SVT – the Swedish national broadcasting company – highlights the impact of research on AI for breast cancer detection in mammography, conducted by MedTechlabs researchers Kevin Smith and Fredrik Strand. Also, Lancet published a scientific article on the study conducted at Capio Sankt Göran Hospital in Stockholm.
What differences in cancer detection and unnecessary recalls are the results of various combinations of AI and radiologists? Lancet Digital Health has published the results of a prospective clinical trial conducted in Stockholm.
MedTechLabs researcher Fredrik Strand is surprised at how accurate the AI was and says that the potential for AI to handle most of the screen-reading is huge.
Karin Dembrower is chief physician and medical director of the Mammography Department at Capio S:t Göran’s Hospital where the clinical study was conducted. She is very satisfied with the results.
– We found slightly more cancer and recalled fewer women who turned out to be healthy, so we made fewer women worried. And we freed up time for the radiologists to do other diagnostics than looking at mostly healthy women’s images. We have managed to cut the queues considerably so we are a queue-free breast centre now, she says in the feature.
The AI-assisted breast cancer detection system has been operational since summer 2023 at St Göran’s Hospital in Stockholm.
– We have shown that our AI technology works on St Göran and their equipment. But that doesn’t mean that any AI will work on all equipment, so it’s something you need to check before introducing it, says Fredrik Strand in the same feature.
Watch the tv-feature here (in Swedish)
Read the Article in Lancet Digital Health here (full text)
On 19 April 2023, MedTechLabs organised a seminar focusing on strategies for researchers to better identify and protect their digital innovations. Also on the agenda were tactics and practical tips on patenting or other ways to extract business opportunities from science linked to digital applications.
Niclas Roxhed, director of MedTechLabs, began by giving a background to MedTechLabs and its focus on interdisciplinary, translative research in later phases, proximity to clinical trials and clinical implementation, and industrial collaboration. He then gave the floor to Måns Marklund, IP strategist and founder of Cascelotte.
Måns then explained how he and his company have analysed results from researchers linked to MedTechLabs, examined the volume of patents and also made an in-depth study of their quality, potential economic value and technical standard. Compared to some of the world’s best research centres, these patents were found to be of a very high standard, especially those linked to Mats Danielsson and his research teams.
– Other research at MedTechLabs turned out to be more linked to digital applications and artificial intelligence, which may be one reason why this kind of research does not lead to as many patents. After discussions with Johan Schuber, Executive Director of MedTechLabs, we therefore decided to investigate this phenomenon more closely, said Mats.
How good is Sweden in international comparison?
Mats explained the importance of looking at the differences between different kinds of technologies. Normally, Sweden ranks 12th in the world in terms of all technology patents. However, many come from Ericsson, which inflates the volume. In life science alone, Sweden is ranked 15th, which is still very high, and an indication of how good Sweden is even without Ericsson.
– In AI, Sweden is ranked 20th in terms of patents, also very high considering we are a small country. But perhaps there is room for more patents, with increased knowledge of how this works and why patents should be granted in this area as well. We also know that if there are patents in an area, there are many other intangible assets there. It could be trademarks, knowledge of data and data use, design and so on. So, the patents indicate a country’s ability to transform innovation into economic value, Mats explained.
He went on to add that AI is also one of the most patented assets now, so it is not true that AI cannot be patented, even if it is slightly different from other technologies. This also applies to the life science area in general, compared to other patent areas.
– It is important to understand which strategies, tactics and practical approaches are good to know when it comes to both AI and life science. Therefore, we chose to invite Michael Kitzler and Malin Keijser Bergöö, who both work as lawyers specialised in European patent issues at Rouse, said Mats and handed over to these speakers.
Positions of control and the commercialisation of innovation
– In a commercial context there are some fundamental questions you need to ask yourself, like what you’re selling – sometimes it seems like you’re offering one thing when it’s actually something else. In the context we are talking about now, many of you are probably in an early phase of commercialisation, maybe you have created a company and you have something to sell to someone, Michael began.
He then described how Rouse usually starts by identifying the intellectual assets that the research team or company has built up over time and then translates these into different “control positions”, which he explained as the aspect of control that creates ownership and/or control over an innovation over time. Some of these can be patented, some may need to be kept secret until later, some can be contracted with other actors, and some are based on a unique combination of capabilities that are difficult for others to acquire.
