Measuring the Impact of AI/ML Technologies on the Current Medical Device Landscape

What impact has artificial intelligence (AI) and machine learning (ML) had on the medical device industry thus far? Have the emergence of these new technologies created new, unanticipated regulatory and quality challenges for industry professionals?

In this episode of the Global Medical Device Podcast, host Jon Speer and guest Mike Drues from Vascular Sciences revisit the topic of AI/ML to identify notable changes and technological advancements that have emerged as a result of these technologies and how industry professionals are responding.

LISTEN NOW:

Like this episode? Subscribe today on iTunes or Spotify.

Some highlights of this episode include:

• How is AI/ML defined in the medical device industry? Mike uses the two terms synonymously, although they are not exactly the same. He defines AI/ML as a product that is not stagnant, but changes and evolves over time.

• FDA’s definition limits AI/ML to Software as a Medical Device (SaMD) and change management. However, medical device companies tend to misuse and overuse AI to promote/advertise capabilities in software that are not intelligent.

• Regulatory and Quality Challenges: A locked algorithm dumbs-down AI technology by disabling the ability to learn from data because the software never changes once it makes it onto the market.

• AI is not new, but the challenge is validating changes not made, yet. Pattern recognition draws broad conclusions with primitive pop-ups. AI turns pattern recognition into intelligence by identifying if/when habits change.    

• Suggestions and Solutions: Come up with a range of changes that AI software is allowed to make and refuse to use regulation as an excuse.

• “Adaptive AI/ML technologies, which have the potential to adapt and optimize device performance in real-time to continuously improve healthcare for patients, do not ideally fit the traditional paradigm of medical device regulation.” - Executive Summary

• To date, only 29 products have gone through FDA processes as AI/ML. For something with so much potential for the medical device industry to improve health care, a lot more devices capable of changing are needed.

Links:

Grasping the Impact of Artificial Intelligence and Machine Learning on Medical Devices

FDA - Artificial Intelligence and Machine Learning in Software as a Medical Device

FDA - Executive Summary for the Patient Engagement Advisory Committee Meeting: Artificial Intelligence (AI) and Machine Learning (ML) in Medical Devices 

FDA Proposes Regulatory Framework for AI- and Machine Learning-Driven SaMD

When Do We Need FDA's Permission To Market Our Device And When Do We Not?

Mike Drues on LinkedIn

Vascular Sciences

Greenlight Guru YouTube Channel

MedTech True Quality Stories Podcast

Greenlight Guru

Memorable quotes from this episode:

“My definition is a product that’s not stagnant. In other words, a product that changes or is at least capable of changing over time as it learns, as it evolves.” Mike Drues

“I see a lot of companies advertising AI capabilities in a particular software. When I look at it, I don’t see anything that’s even remotely intelligent about it—artificial or otherwise.” Mike Drues

“I refuse to use regulation as an excuse to hold me back.” Mike Drues

“For something that has so much potential for our industry and for improving health care, it is a little discouraging to see or to learn that there’s only been 29 products that have been through this path.” Jon Speer

Transcript:

Announcer: Welcome to the Global Medical Device Podcast, where today's brightest minds in the medical device industry go to get their most useful and actionable insider knowledge direct from some of the world's leading medical device experts and companies.

Jon Speer: All right folks, AI machine learning and how it applies to medical device. I had the chance to talk with Mike Drues from Vascular Sciences on this topic. You may recall and if not I would encourage you to go check out episode 98 from April of 2019 where he and I first talked about that but on this episode of The Global Medical Device Podcast we're revisiting the topic. Has anything changed? Is the ball moving forward so to speak? Well, sit back relax and enjoy this episode of The Global Medical Device Podcast. Hello, and welcome to The Global Medical Device Podcast, this is your host and founder at inaudible, Jon Speer. AI, all right we've talked about this a time or two on The Global Medical Device Podcast in the past and joining me today is Mike Drues, president of Vascular Sciences. Mike, you and I first spoke about AI machine learning I think it was on episode 98 way back in April 2019 so it seemed like there's enough going on in this space that it made sense to revisit the topic.

