Can AI shape the future of Regulatory navigation?

by Francesca Davis on 9 Sep 2024

Reading time: 4 minutes

AI is transforming our lives, reshaping industries, enhancing everyday experiences, and driving innovation. Currently, AI is embedded in technologies we use daily, from personalised recommendations to smart assistants. Looking ahead, AI's influence will only increase. Large language models (LLMs) is a type of AI that is designed to understand, generate, and interpret human language. LLM consists of algorithms that can take in text in order to generate new text based on statistical relationships. The most famous of these is ChatGPT, but there are many more LLMs available. They offer the opportunity to understand human text and provide responses to questions.

Jeroen Bergmann, Reg Metrics co-founder, recently published the scientific article: ‘Evaluation of large language models for the classification of medical device software’. I spoke with him to explore his research and find out in detail how large language models (LLMs) may be useful for MDR classification.

To really understand Jeroen’s research, I asked him: what is the primary goal of your research? Jeroen’s response was to understand how we can use data science for navigating regulations. The regulatory procedure makes sure that devices are safe and doing exactly what they are supposed to do. As this is something that is vital it needs to be done correctly and methodically.

One of the most important parts of Jeroen’s research is finding out how we balance the complexity of the regulations that are there for a reason, as well as the need to make it as accessible for companies developing medical devices to comply with the regulations. Finding this balance will allow researchers to bring their medical device to the market much easier, without removing any necessary parts of the regulation and making sure everything remains safe.

Jeroen hopes that what can come out of his research is “ trying to understand what might be able to be developed in the future in order to, hopefully, further improve not only the regulations but how people are using the regulations”

LLMs and device classification

Device classification is determined by how much of a risk a medical product may be to a patient. For example a plaster is a lower risk than a ventilator which is being used in an intensive care unit. The risk classification is different for certain medical products and this means that the regulatory pathways are different for those that are trying to build new systems or trying to innovate within that space.

But just why is risk classification such an important step in the regulatory strategy?

According to Jeroen it is the first thing you want to know about the device because it can help you understand how long it’s going to take to bring the device to the market. A medical device that is a higher risk may require more testing to ensure its safety. This can then be more expensive to regulate and take longer before that device is able to reach the market.

What Jeroen’s research explores is the possibility that a large language model could help navigating the medical device regulations. Large language models have become a popular tool in recent years, and could possibly assist the process of classifying devices. Jeroen thinks an LLM could really help a lot of manufacturers and he sees these models being really useful, if they are being well applied. At some point in the future we may see a development with appropriate large language models being used in the regulatory industry, but we could still be looking at a good few years before this comes really into play, as more testing is needed.

The negatives of using an large language model

The reason why we haven’t already implemented LLMs into medical device classification is because, despite there being a lot of possible uses for a LLM, there are also a lot of negative factors. These need to be taken into consideration before we begin using them for regulatory purposes. One challenge is that because the models are not specifically for medical device regulation, they might pull information from a range of different texts that might not be reflective of the real regulations. This could be dangerous if it is affecting how we classify risk because it can suggest incorrect risk classes to manufacturers.

Another big problem with large language models is the possibility that it might start hallucinating. If this is to happen you will get text generated that looks like it might be true but has just made it up out of the data it has. Jeroen added that, “we have to be very careful if this is to start happening, as although sometimes it may be obvious that the text you are getting is not accurate, other times you might assume that it is when in fact you are not getting the right information.” Again this could cause issues and Jeroen agreed with this, saying that we need to be sure that it is the right risk classification, because we are using the classification as a stepping stone for the regulatory journey. Before we can have a future using LLMs in device classification we need to ensure that the information being generated can be trusted and is accurate.

The future of using LLMs

The possibility of using LLMs for device classification in the future is definitely a possibility and at the very least the future of regulations will be digital across a lot of different fields. However, as Jeroen has already pointed out, it is important that we iron out all the potential challenges that we may face.

A way to try and do this is create test sets, and know what the outcome should be, we can then run different models on the test sets and see how well the models perform. This will help to establish which models can be trusted to generate accurate information and from this, there may be models that prove to be helpful and provide correct information consistently.

In times ahead, Jeroen thinks that there may be some dedicated models which will be developed for different fields, including device classification. Part of Jeroen’s future research will involve exploring what happens if we have more dedicated models to help with navigating this risk classification question. Looking into this concept may mean that LLMs could become more useful to the regulatory industry, as they will be made specifically for this purpose. Researchers like Jeroen are finding out how possible this will be and we will be looking out for any updates on how LLMs could shape the future of medical device regulation.