Medical devices and artificial intelligence: FDA issues first guidelines for developing Good Machine Learning Practices

Thanks to Giulia Conforto for collaborating on this article

On October 27, 2021, the U.S. Food and Drug Administration (“FDA”), in collaboration with the Canadian regulatory agency, Health Canada, and the United Kingdom’s Medicines and Healthcare Products Regulatory Agency (“MHRA”), released a document containing 10 guiding principles to promote the development of Good Machine Learning Practice (“GMLP”).^[1] This document is currently open for stakeholder consultation.

This is an extremely innovative document, as it represents the first time that “good practices” are being developed specifically to be applied to the process of machine learning in the medical device sector. Accordingly, the 10 principles identified therein and the developments that will result from them are destined to provide very useful guidance to be employed worldwide.

Specifically, as detailed in the paper itself, these guiding principles may be used to:

• Adopt good practices that have been proven in other sectors;
• Tailor practices from other sectors so they are applicable to medical technology and the healthcare sector; and
• Create new practices specific to medical technology and the healthcare sector.

Scope of the guiding principles and regulatory framework

The 10 guiding principles follow a path paved a few years ago by the FDA, aimed at the adoption of a regulatory framework able to assist ongoing development of software classified as medical devices and the artificial intelligence algorithms on which many of them are based, while ensuring proper levels of safety and efficiency during the entire life cycles of such devices.

Indeed, it is worth recalling that technologies based on artificial intelligence (“AI”) and machine learning (“ML”) systems intended to be used for one or more medical purposes (i.e., to treat, diagnose, cure, mitigate, or prevent diseases) fall into the category termed “software as a medical device”   (“SaMD”) and are therefore subject to specific regulatory requirements for the purposes of their marketing and subsequent monitoring.

In 2019, the FDA issued a discussion paper outlining a legislative proposal for regulation of changes to AI/ML-based medical devices already on the market, with an eye to development of future guidance on the topic (Proposed regulatory framework for modifications to artificial intelligence/machine learning–based software as a medical device).

The 2019 paper touched on the concept of GMLP, emphasizing the importance of software manufacturers adopting proper standards of quality systems and good machine learning practice (GMLP) at their organizations for the purposes of clinical software validation.

With the aim of defining guiding principles that will lay the groundwork for the development of increasingly informed and appropriate GMLP, the FDA has therefore focused on drafting new guiding principles that should help to create safe, efficient, and high-quality SaMD.

Some of the guiding principles

The principles listed by the FDA include the following:

(i) good software engineering practices shall be continuously implemented to ensure device safety and data management quality;

(ii) the data sets used for a machine learning process should be as representative as possible of the patient population involved, in order to limit the risk of bias and prejudice;

(iii) the knowledge and experience of people and the human interpretability of a model’s output shall be put at the center of the machine learning process (human-in-the-loop approach;

(iv) users shall be provided with clear and relevant information, including complete information about the intended use of the product and directions for its use, the characteristics of the data used to train and test the model, as well as any modifications and updates to the devices.

As indicated above, the principles identified by the FDA include additional elements beyond the provisions governing SaMD in both North America and Europe, such as the human-in-the-loop approach, which also is designed to pursue primary goals and will therefore contribute to the development of best practices capable of reinforcing and incorporating the same provisions in a transversal and supranational perspective.

The upcoming steps

The SaMD sector and, in particular, the part of the sector creating SaMD based on AI/ML is expanding rapidly and is a top priority for global health regulation, since there are still many uncertainties at the regulatory level. At the European level, as well, this category of devices is the focus of great attention from a regulatory point of view, both in the new Medical Devices Regulation and in several interpretative guidelines aimed at supporting players in the sector issued in recent months by the Medical Devices Coordination Group (MDCG).

In this context, the FDA guidelines appear particularly important, not only due to their content, but especially because they trace a possible path for moving forward aimed at providing an increasingly solid and shared basis for the development of SaMD based on ML algorithms.

Moreover, the guidelines are designed to have significant impact at the international level. In this regard, the same paper calls for cooperation between the International Medical Device Regulators Forum (IMDRF), international standards organizations, and other collaborative bodies that could lead to further development of GMLP to help inform regulatory policy and regulatory guidelines. This collaboration is essential to facilitating establishment of common rules to ensure global uniform quality levels, as is already happening with GMP (Good Manufacturing Practice) and GCP (Good Clinical Practice) in other contexts (e.g., in the pharmaceutical sector).

Thus, these guiding principles represent only a starting point in the process of establishing good practices in this sector. The same paper states that the GMLP should grow along with the development and progress of AI/ML-based medical devices, which is expected to be very quick and significant in the next few years.

[1] The document is available at the following address: https://www.fda.gov/media/153486/download.