Mike Drues: Absolutely Jon, we can talk about what, if anything, has changed in the last year and a half since we've had that discussion or are things essentially substantially equivalent today compared to where they were a year and a half ago on my point.

Jon Speer: For sure. No, that's okay, that makes sense. As you and I usually do, it probably would make sense to give people a little bit of context, put a frame around the conversation. A good place to start might be, how would you define an AI machine learning within the context of medical device industry?

Mike Drues: It's a great question Jon. As always thanks for the opportunity to have this very important discussion with you and your audience. The first thing to point out in your question, you're asking how I would define it as opposed to FDA or somebody else.

Jon Speer: That's right.

Mike Drues: So let me give you my definition which admittedly seems to be a bit higher in terms of where we set the bar then most other's including the FDA. But to me, artificial intelligence or machine learning and for the purposes of our discussion Jon, I think we can use those two synonymously even though they're not exactly the same. But to me, my definition is a product that's not stagnant. In other words a product that changes or at least it's capable of changing over time as it learns, as it evolves. Almost in a winning evolution sense of the world. Most of the products that we have today in medical devices for example, we develop the product, we get it to the point of design inaudible and then we do all of our final testing on it, our final V&V testing and we release it out to the market and that product never changes unless and until the manufacture decides to change it. But with AI, what I call true AI, artificial intelligence products and notice that I'm not limiting it to just what FDA refers to as SaMD, software as a medical device because some of the AI technologies I'm involved with right now are incorporated too are incorporated into physical devices as well. Thinking of AI limiting as FDA is doing right now to SaMD, I think is already myopic or limited thinking. But basically, as I said, a traditional device does not change or evolve once it gets to the point of design pres, whereas a product that has true artificial intelligence learns, evolves, with use, with more information and so on. Does that make sense Jon?

Jon Speer: Absolutely does. There's a couple threads we can pull on and maybe we'll dive into those but I guess sort of a rhetorical question, and we'll get into this in the ensuing conversation here, but it seems like if we ably the Mike Drues definition of AI that we might be overusing or misusing that term with respect to medical devices. What do you think?

Mike Drues: Well, regrettably Jon, I could not agree with you more. In fact, I think you're putting it quite kindly. I see a lot of companies advertising AI capabilities in a particular software and when I look at it I don't see anything that's even remotely intelligent about it, artificial or otherwise. So I do think it's an overly used phrase. In fact, if you look at the medical device products, the software that have come through the FDA thus far, there are about 29 devices that have been cleared or approved or granted that include AI but this does not fit my true definition of AI. All of them use this, what I think of as an archaic concept that FDA introduced in its original discussion paper back in April 2019 when we had that discussion. This concept of a locked algorithm, where basically you use the artificial intelligence during the development process. For example, you feed data into your software from a canned data set that you purchase or you get from somewhere. You alow the AI to learn if you will by chewing on that data which is fine but then you lock the algorithm. In other words, you disable its ability to continue to learn. You get it to the point of design pres and then you release it out to the market and the software never changes again. And to me that's really dumbing down our technology. What I would really like to see Jon, and we can get into this in more detail if you want, is not just allowing the software to learn during the actually development process but let it continue to learn after it's already on the market and as you can imagine Jon, that has some pretty interesting regulatory as well as quality challenges along with it as well.

Jon Speer: Yeah, for sure. That totally makes sense and that's kind of my next thought. There are some big challenges with respect to AI and medical devices. You started to hint a little bit at some of the regulatory framework but what are some of the other challenges that you see facing AI and med device?

