Dr. David Friedenberg is a Principal Data Science and Neurotechnology and the Team Lead for Machine Learning/AI in the Health Analytics group at Battelle. He's the Principal Investigator on several neurotechnology efforts developing new AI-powered technologies to help improve the lives of people living with motor impairments due to neurological injuries like spinal cord injuries and stroke. An experienced data scientist with consulting experience across several disciplines he is passionate about developing AI/ML-driven solutions to challenging problems for the betterment of humanity.
BCI Pioneers Coalition: BCI Pioneers Coalition
Episode Description
About 5.4 million Americans live with some form of paralysis. Sometimes that's just a temporary loss of mobility, but for the Americans whose paralysis is caused by a spinal cord injury, that loss of movement is often permanent, as there's no biological way to heal an injured spinal cord. There are efforts to see if technology might be able to help these individuals regain use of their limbs, and that's the focus of this episode of Stats+Stories with guest Dr. David Friedenberg.
+Full Transcript
Rosemary Pennington
About 5.4 million Americans live with some form of paralysis. Sometimes that's just a temporary loss of mobility for the Americans whose paralysis is caused by a spinal cord injury, that loss of movement is often permanent, as there's no biological way to heal an injured spinal cord, there are efforts to see if technology might be able to help these individuals regain use of their limbs, and that's the focus of this episode of stats and stories, where we explore the statistics behind the stories and the stories behind the statistics. I'm Rosemary Pennington, stats and stories is a production of Miami University's departments of statistics and media, journalism and film, as well as the American Statistical Association. Joining me, as always in the studio, is regular panelist John Baylor, emeritus professor of statistics at Miami University. Our guest today is David Friedenberg, a Principal Research statistician at Battelle. Friedenberg has been the algorithm and data analysis lead for the neuro life, neural bypass collaboration between Battelle and the Ohio State University for the past three years. He earned his PhD in statistics from Carnegie Mellon in 2010 and a BS in mathematics and statistics, with a minor in computer science from Miami University in 2004 Friedenberg is also the author of a recent chance article explaining efforts to help people living with paralysis regain mobility. Dave, thank you so much for joining us today.
David Friedenberg
My pleasure. Thanks for having me. It's great
Rosemary Pennington
to have a Red Hawk back.
John Bailer
Yeah, you know, Dave. I'm just delighted to have you on talking about this really very cool work I you know, and rosemary mentioned, sort of when she said recent chance, it was like 2016 that you did this work. That's, that was, what was fascinating to me, is to realize that that some of this work on brain computer interfaces has been going on for more than, well more than a decade. And so I guess let's, let's start back at kind of the a little bit of the origin story of some of this work that goes with with this brain computer interface. Can you talk a little bit about, give us a sort of a the flyover of kind of the general the general problem and general approach, absolutely
David Friedenberg
so when we talk about brain computer interfaces, that technology has been hypothesized and developed for quite a while, for several decades. When my team started getting into it, as you said, it was in the early 2010s and we started to think about, how can we record signals from the brain? So that typically involves surgically implanting electrodes on the surface of the brain and recording electrical activity of the brain when someone's thinking about moving. And then, how do you take those signals out of the brain, run them through what my expertise is, is the algorithms to understand what they're attempting to do, and then use that intended movement to try to affect some end effector device that helps them do something that they want to do. So we were thinking a lot about spinal cord injury. So when someone has a spinal cord injury, the signals that are going from their brain down to their body are disrupted by the injury, and that often results in a loss of movement or paralysis, and specifically when you lose the use of your hands, it's a devastating loss for most people, because you need your hands to do so many different things, and there's lots of research about the priorities of this population, and hand loss is one of the main ones that's a target for restoration. And so what we thought was if we could understand when the user is intending to move their hands through the brain computer interface. We knew there was some existing technology called functional electrical stimulation, where, if you stimulate the arm, you can actually generate movements of the hand. And so we thought, if you can combine those two pieces together with the algorithms in the middle, then we could have someone who can't move their hand, but can think about moving their hand. We can understand that movement, and then we can stimulate their hand to actually generate the movement that they're intending. So system works as intended. They think about moving in a split second later, their hand moves the way that they intend. So that was the idea behind it. There was a lot of technology development to get there so that functional