He explained that different companies/teams need different mixes of control over assets and gave the example of a company that had 55 per cent of its assets covered by patents, 25 per cent by external agreements of some kind and 10 per cent by secrecy-based control. Software and IT companies tend to have fewer patents and are more likely to use other methods, while high-tech companies put a lot of effort into getting patents. Michael argues that medtech companies are somewhere between high-tech companies and software companies in terms of control positions.
– Looking at the trends in medtech, the statistics show that there are more patent applications in this industry than in others. So, is it worth the money to make these applications? Probably, said Michael.
The largest patent owners in medtech include companies like Siemens Healthcare, Philips, Fujufilm, Canon Medical Systems and Merative. Then there is a “long tail” of companies that may not have as many patents, but these can be very important. Michael gave the example of Elekta, with only 12 patents in the field of AI, but which are very significant. Even further down the list, there are companies with maybe four patents, which can be very relevant portfolios despite their small number. To summarise, it doesn’t take many patents to make a difference and achieve commercial success, it’s all about the content of those patents, Michael explained.
Michael then listed the commercial values that patent control positions create for medtech companies, from limiting business risks to contributing to the company’s participation in key ecosystems (see image below).
Patenting medtech and AI in practice
– Your field is one of the toughest to patent. Partly because it has to do with things you do with the help of computers, which the patent authorities are very careful about before approving, which makes it complicated and requires more work and the right strategies, Malin began.
She explained that in some countries there is a ban on patenting treatments, which can affect companies, and that you are then allowed to patent the devices you use, not the methods. This differs depending on where you are in the world; in China, for example, it is easier to patent treatments. In Europe, on the other hand, it is often impossible to patent treatment methods. In the US it is more unpredictable, where a single court decision can change the situation so that it suddenly becomes possible to patent something that was previously impossible, and vice versa. Malin emphasised that having said that, it is still often possible to patent – in some way. You may have to go through the equipment instead of the method, or do it in other ways, but it is always possible to patent something.
– One thing we saw when we looked at medtechlabs was projects in bioelectronic medicine where inflammation is reduced by electrical stimulation via an implant. If you want to patent at an early stage here, the concept constitutes a treatment method, which means that you might choose to patent the implant instead. But competing companies may then choose to use a different implant. So very broad concepts can be difficult to patent with commercial success, Malin explained.
She then gave an example of pacemakers. A pacemaker stimulates the heart and is therefore considered a treatment, which cannot be patented in Europe. But the actual method of identifying which signals the pacemaker should use to stimulate the individual patient’s heart in the right way can be patented – if you state that you store this data somewhere and do not use it directly to stimulate the heart. Then, of course, you can also use this stored data to stimulate the heart. An example of how you can adapt the design of patents to the legislation and get an effective patent that provides indirect protection for the treatment method anyway.
Malin also gave the example of a company that has patented a detailed description of how they trained their medical image recognition algorithm, thereby gaining intellectual property protection for their innovation. This protection has been granted worldwide, including in Europe. Another example involved wristbands used to alert an elderly person to a fall. The company used AI and machine learning to correctly interpret the signal from the bracelet. The solution to obtain patent protection here was to focus on specific parameters that were used when the algorithm analysed the signal to distinguish between an accident and the wearer just lying down. Such parameters could be, for example, which room the wearer was in, whether the wearer had moved, slept recently, run, etc. The combination of parameters provided a basis for the way the algorithm works, and this was the basis of the patent application.
– One challenge when you want to patent AI solutions is that you can’t always describe why the solution gives the answers it does, because they are based on an algorithm that has trained itself to find the answers on huge amounts of data, but in patent applications you have to be able to describe this, said Malin.
Michael summarised the need to contextualise AI and gave some key takeaways: patenting can be a powerful tool to control the value-creating phenomena that arise in digital medical technology. However, a systematic approach is required by those who want to create control through patents and other control measures. Finally, patenting in digital medtech is more complex than in most other fields and therefore requires well thought-out strategies. To achieve powerful patents in medtech, you must do the right thing, which often requires a more systemic approach. You also need to understand the full range of assets you have and how they can be used and controlled, concluded Michael.
Questions from the audience
Niclas then moderated questions from the audience, to which the participants provided answers.
Question: What is required to be able to patent an AI model for radiological images?
Answer: The way to build and train the model, e.g. based on specific parameters, if these are very different from other models, then it is easier to patent it, provided it is new. Just using a general model trained on many images is more difficult.
Question: What happens if a researcher has published an academic paper describing such a method?
Answer: It depends on how long ago it was published. In the US, you have one year to apply for a patent after publication without blocking yourself from patenting. In Europe, you must apply before publication. It should be added that academic papers do not always disclose the whole innovation, they can be more general and do not necessarily describe details that make patenting impossible, in Europe.
Q: What happens if you have patented and published something and then want to patent something that is slightly different from what was first patented?
Answer: It always gets complicated if you change things in a follow-up patent application. It is better to have everything worked out and right from the start. But the reality and needs of publishing an innovation may take priority and you should generally file an early patent application; it is better than not claiming your innovation at all.
Q: What does protection mean in practice if a large company infringes?
Answer: Unfortunately, those with the most money win. But if you have a patent, you can enlist the help of another big company to help you win against the infringer. The patent can also act as a deterrent to large companies who don’t want to get into trouble. Large companies may also appreciate that a smaller player talking to them has a patent for their innovation, so that it is clear what it is about.
Question: You can’t patent data, but can you patent the way you curate data when training an AI?
Answer: It depends on whether the approach is very different from how the same thing has been done before. If it is done in a different way, it can work, compare with the example of the bracelet that would warn of falls, where the choice of parameters was part of the curation. Then there is the actual database protection, that you own this data. And when you process the data, a new data set is created, much like the translation of a literary work from Swedish to English. In some cases, it can be difficult to prove who originally owned the data. The EU also requires that certain health data must be made available to other users, known as data sharing.
Q: Does a patent application need to have proof that the innovation works?
Answer: No, normally not. A classic example is the patent protection of perpetual motion machines.
Question: Can recommendations for treatment be patented?
Answer: Yes, the example of the pacemaker is such a case, where the signal to be used for the treatment was first stored and then used.
Question: Is it common for sufficient preparatory work to be done on IP at the pre-commercial stage? Answer: Yes, I would say so (Michael). But you should always make a business case before applying for a patent, because it can cost around one million to do all the IP work, and then you need to know that you will earn at least that amount to make it worthwhile (Malin).
KTH Innovation: we can help in these phases too, for the researchers who need it.
Nicklas: And KI Innovation can also contribute in a similar way.
Niclas: and to present a credible business case, you need to try in the IP area beforehand. This is at least my experience from this area.
Michael: And we’ve said here today that all patents can ultimately be broken by big players, but if you don’t have a patent right from the start, you have no control over what you’ve come up with and no negotiating position.
Niclas: What belongs to the field of medtech is also the long lead times, not least when the innovation is to be tested clinically and then implemented, generating the need for a lot of additions to the patent application.
Question: What are the three most important things to consider for a research group that has developed an innovation and wants to commercialise it in the future?
Niclas: To patent at all. I always suggest it to my postdocs – for every paper, the legal requirement that it is an innovation is there. There should be potential there. For every paper we publish, we submit a patent application, in principle. You also have to consider whether you can get your money back on such an investment. If you don’t see that possibility within a year, you might drop the application.
Question: We talked earlier about how you can’t protect a biomarker, but you can protect technical equipment. Nor can you patent a single molecule, but is it possible to patent a combination of molecules?
Malin: It depends on the country. In the US, you used to be able to patent treatments using a combination of biomarkers, but that is no longer possible. Nowadays this is considered a natural phenomenon that cannot be protected.
Question: how to protect AI innovation in early collaborations between researchers and companies. How do you protect the agreements you make there, when you may not even have published anything?
Michael: This is an important phase where you need to define the origin of the innovation through agreements. Even if it is too early to apply for the patent, the agreement is valid in court. Unfortunately, research teams often settle for agreements that are too general and do not go into the details of who owns what and how this will be handled at a later stage. Again, researchers need to get proper help from lawyers who know this area, because creating these agreements is complicated. I have seen so many cases where it has gone wrong.
Question: How broad can patents for diagnosis based on AI be?
Answer: It depends on whether you put in things like what parameters you used. Just using general descriptions that you used AI is not enough. Then your patent may only cover, say, 17% of the solution you developed. But then you still have protection for that part of the window. Many patents are quite narrow, but if you have the smartest solution, the market will probably not settle for an inferior solution just because it is not protected. And then you have something to sell.
In episode no 121 of Karolinska Institutet’s podcast Medicinvetarna, Magnus Boman gives his views on potential applications and ethical challenges in the use of AI in medicine and research.
On behalf of the MedTechLabs steering committee, Magnus Boman is mapping healthcare needs in terms of AI, and possible synergies between research centres and suggests training courses in the area that MedTechLabs should be responsible for.
You can find the episode here (in Swedish).