Mike Drues: Well, I'll come to that in just a moment Jon. I'd like to mention one or two other things that will lead into that. It's fascinating to me how so many people, including my friends at the FDA, seems to think that AI is so new, so revolutionary compared to what we've been doing in the past and yeah there are some differences, some new challenges but when you really try to understand the medical device product development process and change management and so on, there's really not a lot new here. Let me use a simple medical device example. I know in your past Jon you used to be involved with catheters a lot. When you're developing a new catheter, it's very common for you to make a prototype and then to test it, maybe using it on a torta trek model to see how that catheter's going to track to a particular part of the body. In other words, you're exposing that catheter to a data set so to speak. And then you might make adjustments to the design, maybe you might change the length or the diameter. Maybe you might change the material properties, the mechanical properties. In other words, you tweak it, you iterate it, you test it again and then at a certain point you get to the point where you say, okay, this catheter is pretty good, we're going to declare design pres. We're going to do our final V&V testing. We're going to submit all of this to the FDA, we're going to get it all onto the market and then we're done and we're not going to change that catheter again unless and until the company decides to do so and then they repeat the whole process. How is that any different than what we do with AIs? Especially with the concept of a locked algorithm where your catheter, once you get it to the point of design pres, that is essentially a locked algorithm and that catheter is certainly not going to continue to change once it's on the market. But I think, as I said a moment ago Jon, as a biomedical engineer, never mind as a inaudible career professional but as an engineer I think that is really dumbing down this technology because it's preventing the technology from doing what we really want it to do. Does that make sense Jon?

Jon Speer: It totally makes sense. Here's this awesome opportunity or potential benefit of an AI product but it's being handcuffed from actually doing what it was designed or intended to do.

Mike Drues: Exactly. To use the catheter metaphor a little bit further. If a catheter was intelligent, if a catheter would be able to change either mechanical properties or geometric properties while it was in a particular patient, in other words after it's gone through the FDA. And by the way Jon, even in the catheter world we have examples of this, we have materials. For example, PVO electric materials that we can use that do change mechanical properties of a device or electrical properties of a device on the fly. So this concept is really not that foreign but the challenge of course is, how do you validate changes that have not happened yet? I have some solutions to that I'll get to in a moment Jon but one other metaphor that I want to use because I think there's a lot of confusion here in artificial intelligence versus just simple pattern recognition. Let's use not even a medical device metaphor, let's use a simple metaphor that everybody can relate to and that is the TV guide on your television. If the TV guide in your television had artificial intelligence, it might notice for example, and I'm just going to use and example that might appeal to some geeky engineers out there, a TV show like Big Bang. And if you watch Big Bang a few times, you might get popups reminding you that there's an episode of Big Bang, you're not watching it now, would you like me to show it to you? But that's very, very primitive. That would not be what I consider to be artificial intelligence, that's just pattern recognition. To me artificial intelligence would be not simply noticing that you're watching Big Bang but drawing a more broad conclusion like, hey, this person seems to like to watch comedies in the evening as opposed to maybe news shows in the afternoon and the morning and taking it a step further. The software, if it's truly intelligent, it will say, this person likes to watch comedies in the evening when the weather is nice and so I'm going to offer not just Big Bang but... I don't know, I'm dating myself here Jon, like Freeze Comedy or some other kind of a comedy. On the other hand, if it's in the evening and the weather is bad then this person might prefer to watch a movie instead of a comedy like Big Bang. To me Jon, that's truly artificial intelligence and taking it one step further in terms of the evolution here. That software should also be capable of determining if your habits change over time. For example, maybe when you first start using this TV guide it first recognizes that you like Big Bang, then it recognizes that you like to watch comedies in the evening. Then it recognizes that you want to watch comedies in the evening when the weather is nice as opposed to movies in the evening when the weather is bad. And finally, maybe six months from now you might have changed your viewing habits and you no longer watch comedies in the evening, maybe you watch dramas or maybe you watch political shows or something like that. To me Jon, most if not all of what we're doing on the technology side when it comes to AI or ML, let's just put it this way. It's pretty hard for me as a professional biomedical engineer to become really excited about what we're doing right now. But that's okay, I understand that this is a starting point obviously. There's an evolution just like AI is a revolution. The question is, now we can come to those regulatory and quality challenges, how do we develop medical devices that allow us to do some of these things and at the same time still address some of the regulatory and quality challenges that you alluded to a few minutes ago.

