Superpowering the Human Body : Episode 2
The second episode of CIONIC’s podcast Superpowering the Human Body, features a conversation with Akshay Chaudhari, Assistant Professor and Research Scientist at Stanford University.
It was a privilege to talk to Akshay about his research on neuromuscular disease progression and to hear his thoughts on how machine learning will help bring about a future of precision health.
Listen on Soundcloud
The full transcript is below, edited for readability.
Jeremiah: Hi, my name is Jeremiah Robison, founder and CEO of CIONIC. We build bionic clothing to enhance human performance and overcome disability. Welcome to our second episode of our podcast called Superpowering the Human Body, where we explore the science and technology of human augmentation.
I’m joined today by Akshay Chaudhari, Assistant Professor and Research Scientist at Stanford University.
Akshay, thanks for being here with me.
Akshay: Hey, thanks for having me Jeremiah. I’m really looking forward to it.
Jeremiah: Now Akshay, you work in the Department of Neurology — you have your own lab there. Why don’t you tell us a little bit about your work and your journey in getting here?
Akshay: I’m an Assistant Professor in Radiology and my primary interests really started off with trying to understand the human body. What is healthy? What is disease? How do we really define the transitions from healthy to disease activities? Those are the sorts of questions that really drive the research that I do and the research collaborators that I have.
I got started in Radiology because I had a technical background. I wanted to do something with my engineering skills, but I didn’t really know how to make an impact in healthcare. And, I am very impatient so I don’t necessarily like working today on something that will maybe see a clinical impact next decade or two decades down the line. So, when I was trying to decide what to do with my PhD, I really wanted to work on something that could be readily translatable. I thought medical imaging could be a really nice way to do this where we can build things, and we can deploy them in the clinic, and hopefully have an opportunity to actually impact patient lives. So, that’s one of the reasons I decided to start with the field. I didn’t necessarily know what the field entailed but it seemed exciting and it seemed fun at the time, and looking back — that was almost 10 years ago now — I’m very glad I’m in the field.
There really is a potential to impact lives at a population level perspective. Be it by developing new ways to look into the body, be it into developing new ways in which we can actually analyze what’s going on, and really trying to understand what are some of the early manifestations of disease. Once we understand that, how can we effectively intervene. That’s what has always inspired me in the past and I think that is what will inspire me in the future — time will tell.
Jeremiah: That’s fantastic. One of the areas that you talked with me about is this idea of quantitative MRI examinations; the ability to combine deep learning, computer vision, advanced MRI data acquisitions — and combine those into an advanced reading of MRIs that is quantitative. Describe to me how that works, how it’s different from the current standard of reading MRIs, and what can really be gained from this approach.
Akshay: The way I try to think about all of these advances in MRI, which is what I typically end up focusing on, is through this lens of value. There’s a large emphasis in the healthcare system about value-based healthcare. At a simplest level ‘value’ can be defined as some benefits over some costs. If we can maintain the same level of benefits that a medical test can provide by lowering costs, we’re creating value. Or maybe we can keep costs the same and increase the benefits of an exam, that increases value.
What I try to do is I try to come up with ways in which we can have our cake and eat it too. Can we increase benefits and can we lower cost simultaneously? That’s where I think quantitative MRI can really come into play.
Normally, when clinicians or radiologists are interpreting MRI images, probably 95–99% of the examinations are qualitative. So they’re looking at these gray scale images and typically a radiologist may say “hey, I think this region has a tumor or there’s some abnormality here” because it looks a little bit darker than the region next to it or maybe because it’s a little bit brighter than the region that it should be next to. Can we go beyond that? Can we have more of a quantitative assessment of what is going on physiologically? And if we have this quantitative assessment — if we have any sort of serial imaging — can we see how the body changes over time? With these qualitative assertions, it’s really hard to understand longitudinal variations. But once we come up with quantitative metrics, we know that there is some correspondence to a physical quantity. Modalities like x-ray or computed tomography or positron emission tomography, all of those are fundamentally quantitative because we know what the value of the pixel corresponds to in terms of the physical unit. But the challenge with MRI is that the signal that we generate is just a function of so many different variables, that it’s hard to single out one specific variable. That’s what my research has been focusing on for the last five or six years, specifically in the context of musculoskeletal imaging.
I’ve always played sports — all my friends are athletes — I ski a lot so I’m always thinking about my knees. That’s been my primary focus : how can we understand different injuries and degenerative diseases that can occur in the musculoskeletal system, but with an emphasis on the knees.
Typically if we get an MRI scan of the knee, it takes 20 to 25 minutes at best and everything that would generate is qualitative information. So what I’m trying to do is answer “can we reduce the amount of time that it takes to create that imaging without compromising diagnostic information?” That reduces costs in our value equation, but at the same time, can we come up with additional metrics that we can use to look at very early changes in the tissues.
One of my research focuses has been to try to study early osteoarthritis. Osteoarthritis affects around 30 million adults in the U.S. Costs around 200 billion dollars per year, but we still don’t know what causes the disease and because of that we don’t know of any therapeutic interventions that we can have. So we’ve started looking into some ways where we can quantify some of the changes in the articular cartilage using MRI-based methods and because of that we’re seeing very early manifestations and change over time, with tissues like the cartilage and maybe the meniscus in the ligaments. The way I think about that is that it can give us more benefits from an imaging examination. We’re just getting more value per unit time or per unit cost. So that’s really a nutshell of my research.
