemiological studies or registers.
We've got our answer: No, not every AI that could have medical implications is necessarily a medical device.
However, it should be noted that there is a wide grey area in this regard. For example, if an AI system is used to search for and compile literature on a specific disease, it may not be considered a medical device. On the other hand, if the AI does the same task but looks for patients similar to the patient at hand, it is highly likely that this software would be classified as a medical device.
After reading this article, you should have a better understanding of when your AI product might be considered a medical device. If you are still unsure, please feel free to reach out to us for a free micro-consultation. We are happy to discuss your product and the qualifications needed to bring it to market.
If you have determined that your AI product is not a medical device: Congratulations! In some cases, our clients need assistance in formulating solid arguments to explain why their product is not regulated in order to reduce the regulatory risk for various stakeholders. If you need assistance with this, please do not hesitate to contact us.
If your AI product is a medical device, you may benefit from our upcoming article, "MDR classification of AI software", which will help you to determine the class of your AI product. We hope that this article will provide you with the resources you need to ensure the success of your AI product.