Jon Speer: Actually, when you said TV guide I immediately thought of the old TV guide and then you make a reference to inaudible company so I get those references but I know what you were talking about on screen. Taking that analogy or metaphor, whichever is the right descriptor there. If I like comedies and the AI is suggesting that, hey maybe you would also like Freeze Company if you're a Big Bang fan but then it notices my patterns over time and notices that I'm shifting to more dramas or whatever the case may be. I could see where that would be challenging from a regulatory perspective, bringing it back to a medical device because if I shift from comedy to drama does that mean that my indications for use has changed or my application of use has changed? What do you think?

Mike Drues: That's a very good question Jon. Let's look on the technology side and then if you want we can come back to the labeling. Let me go back to my catheter example because obviously that's a little closer comparison than the TV guide. We don't have this capability today but I'm just using this to make the metaphor work. Imagine that we have a catheter that has some" artificial intelligence" in it that will allow the catheter to change in length or allow the catheter to change in mechanical properties maybe like the PVO electric example I mentioned a moment ago where it becomes stiffer, where it becomes more flexible on the fly based on what is happening, the information that the catheter is gaining when it is in the patient. Obviously that's going to make a lot of people, including FDA very, very nervous because as I said, how do you validate a change that has not happened yet? If the company decides to design the catheter to be stiffer than they're going to be doing all the appropriate V&V testing to support that that stiffer catheter is still safe and effective. So the question is how do we allow the catheter to do that and still have the regulatory and the quality controls in place so that we don't have crazy, weird catheters evolving? Here's one solution Jon because one of the common themes, one of the things that differentiates my approach from so many others is I refuse to use regulation as an excuse to hold me back. Even though this concept of a locked algorithm in my opinion, for whatever it's worth is truly an archaic concept, it's really holding us back, we've got to come up with ways to figure out how to get our more advanced technologies through this regulatory environment that is clearly not intended to do this. How do we do it? Well, one of the suggestions, and I have to be a little careful what I say here Jon because we're actually taking a few AI devices to the FDA right now using this approach but one of the suggestions that we're trying is we're coming up with a range of changes that the AI software is allowed to make. In other words, again using the catheters as an example, you can change the length between X and Y. The software can change the length between X and Y or the stiffness between A and B and we lock that much into the catheter in advance. If you want to make an adjustment to the length of the catheter fine, based on what you've learned but that adjustment has to be between X and Y. And then what we do as part of our submission is we do a classic validation, Jon, that you're very, very familiar with, probably more than most. You validate the minimum, you validate the maximum, you validate a few points in between and then we can be confident that as long as the software evolves and makes changes within that pre- determined, pre-validate range, we can be pretty confident that those changes are going to be appropriate. Those changes are going to be safe and effective. Is that truly the solution? In the long term, probably not but at least it takes us a step past this notion of a locked algorithm where we don't allow any changes to occur whatsoever once we reach design pre and the product is on the market. Does that make sense Jon?

Jon Speer: Yeah, I'm tracking with you. I want to get back to the conversation here in a moment. Folks, I want to remind you, I'm talking with Mike Drues. Mike is president of Vascular Sciences. He is a regulatory expert and specifically with respect to the topic of today, of AI and med device. This is one of his sweet spots and domains of expertise so if you're exploring best practices and regulatory strategy with respect to AI and your medical devices then I would encourage you to reach out to Mike Drues. And while we're at it let me remind you that Greenlight Guru is here to help as well. Go to www.greenlight.guru to learn more about the medical device quality management system software platform designed by medical device professionals for medical device professionals from Greenlight. Go check that out. We have quite a few customers that are in this space that are developing AI different related products in that nature and thankfully we have partners like Mike Drues who are there to assist if and when those companies that we're working with need that assistance. So again, go check out www.greenlight.guru to learn more about the medical device quality management system software. All right, Mike I was just thinking about the regulatory, maybe peeling it back a little bit further but the framework that's in place. FDA's had the pre certification program around for a bit, they've had the digital health initiative and even as recent as back in September they released the formalization of the Digital Health Center of Excellence. I don't mean to be overly critical here but it seems like there's not been a lot of substantial progress or movement. I get that it's really complicated but man it seems like there are other things that are a little bit of a head scratcher. To your point, I'm curious to find out on your example, how things go. I know you're an artist when it comes to these sorts of things. But being able to design in acceptable parameters for allowing dynamic changes to the AI I think is a really, really smart move. Is there any sort of precedent that you're aware of from a regulatory perspective? And I guess this is maybe a 10 part question rolled into one but how to you think things are going from a regulatory perspective?

