2 October 2019
Machine learning promises to change the way we research, the way we diagnose and ultimately treat patients, but at the same time it poses new challenges for the regulation of medical devices.
Machine learning is a form of artificial intelligence that uses algorithms to allow a system to learn from data. There are many potential applications of machine learning for health, for example apps (software) that help health professionals to diagnose fractures more accurately from radiograph images, or that can be used by consumers to detect abnormal heart rhythms via their mobile phones.
Many machine learning applications will be regulated as medical devices. This is not only important for developers to be aware of, but also poses important questions around precisely when a digital health tool qualifies as a medical device, how machine learning systems will be regulated as medical devices, and what special problems machine learning poses for regulation.
Our new report, Algorithms as medical devices addresses these questions and offers pragmatic recommendations for policymakers and regulators. The report considers how digital health devices are likely to be regulated by the Medical Devices Regulation (MDR), In Vitro Diagnostic Medical Devices Regulation (IVDR), and associated harmonised standards.
The first section of the report outlines what medical device regulation is, how software fits into such regulation, and why such regulation matters.
The digital health market is expanding in both volume and variety. As software is an integral element of increasing numbers of digital health tools, this market shift will bring new regulatory challenges as increasing numbers of devices fall within the scope of regulation. For example, proper evaluation of software as a medical device will in many cases require new techniques, and new skillsets for regulators.
New market participants from the software industry may find that products are subject to medical device regulation, often for the first time. The report details how devices will qualify as a medical or in vitro diagnostic medical device in the UK and EU under the MDR/IVDR once they become fully operational in 2020 and 2022 respectively, and compares this position to the US method, evaluating the significance of these changes.
This evaluation is becoming less straightforward because the line between what qualifies as a medical device (and is therefore regulated as such) and what counts as a life-style or well-being device is increasingly hard to determine. The manufacturer’s ‘intended purpose’ assists in distinguishing between the two, but is more and more being influenced by the function and risk that the device poses.
In the future regulatory environment, the potential risks posed by the device will become increasingly important in determining ‘intended purpose’ and ‘specific medical purpose’, something that developers will need to bear in mind.
The report identifies two special (though not unique) issues that the machine learning paradigm poses for medical device regulation, and how such regulation might fit with this new generation of devices. The first is that machine learning models are ‘human uninterpretable’ - that is, it can be difficult to understand why the model produced the output or conclusion it did. No harmonised regulatory standard directly addresses this issue, even though human interpretability may be relevant when assessing the safety, effectiveness, and risk of some devices.
Second, machine learning models adapt and retrain, changing over time. Such dynamic devices will struggle to comply with the change management processes that current medical device regulation contains.
Clearly, in these changing times, medical device regulation needs to keep pace in order to appropriately balance the potential health benefits of machine learning against the risks. Our new report questions whether existing regulations are sufficient to address these novel challenges, and surveys a number of regulatory strategies for medical devices in the EU context relevant to machine learning. Other resources focus on different regulatory issues. Together, these publications provide a guide for developers, regulators and policymakers to help underpin robust regulation of algorithms as medical devices.
For more information, including summary and recommendations, see: