FDA recently published a set of 10 guiding principles to inform the development of Good Machine Learning Practice.

MDDI Staff

May 23, 2022

5 Min Read
IMG_2022-5-23-130309.jpg
Image courtesy of JL / Alamy Stock Photo

Machine Learning has been an ongoing focus of FDA activity for some time. The first Industry Discussion Paper on the topic was launched back in 2019, followed by an Action Plan in 2021 aimed at offering guidance on the topic.

Off the back of these developments, FDA recently published a set of 10 guiding principles to inform the development of Good Machine Learning Practice (GMLP) in collaboration with Health Canada and the UK Medicines and Healthcare products Regulatory Agency.

According to FDA, these principles are “intended to lay the foundation for developing Good Machine Learning Practice that addresses the unique nature of these products” as well as “helps cultivate future growth in this rapidly progressing field.”

To better understand the real impacts of these guiding principles – and what they may not cover – we spoke to industry veteran Bill Betten, Director of Solutions, Medtech at S3 Connected Health.

Here’s what he had to say.

Interviewer: How will these guiding principles help manufacturers of medical devices that use artificial intelligence and machine learning (AI/ML)?

Betten: These guiding principles are useful for medical device manufacturers insofar as they codify what needs to be developed and how. They provide much-needed rigor and structure, and that’s no simple task given the wide set of conditions they cover and the speed of development in this area. It’s a good, very necessary, first step.

With that in mind, this guidance remains very high-level, and gaps exist which warrant further consideration.

Interviewer: What actions can/should medical device manufacturers take now that FDA has shared these principles? Do you see any gaps for future resolution?

Betten: One of the key challenges now is for medical device manufacturers to be extremely clear on the methods of data capture or generation. We can’t just assume that data is gathered and analyzed by the same company, since frequently, it isn’t.

What’s more, it is highly likely that data from a variety of sources will be aggregated for analysis: you may have one analytic organization collecting and analyzing data from multiple manufacturers, meaning data normalization and temporal synchronization is important.

Ultimately, the veracity of data is key, and this veracity needs to be established and maintained all the way from generation through to transmission and storage. If the data cannot be trusted, then the analysis will be faulty. These principles need to apply to everyone, or every system, involved in the process.

Interviewer: Are there any surprises or new regulatory approaches in these principles?

Betten: There are no real surprises or approaches in the current principles. They’re general, and they target the same key areas of focus that many in the industry are already examining, namely: how do we gather real-world evidence and use it properly in a machine learning system? How do we create a patient-centric solution that’s relevant to the individual, in an environment where AI is based on broad data sets?

Again, the principles are a great first step, but the devil will be in the details and the challenge will be in the implementation.

Interviewer :FDA states that the “10 guiding principles identify areas where the International Medical Device Regulators Forum (IMDRF), international standards organizations and other collaborative bodies could work to advance GMLP. Areas of collaboration include research, creating educational tools and resources, international harmonization, and consensus standards, which may help inform regulatory policies and regulatory guidelines.” Could you explain the remaining work the IMDRF has to do in fact to advance GMLP?

Betten: The challenge of using the appropriate training data set has become more apparent lately, particularly regarding bias in the data. This remains critical to having a workable system. 

One of the ongoing issues yet to be addressed has to do with self-modifying code. Device manufacturers have historically performed rigid verification testing and configuration management, but the issues associated with confirming operation in the context of machine learning and AI are very different.

It’s a bit like having two identical twins raised in different environments: while the systems are identical, they will “evolve” differently depending upon their experience and exposure, and give different results. That presents some interesting challenges to the medical device manufacturer, particularly with regard to aftermarket support and troubleshooting.

Interviewer: FDA also states that it “envisions 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

  • Create new practices specific for medical technology and the healthcare sector”

Can you offer any examples of how the principles may lead to those achievements?

Betten: Other sectors can certainly provide some guidance here. Take social media, for example, an industry that understands very well how to mine data and make personal recommendations. Or facial recognition, which has highlighted the importance of better training data sets to improve performance.

At the same time, it’s important to remember that the significance of the decisions made in healthcare are far more critical than in other industries since they may involve literal life or death scenarios. This necessitates a need for discipline in data collection and processing that may not be matched in other sectors. The challenge is balancing this discipline with the flexibility to accommodate systems that must evolve.

An additional consideration is that to truly realize the promise of personalized medicine, we need to make data personal and relevant to the patient while finding a way to avoid the significant security and monetization challenges other industries face.

This will no doubt require a more general overhaul of healthcare infrastructure to incentivize prevention rather than treatment, and that in turn means business models around data collection and analysis in healthcare – still in their infancy – will need to better reflect financial rewards as well as improved patient outcomes.

At the end of the day, it’s still about turning data into insights, and then into action. We need an environment that’s secure but enables machine learning to drive improved healthcare and better outcomes. These guidelines are indeed a step in the right direction, but there’s more to be done.

Sign up for the QMED & MD+DI Daily newsletter.

You May Also Like