Mike Drues: Well, good question Jon and by the way, thank you for your kind words. I greatly appreciate that. Let's start out with the last question, is there precedent? I guess it depends on how broadly you're looking for precedent. If you're looking specifically at AI as FDA defines it right now products, no. If you look even a little bit more broadly at software as medical device SaMD, no there's no precedent. However, when you look more broadly than that, at the medical device universe in general I think there's a tremendous amount of precedent. And again, I have to be a little careful because we've already to presented some of these ideas as pre- subs at the FDA for some of the products that I'm involved with but you know as well as anybody Jon that when you come out with a catheter or a stint or a heart valve or an in vitro diagnostic or whatever it is, typically what happens is you come out with this device in different sizes, different increments. Catheters may be different lengths or diameters or something and then you do that validation of those different sizes or at the very least the smallest, the biggest and a couple of points in between. The short answer to your question Jon is no there's no precedent within the AI or software per se but there's a tremendous amount of precedent across the broader industry and this is why it's so frustrating to me and like said earlier, people think this is so new but when you understand not just the regulatory logic, obviously Jon you've heard me talk about regulatory logic all the time, but engineering logic, there's so much precedent. When you understand the regulatory logic and the engineering logic, to me Jon, there's just not a lot that's really new or different here if you truly understand what we've been doing in the past.

Jon Speer: No, I totally get that. I remember back in the day when I was working on catheters and this is where some of the art would come into the regulatory strategy and submission process. We knew where we were going as a company and we knew where we wanted that particular product line to go and we knew that we wanted to try to minimize the burden so to speak, on our behalf from a regulatory submission standpoint so that we weren't constantly going through the cycle. We tried to look in the scope of what we thought was going to be within the realistic possibilities in the foreseeable future when we did that. There was good engineering, good science behind it too but yeah I agree. When you put it that way it seems like there is nothing new here. It just so happens that instead of it being a catheter or other tangible types of products, we're talking about software. But the same methodology, the same logic can apply.