electrical stimulation system didn't really exist in the way that we needed it to. So that was something that my company, Patel had to build, and so we built this array of electrodes that spits on the arm and allows us to get all sorts of different functional hand movements. And then the study with Ohio State, we had a participant who was fortunate enough to volunteer, and his name's Ian Barker. He's disclosed his name's secret anymore, but he volunteered to have brain surgery to get one of these brain computer interfaces implanted. It's called the Utah array, made by a company out of Utah Blackrock neurota, and so he had that system in place for about seven years when he was using the system. It's an investigational device. It's not something that's out there for the public. But when he was using the device, he was able to regain some hand function, do things like opening his hand, closing his hand, picking up objects, manipulating different kinds of objects. We would play battleship. We would be doing, you know, simulated cooking tasks, simulated feeding tasks, those kinds of things. And then we did some other stuff that was unique to the fact that we had access to this brain data, which is very hard to come by. So we also had him do things like driving a simulator. So he would be navigating on a simulator on the computer, and he'd be navigating on the road. And so we did lots of those kinds of experiments, and one of the things that we found that we weren't expecting was that he actually started to regain some function when he wasn't using the system. Wow. And so as you mentioned, that's not something that you typically expect to see with someone with a spinal cord injury, especially someone who's so far removed from their injury. And so we started to think then about, well, how might we be able to design a device that's intending to get those kinds of effects but eliminates the need for the brain surgery and the implant and everything, because that's a very serious and significant expense, and it also adds a lot of risk. And so past several years, our team's been thinking a lot about, how can we sense the user's intent to move whether they have a stroke or a spinal cord injury or some other kind of neurological injury? How can we get that whole system working with the intent to move, and then the movement, but taking that invasive surgical piece out? And so we've come up with a system now, also an investigational device, but that system is actually recording electrical activity from the surface of the arm we call electromyography, or EMG, and similar to the brain computer interface, the users attempting to move were recording in real time, interpreting in real time, and then stimulating to help them accomplish movements that they have trouble accomplishing on their own. When John
Rosemary Pennington
and I were talking about having you on as a guest, I was telling him, when I was still a working journalist, I did a reporting workshop with an organization out in Maryland on the human brain. And we went out to NIH, and this would have been in the very early 2000s like maybe 2000 to 2003 and we were visiting with these researchers at NIH, who had also been working on something related to this, but who had said at the time, it was so far distant from being able to actually help people regain movement that they were really just now trying to figure out how that might be possible. I wonder if you could sort of explain, you know, for our audience, who knows nothing about this, what has changed over the last, you know, 20 years to actually allow the like things for like the development of the technologies you're using and the sort of data analysis you're using to be able to do this?
David Friedenberg
Yeah, it's a great question. And I think there's a couple different pieces that have have spurred this. You know, first is, as with a lot of fields, there's been a couple of pioneers, and Ian was one of the pioneers in terms of being one of the first people to go through this process and and, you know, be a guinea pig for us, in the sense that, you know, he was trying out some technology that had never been tried out before, trying to do some things that no one in his condition had ever done before. And so when we have success there, it encourages everyone in the field to keep pushing the boundaries and keep trying new things. So I think there's now upwards of about 50 people who have been implanted with these types of devices. And you know, as we collect more data, you know, the safety profile becomes clear and and fortunately, there haven't been a lot of significant safety events with these kinds of devices. And I think that's also encouraging to people. And then I'm sure many of your listeners have heard, you know, Elon Musk is investing in this space. There's lots of money flowing into this space. And so there's now several companies that are actively developing these types of devices. Each one has a slightly different profile in terms of how invasive the surgery is and how many electrodes they have and some of the technical parameters around it. But I think with all that investment, it's really pushed people to push the boundaries of what's possible and try out some new things far as the algorithms go. That's something that's also, you know, really transformed from when we started to where we are today. So just like in lots of different areas, you know, things like deep learning have have come into play and have really transformed some of the things that people are able to do with these devices.
