Artificial Intelligence / Machine Learning

It is almost summer and as ever, challenges in the regulatory environment define our everyday life. One year has already passed by since Brexit and the application date of the IVD regulation in May of 2022 is upon us. 
Another important topic is the current debate concerning legal and regulatory considerations on the application of artificial intelligence and machine learning (AI/ML) in medicine. This newsletter includes a short reflection on Artificial Intelligence and Machine Learning components in medical devices, written by one of our consultants. 

As technology evolves, new and more powerful tools can be implemented into medical devices and made available to healthcare professionals and lay users, for example to support image analysis, clinical decisions, guidance for self-treatment and life-style recommendations. Even though Artificial intelligence and Machine Learning (AI/ML) have been seen in decision supporting medical devices for decades, the technology has now been made widely available through smartphone apps, wearables, web-services, and other platforms.

Looking into the regulatory framework for AI, there are no specific regulations yet. There is an AI Act proposed by the EU commission (April 21, 2021). This proposed Act defines products regulated by MDR and IVDR as high-risk AI products. The proposed Act includes some overlaps and misalignments with MDR/IVDR, that need to be clarified, related to e.g. classification, certification and documentation. For AI/ML-based medical devices, MDR applies in Europe and QSR in the U.S. and there are ongoing activities in both markets to gain knowledge and to publish guidelines for AI/ML-based devices. As any medical device, AI/ML-based medical devices must be proven to be safe and effective for their use. The below highlighted aspects of quality assurance for AI/ML-based medical devices are only parts of all activities that need to be done.

Both the FDA and the proposed EU regulation emphasize the need of transparency of AI/ML-based medical devices, meaning e.g. transparency of algorithms, intended use, benefits, risks and limitations for the component. Transparency is nothing new to medical devices, but manufacturers will need to put high effort in describing the AI/ML component product thoroughly and indicate the presence of AI/ML components.

Due to disclaimers of some AI/ML-based devices not taking responsibility for clinical decisions, the intended purpose of AI/ML-based devices might be complicated and, if misunderstood, might lead to patient harm. Also, complex software might cause perception and cognition errors leading to use errors. Therefore, usability engineering is of high importance. FDA puts efforts in monitoring of devices from this perspective and has already issued information letters to healthcare providers related to intended use for certain AI/ML-based devices.

Training of AI/ML-based devices require a large amount of data, having a large impact on the outcome of the device. From a QA perspective, this data can be seen as design input. Therefore, the data needs to cover the entire intended use, it shall be relevant, unbiased, and specific. Data is also needed for verification and validation of the AI/ML component. This data must fulfill the same requirements as, and must not be redundant with, the design input data. To manage validation of training data, and ensure required change control of devices, training of the AI/ML component is preferably done before release. Quality assurance of an AI/ML component trained after release is a huge challenge, if even possible at all.