Mike Drues: Exactly correct Jon and ironically I just did a webinar for Greenlight on what is a regulated medical device and one of the things I tried to stress to the audience is the form factor if you will. The physical form of the product whether it's a physical device like a catheter, whether it's a in vitro diagnostic, whether it's an electrical device, whether it's a piece of software, whether it's a liquid. Quite frankly I could care less, the form doesn't matter, it's does it fit that quota inaudible regulations definition of a medical device. And one of the things that we started out this discussion with Jon is what has changed in the last 18 months or so since FDA put out its original 20 page discussion paper on artificial intelligence and machine learning and they limited it pretty much to softwares medical devices and changed managements and so on. But just recently, just a few weeks ago, they put out sort of an update to that and we can put a link on the website, an executive summary from the patient engagement advisory meeting on artificial intelligence and machine learning in medical devices. I don't have time obviously to go through the details of this but to be honest with you Jon, you mentioned something a moment ago about being critical, obviously when a company makes a submission to the FDA, the FDA's job is to be critical. As I like to say, the FDA's job when they're doing their job is to criticize everything so if you say, " This guy is blue," the FDA's job is to prove it. Well, with all do respect I'd like to be a little bit critical with our industry as well as the FDA here when it comes to AI because I think we can do a much better job at this. When FDA put out this most recent update in October of 2020 just a few weeks ago I read through this and to be critical, there's a lot of marketing height in here that I see, probably written by the politicians. AI is this greatest thing since sliced bread, it's going to solve all of our problems but when I put my engineer head on there's really little or maybe nothing or what I would call engineering substance. How do we do the things that you and I are talking about here? How do we get past this locked algorithm in allowed devices to truly utilize artificial intelligence and make changes not just pre- market but post- market as well. I would just like to read one small section from this most recent document out of FDA on AI just last month. It's maybe two sentences. This is a direct quote." Adaptive AI and ML technologies which have the potential to adapt and optimize device performance in real time to continuously improve healthcare for patients, do not ideally fit the traditional paradigm for medical device regulation." Let me repeat the last part. " Do not ideally fit the traditional paradigm for medical device regulation." I hope at least some in our audience that feel like I do, that we refuse to use regulation as an excuse to hold us back because let's be honest Jon, I have a lot of companies come to me with really cool- I used to live in Boston for 25 years, what I would call wicked cool technologies using AI but they're hesitant to include it in their product because of this concept of the locked algorithm and everything else. We need to make some significant changes here and with all do respect to me many friends both in industry and the trade association as well as in FDA, we don't need more people just waving the flag and saying, " Oh yeah, AI is going to be the greatest thing since sliced bread." That doesn't serve anybody when you're honest. We need to get into the engineering detail, the regulatory detail, the quality detail. Like for example, I mentioned a moment ago the concept of the pre validated range of changes that a device can make. I have other suggestions I've made to companies, other potential strategies as well but that's the only one that I think I'm wanting to share at the moment publicly. I don't know Jon, is it just me or do other people share similar concerns as well?

Jon Speer: I'm glad you went down that path because in preparation for today's conversation I did a little bit of research on some things. You mentioned this earlier and I thought it's worth discussing maybe a little bit further but to date I think this is through maybe September, October but there have only been 29 products that have gone through FDA processes as AI machine learning. This is going back to 2016 through 2020. For something that has so much potential for our industry and for improving health care, it is a little discouraging to see or to learn that there's only been 29 products that have even been through this path and I think you're right, I think people are hesitant to pursue that because of the real regulatory obstacles and hurdles to pursue these types of opportunities. I do think we need to do more on all sides of the table to continue to advance this because you're right, just saying, " AI is amazing." Well, no kidding but that doesn't add any value so how can we level up? How can we get to a point where both sides are collaborative in a way, such that we can actually have many, many more products above and beyond 29 devices. In context, 29 devices from 2016 to 2020 as AI, that's probably a fraction of a fraction of a percent of the total devices that have been cleared or approved by FDA.