John Bailer
This is really amazing. And I was thinking about the the just the amount of information that you said was being processed in real time. You know, I think, you know, you say this so lightly, but it's, it's mind numbing to think about how much information you're really needing to process. And even, even 10 years ago, you're describing a system that was generating, you know, 2.9 million samples that were being processed every second. So, you know, this is an amazing amount of information. And as you noted in there, you can't have a system that's going to lag the signal from the brain, you know, so you're you have to be able to quickly decode what is being intended by the person and then encode the signal for the muscle. Muscle movement that you wish so, so what? What are some of the the challenges and thinking about that much, that much information coming in, and basically boiling it down to what you need to act so,
David Friedenberg
as you mentioned, you have to be able to process this stuff more or less in real time. I say real time. We process in a loop about 100 milliseconds at a time, and and the reason for that is, if the user is thinking about moving their hand, and they don't get that response almost immediately, then they start wondering, is the system working? Am I? You know, they start changing and adapting. And that's one of the biggest challenges that we have with these kinds of systems, is, you know, we think about machine learning and statistics typically as I have a data set. That data set is usually kind of static, and I have a train and a test and I do my analysis, but here, like we're collecting the data in real time from the user, but the users also adapting to us. So every time we train an algorithm and they're using the algorithm, they're trying to adapt to figure out, how do I think about doing this so that I can do it better? And so there's this kind of CO adaption process that's a very challenging problem for us to figure out, because sometimes we don't change our algorithms, or don't adapt our algorithms as much as we might normally, because we want the user to adapt, or we don't want to change things up on using too too much. So obviously
Rosemary Pennington
this is a large team has been working on this. What is it like to collaborate on something like this, and sort of, what has that? What's that entailed, and maybe, what have you learned from one another?
David Friedenberg
Yeah, this is a great example of team science. So the team for this project is very large. It involves neurosurgeons. It involves physical Med and rehab docs. It involves therapists. It involves all disciplines of engineering. We had electrical engineers, systems engineers, computer scientists, data scientists, I'm sure I'm missing a few in there, but all of us having to work together. You know, personally, as a data scientist and a statistician coming into this project, I had no real knowledge about spinal cord injury, and so, you know, I think you also be remiss to not acknowledge that the participants are a very, very big part of this process, and, you know, we're designing this technology for them, and so their voice and their needs, you know, they're an integral part of the team, so you'll hear, you know, I usually don't refer to them as a subject or a patient. They're participants in our studies, and they're really collaborators with us. You're
Rosemary Pennington
listening to stats and stories, and we're talking with Mattel's David Friedenberg. David Friedenberg, you know.
John Bailer
And it's, it's amazing to think about the, you know, just the investment that the participants have in terms of doing this work. I mean, the how, the kind of the brave, the braveness and the boldness to basically say, okay, yeah, do this. Do brain surgery to help. Let's, let's see if we can, can move this forward. I find it. I I'm just trying to think about just the unbelievable amounts of approval processes that might go into doing this. I mean, you know, you're, you know, when we think about clinical trials with with just things that are, are well known meds, or kind of meds that we're going to be operating in ways that are our pharmaceutical agents, that we kind of know the way things are going to function, but But now you're talking about, ultimately, the the design of a medical device, and that's that I just wondered how challenging that was to kind of make the case for doing this type of work, when in human subjects or and I mean human subjects in the sense that there's human subject approval processes that are a part of any type of work like this,
David Friedenberg
yeah. So to do this, there's a lot of interaction with the FDA. We have, you know, investigational device exemption to do this project. There's a lot of oversight. And so it is a very complex process. No doubt. I think as more and more groups are having success in this space, I think it becomes a little bit easier, because there's a lot more precedent that you can say, hey, you know, there's been 20, 3040, people who have been implanted with this type of system, and this is the safety profile, but, but absolutely, and I think there's a lot of ethical issues that we have to take very, very seriously when we do this kind of work as well. You know, when someone has a spinal cord injury, it's a traumatic loss of of ability and function, and when we're providing these technologies that help to hopefully restore some of that functionality, and then at some point, you know, that trial ends and that that's taken away, you know, that's an emotional toll on that person as well right there, hopefully using the system and doing some things that they haven't been able to do for a while, and then all of a sudden, when the trial ends, you take it away. So that's something that we think about about too, is, how do we transition people off? And, you know, are there ways that we can offer them, you know, maybe additional study where they get to use parts of the system, or have them involved in other ways so that to kind of ease that transition back. So yes, and FDA has been an awesome partner on this. They're, you know, they're very interested in moving this kind of tech forward, but obviously have a very important responsibility to make sure everything is done safely and by the book. So you mentioned
Rosemary Pennington
earlier some of the early, early. Tech with sort of, you know, brain implantations that involved sort of a, you know, linkage between that and technology might have, that might have helped an individual move. And now there's this move towards, I think, what you've mentioned that your company is working on this sleeve. And I wonder, you know, how, how quickly does that development take place when you move from something that you know is is so invasive as the implantation, to something that's less invasive. And how far out I guess, do you think we are from things that are actually wearable in the real world, and not just things that are used in these controlled environments?
David Friedenberg
Yeah. So as I mentioned, we were very interested in seeing if we could get that similar sort of volitional control of movement while taking the surgical piece out, just because it's, it's expensive, it's risky, it's, it's very complicated. As you mentioned, it's a one of the most highly regulated types of trials that you could possibly imagine. And so as we were building some of the sleeve technology that went on the arm as part of the brain computer interface system, we also started to think about, well, we have all these electrodes on the arm. What if we started listening on those electrodes as well as stimulating and so we started co developing this as that as the trial with Ian was ongoing. And so it took a few years for us to get to, sort of our prototype state, and we put the prototype on. A couple of our participants had some good success there, and then it's gone through a couple iterations of improvements and refinements. And so some of that stuff is, how fast can you move? Is a lot of it's based on, you know, how much money you have to move that fast. The more people you can put on it, usually, the faster you're going to move. But that, you know, the development of that system is, I would say it's probably been about five or six years, and we're now testing that system out in a large trial that's funded by the Department of Defense, where we hope to enroll 12 participants with spinal cord injury, and they're going to use that system for about three months at a time, and we'll be testing to see how well they can use that as both an assistive device where they're it's helping them actively to manipulate objects in their environment. And then we're also evaluating, over the course of the three months, do we see improvements in their performance over time when they're not using the system? So thinking about is potentially something that you could use as a rehabilitation device where someone comes in for a fixed amount of time, they use the device, and then they go on with their lives, and hopefully they're better than where where they started.
John Bailer
So So with this new device, this sleeve that's being used there, there isn't the brain surgery, there isn't this kind of implantation of this brain computer interface. Is that? Is that correct?
David Friedenberg
That's right, and that's one of the big benefits. As our participants are able to roll into the clinic, we're able to put the sleeve on them on day one and begin using it. So wow, to get into the brain computer interface, there was, there was a long, long process. There was a lot of screening, there was a lot of imaging, there's surgery, you know, you have to actually clear for surgery, all these different steps. And so you also mentioned the the regulatory components. This is classified as a non significant risk trial, which means that the oversight is more on the IRB. The FDA defers a lot of that to the IRB, and so it simplifies some of the regulatory pieces as well.
John Bailer
You know, I was just thinking about, well, the initials, the initial study, where you have kind of these two components of complex models that are processing information and then yielding output that made to another model. So I was, I was thinking, Is it really hard to determine when you're not getting the activation you want? What's causing that? I mean, there's, it seems like there's multiple places where, where models may may be failing you, because you're you were trying to basically track what the brain signal might be, or that, that you're intending, and then this muscle activation that you want, and is it? And in both cases, you have some processing of information. Is it tough to figure out kind of where things aren't working when, when you're not getting the desired result?