Mike Drues: I would agree Jon and taking it a step further, on one hand the glass is half full, at least we have 29 devices but on the other hand when you look at those devices as we talked about earlier, to me all of the artificial intelligence has been sucked right out of them because they all have had to implement this concept of a locked algorithm. So again, I can't stress this enough, this is an important point for our audience, the AI that these 29 devices have used have been in the development process. And once a product is developed, just like a catheter, you lock it in i. e. design freeze and then you never let it evolve anymore and that's a problem. I also find it interesting Jon, I don't want to attribute quotes to my specific friends at FDA but there have been some pretty high level folks at FDA that have been quoted as saying, " We have a lot of questions when it comes to AI, we're not sure how to proceed when it comes to AI." This is where I think we as an industry and we as individual companies, have to be much more proactive and we have to not be content with this limitation of a locked algorithm and instead take a product that has true AI capability, has the capability to for example, truly learn in a box, frolic on the market. And taking it one step further Jon, why do these devices have to evolve on their own? Again, using the catheter as an example, why not have all of your catheters being able to talk to one another so to speak? So that they can learn not just from their own experience but from other similar catheters being used in other households by other positions, sort of a herd of animals if you will. And by the way Jon, here's another pragmatic solution, again I have to be a little careful what I say here but this is also something we've proposed to the FDA in some inaudible meetings for some specific devices. Instead of allowing the AI software to make the changes post market by itself, we allow the software to make the recommendation to the change back to the manufacturer. In other words, the software can't implement the change, it can't learn or implement the change on its own. It recognizes the change and it sends some sort of a message back to the manufacture, again using the catheter as a metaphor." Hey, we probably would get better results if we made the catheter a little bit stiffer or a little bit longer." Allow then the company to evaluate that change, to go through their normal change management process which you and I have talked about many times before Jon. Again, nothing new here to me whatsoever. The company can do the V&V testing on the changes that the software is suggesting and then the company can decide, okay do we need to notify the FDA via a special 510(k) or a PMA supplement or something like that or do we not notify that FDA and handle this via a simply letter to file. And then the company maybe issues a software update for example that implements those new changes based on what the AI software recommended but the software can't do it by itself because there isn't another way that we can control these changes and still have at least some of the capabilities that AI allows us. I hate to keep harping on this point but confining us, using this locked algorithm approach because to me Jon, your point is very well taken. We shouldn't have 29 devices, we should have a heck of a lot more than that an more important to me than the number because to me these 29 devices from an engineering perspective, they don't really excite me very much because we've dumbed the technology down. We need a lot more devices and we need a lot more that are capable of changing as we've talked about.

Jon Speer: Yeah, and just to kind of follow up on that last little bit there. To your earlier point, there's already regulatory precedent for handling design changes, why should this be any different? If the company gets a recommendation based on the feedback that their products are gathering, the data that they're gathering and the company goes through the design change process to determine impact on the product or regulatory clearance or approval or whatever the case may be and they determine it's letter to file, and they can push out a software update. Well why not? There's precedent here.

Mike Drues: I'm sorry Jon, if I could just pick up on that just real quick when it comes to precedent because again we have to nail it on the head. There's so much precedent across the medical device industry. How many times does it happen where a physician or a surgeon is using a device, a catheter, a laparoscope, whatever it is and he or she provides feedback to the company. Hey, if you make this a little longer, a little shorter, a little fatter, a little thinner, I can use it better, I can use it for something else.

Jon Speer: For sure.

Mike Drues: It's just now, the software is doing that as opposed to the physician or sometimes the patient and you also mentioned Jon, there's a tremendous amount of precedent when it comes to change management which I would agree but I would also remind you and our audience that one of the most common reasons, perhaps it might even be the most common reason why FDA issues inaudible observations and warning letters is because of change management or the lack thereof.

Jon Speer: Touche.

Mike Drues: We should learn from those problems and apply it to our AI products as well.

Jon Speer: The software feeding information back to company. This is a pure form of real time customer feedback. It's actually a better mechanism than what we are currently quote stuck with as an industry because we have to rely on feedback and even if we're being proactive to go it a lot of times it's a dead end. This is a mechanism potentially that can allow companies to be more proactive and you and I have talked about cappa in the past and reacting and all that sort of thing and how that's now the ideal state.

Mike Drues: Yeah, we should wrap this up but just to be clear, I think there's solutions to all of these problems that we're talking about and we're just barely scratching the surface but I think there's solutions to all of these. But the solution in my opinion is not to just simply have more people reading and following the regulation that we already have because that's not going to solve any of our problems, if anything it's going to propagate the problems that we're already having. The solution is to get people to think. And the solution is to say, okay, this is the intent of the regulation and how do we apply the regulatory logic or the engineering logic to meet the intent and at the same time still be able to take advantage of all of the bells and whistles that our new and wonderful, as all of the politicians say, " greatest thing since sliced bread," AI technology can offer because right now Jon, I hate to say it but this is the lowest of the low hanging fruit. That's okay to start but we need to make progress from here.