David Friedenberg
It can be very challenging. Yes, you know the good thing with our with the newer system that that's used in electromyography, because it's non invasive, like we can also put the sleeves on people who don't have injuries, and so that kind of gives us a ceiling on what we could expect. And then as soon as you put it on someone with a spinal cord injury, we also have trials going on with stroke. The reason that they're in those trials is because of those injuries. And so necessarily, a lot of times those signals are either very faint or kind of scrambled, and from the way that you would expect if there wasn't an injury, or in some cases, just non existent. So some people with very severe spinal cord injuries, there's just no signal getting from the brain to the arm, and so this device isn't well suited for that. But what we found is that with most of the participants in our trial that have a spinal cord injury, there's still some signal getting through, and even if it's faint or a little garbled, we're still able to interpret, you know, differentiate between you're trying to open your hand, you're trying to close your hand, you're trying to do a key pinch like you would to put a key in a lock and turn it. And then we're able to translate that activity using our algorithms. You know, the other piece of that, which was exciting for me as a data scientist and a statistician is, you. We're collecting the training data as we go, and so part of that is also coming up with that paradigm of, how do you collect the training data? What are you asking the participant to do, and making sure that you're setting yourself up for success on the data side, the counterweight to that is that whenever you're asking your participant to collect training data, that's time, right? And so you're you're trying to balance not having allowing the participant to do the kinds of things that they want to do, which is being in control of their hand and using the system. But there's this calibration and training piece that comes in at the beginning. And so that's actually one of the things that we think a lot about as our data science team and algorithms team is, how can we get good performance while minimizing the amount of time that they're spending calibrating I'm
Rosemary Pennington
glad you brought up this sort of data side of this, because I was I and I think maybe in a chance article, you mentioned how in these early studies, this data sort of was protected, right, like you had people who were doing this research, But it wasn't shared widely, and how it sort of created challenges for researchers who are trying to do this work, because you're not quite sure, like, what is the base, right? There's no kind of shared understanding of what that base is. So how do you, I mean, you've mentioned it, I think, a little bit here, where you're sort of working with these people, but how do you get around that challenge and figure out, like, what is the base that we're looking for to be able to compare over time, and how has that sort of protection of data changed at or maybe it hasn't, I don't know.
David Friedenberg
Yeah, those are really important questions. And it's also something that, you know, NIH and other funding agencies have weighed in on heavily. So there's always, especially with the brain data, there's a privacy aspect to that. So one of the things that that we observed with our participant is we could kind of tell from the brain data when he was about to get sick, but, you know, there's we're recording from the brain, and you know, the brain, we're listening for, for motor intent. We're looking for arm and hand movements, but you're recording whatever you're recording, and so you would see changes based on anything from like, you know, medications, caffeine, sleep, stress, all those kinds of things. And, you know, at some point we don't really know what. We don't know of what someone might be able to understand from that data. And so, for the most part, that kind of data, we tend to hold it very close, just because, you know, we don't want to be violating anyone's privacy or or putting out something that that someone could find out, something that the participant might want to keep private. There have been some, some groups that have released some bits and pieces of brain computer interface data, and I think that's helpful for moving the field forward. But I, you know, I'd be remiss if I didn't say there wasn't some privacy concerns around that. When we talk about the the EMG data, I think it's a little bit less controversial, and there's something that we hope to as we complete some of our studies, to actually publish some of our data sets and allow other researchers to look at it and and try to beat our algorithms, or try to come up with other ways of doing things or extracting different biomarkers. It's one of the fun things we've been doing with our EMG data recently, is there's lots of information in there, and can so can we track progress in terms of their ability to move just based on on that raw EMG data? And so far, the answer seems to be yes. We're still, still doing some research and putting together the papers on that, but we're able to extract all sorts of different information from that EMG activity, and see how strong that communication channel is, and see how that communication channel is changing over time as they go through our interventions.
John Bailer
So you've been involved in this for now quite a while, and you've seen just dramatic change. I mean, you've seen this change from from brain surgery and this connection to, now the sleeve being the primary driver of this. I'm going to ask you to put on your your prediction hat, you know. So can you know, are you seeing a future where, where this is, this is now kind of going to be a, an approved medical device that someone will be using in their home to help with activities of daily living. You know, it seems like an aspiration, but do you see that in the future, and how far is way? Do you think that future might be?
David Friedenberg
Well, there's two pieces of that, so the brain computer interface stuff is absolutely moving forward in lots of different ways. So there's a really exciting paper that that recently came out in New England Journal of Medicine about using a similar kind of brain computer interface to what we were using, pairing it up with large language models and using that as a speech prosthetic for someone who had lost the ability to speak due to ALS. And it was incredibly emotional. I think we talked about it in our journal club, and everyone was kind of teary eyed watching the videos. You know, I think that technology is ready to help that population very, very soon. I think there's some other questions around, you know, how do you create a sustainable business around that? Because some of these populations are very small, but from a tech perspective, I think, I think some of those are very close.