Jon Speer: Yeah, to kind of wrap things up today folks, I think we should head Mike's advice. I think we as an industry can take the lead on this. I do not envy FDA at times. It's an awesome challenge that they're in. Mike, what is your quote about FDA as far as clearing a device? I know you and I have chatted about this before. If an FDA reviewer clears or approves a device that can have tremendous impact if it goes wrong.

Mike Drues: If it goes wrong, and it can also have tremendous impact if FDA doesn't clear or approve the device because that means that patients are not going to have the potential benefits of it so that street runs in two directions Jon.

Jon Speer: Yeah.

Mike Drues: To wrap this up, I think they are solvable problems and I've offered a couple of tangible solutions. Just to reiterate very quickly, one is my concept of the pre- validated range. Where you allow the software to make changes between X and Y and you've pre- validated X and Y and a couple of points between. I think that's a very reasonable, very logical strategy. The other strategy that I mentioned is allow the AI to come up with recommendations but not allow the AI unilaterally to implement those recommendations but instead send that information back to the manufacture. The manufacture evaluates whether or not they should implement those changes and if so how. And then I think a third idea that I mentioned is allowing the devices, because whether there is a software or hardware or something else it really doesn't matter, not just to learn by their own experience but to take advantage of the experience of the group of devices, the herd it you will. Those are three very, very tangible recommendations I think and I've got others if people want to talk to me. Like I said I've got several devices that in my opinion are utilizing true AI that we are trying to get through the FDA right now. Have we done it yet? Not a hundred percent but we're making progress because I just refuse to use regulation or the FDA as an excuse to hold me back. Those are just a few of my final thoughts Jon, anything that you would add?

Jon Speer: The last thing that I add and then we'll put a wrap around this one. There's a couple quotes that I jotted down from one of the articles that I reviewed on this topic. Let me share those with you." While agency has experience in softwares and med device, pre- specifications and algorithm change control, FDA is looking to industry and stake holders for input on good, machine learning practices."" FDA has so many question about what good practices look like for algorithm design, development, training and testing." As Mike and I have chatted about today, this is not a unique thing for software and for AI. You as the developer of your product, should have I hope the most expertise around your products and technology. You should be the one that's leading the conversation as to why this technology is applicable, appropriate and provide that guidance and direction to the FDA as to what those best practices are. I think we as an industry can band together to help lead the charge on this because there's so much opportunity certainly in the AI space to improve the quality of life. As always, I want to thank my guest.

Mike Drues: I could not agree with you more. Just very, very quickly Jon. What you're reiterating there is another of my common or my core philosophies and that is, we always have to look for similarities where no similarities seem to exist. So on the surface it may sound like software as medical device and especially AI has a lot of questions and unique challenges and no doubt there are some differences but I would argue as you and I have talked about today in this discussion Jon, there are a heck of a lot more similarities than there are differences.

Jon Speer: For sure.

Mike Drues: And if we take a broader view of the universe and say, how does AI software, how do the challenges either pre or post market compared to catheter development, IVD development, hip implant development? There's a heck of a lot of similarities so we just need to look for those.

Jon Speer: For sure.

Mike Drues: That's another take away from the discussion.

Jon Speer: Absolutely. Mike, thank you so much for this intriguing conversation and folks remember Mike Drues, Vascular Sciences if you're interested in exploring AI or frankly any other medical device technology and need some support and assistance with your strategy and all the different options and pathways ahead of you. Mike is a guy that you want in your corner. As always, thank you so much for being a loyalist of the Global Medical Device Podcast. It's because of you that this is still the number one podcast in the medical device industry so keep sharing this with your friends and colleagues and till next time I hope you all are safe and doing well and we'll talk to you real soon on the Global Medical Device Podcast.

ABOUT THE GLOBAL MEDICAL DEVICE PODCAST:

The Global Medical Device Podcast powered by Greenlight Guru is where today's brightest minds in the medical device industry go to get their most useful and actionable insider knowledge, direct from some of the world's leading medical device experts and companies.

Like this episode? Subscribe today on iTunes or Spotify.