Jeremiah: I talked to some other folks who are doing similar things in the oncology space, where they’re trying to bring down the time in the MRI machine and take advantage of some of the time slack that you have within these MRI centers, and do more frequent measures over time. What will a clinician do with that information once they have it? How would that change the course of treatment?
Akshay: You really hit the most challenging part of this. We can always create more information, but how do we utilize it in the context of a person’s health and any sort of interventions they can get? I think every field will have a different perspective. With Oncology, the way you interpret the information may change what sort of chemotherapy you may provide or what sort of surgical intervention there might be. But maybe in the realm of musculoskeletal imaging, if I have information about my cartilage showing some early wear and tear, maybe I’ll reduce weight bearing exercises. So I think it’s still a nascent question because we’ve never really been able to acquire this information in the first place at a large scale. We’re trying to figure out how to answer the question and there’s no easy answer.
The other challenge that it poses is whenever we’re generating new information, someone has to sift through and try to interpret it. Our poor radiologists, they’re burdened by how many images we’re generating. They’re always trying to increase the workload that radiologists have. So, in addition to just creating more information I think there should be an emphasis on creating ways to parse through this information. I think that’s where a lot of advances in machine learning can come in. If we can automate some easy tasks that a radiologist may have — like growing some regions of interest or filtering out all the normal cases — that may free up some time where they can actually spend more time with the specific patient.
Right now, a radiologist just doesn’t read an image, they try to interpret what the image means in terms of the symptoms the patient has, in terms of the medical record. If we can actually allow our clinicians to spend more time looking at a patient more holistically rather than just through a single snapshot that imaging provides, I think that has the potential to have some longer-term consequences for the good.
Jeremiah: Absolutely. Yes, we talk about human augmentation and what it can do for the patient. But in some ways we’re talking about augmenting the doctors with insights — with actionable insights — on this data. Obviously machine learning plays a big role in that. Talk to me about what those advances have meant for you and your work over the last five or six years and what you look forward to in the future in terms of machine learning and its capabilities to augment healthcare.
Akshay: I think the last five or six years, like you mentioned, have been a really exciting time because that’s where a lot of advances driven by deep learning have come into play. The first big deep learning paper with modern hardware only came out in 2012 and we’re seeing so many advances that that has spurred. So it’s a really exciting field. It’s hard to keep up with it sometimes, but within the context of medical imaging, I really see two major areas where deep learning and machine learning broadly can really help us. One is before an image acquisition process and one is after the image acquisition.
Most of the machine learning that we’ve seen has revolved around questions like “will machine learning replace radiologists?”… “will they automate all the roles that clinicians and radiologists have?” We can try to ask those questions after an image has been generated — and I don’t think we’ll be replacing our radiologists; we will be augmenting them to make sure that they have reduced workloads, less burden of how many scans that they have to read through, and just being able to free up how much time they can devote per patient.
I think that is the one potential that machine learning can have is trying to interpret some of these images in the context of other similar patients, in the context of the patient’s past and in the context of a multi-modality work stream, too. Most machine learning algorithms for image interpretation right now only start with an image. There’s companies that are trying to do digital mammography interpretation, etc. At that point they only look at the image, but a person is more than just the image. They have a medical record that’s associated with them. They have a prior history. What I look forward to in terms of some of these advances is how can we really combine information from different modalities and try to create patient-specific insights.
Right now, I think that’s a little bit too early because most of the papers and most of the products that I see with machine learning always treat a test set. They’ve acquired maybe a thousand patients and then they do these group-level statistics. But a patient is not a test set; we’re not treating a test set in a hospital. How do some of the machine learning predictions work on a per patient level? If we were to simulate what their treatment would be it’s important to understand what are some of the longer-term outcomes. Being able to integrate different sources of information like the clinicians do on a daily basis is, I think, the next frontier in this image interpretation standpoint.
The other perspective is even before that image is created. So with MRI, like I described, it takes around 20 or 25 minutes to just get the images that we need so that we can interpret it. And that’s just because there’s a finite amount of data that we have to capture. In MRI, we refer to this as “k space”. But essentially if we’re trying to interpret a knee MRI scan, I need to acquire a hundred percent of the raw data which is required to build that image in the first place. Well, acquiring that hundred percent data takes 20 minutes. So what if I cut corners a little bit? What if I only acquire 75% of the data and then come up with some ways in which I can still use that under sampled data to reconstruct the images without losing image quality. And then, can I take it one step further. Can I only acquire 50% of the data? Can I only acquire 10% of the data? And I think that is really being pushed by machine learning. Where we can reduce the extent of data required to build the images? And then once we have the images, we can do whatever we want with it. We can send it to the radiologist… we can send it to an algorithm. But now we’ve fundamentally reduced the time required to build that image. So instead of a MRI scan taking 25 minutes, maybe it will take 10 minutes — and a hospital likes that because they can double how many patients come in, we can decrease the waiting lines and maybe we can just make our Healthcare a little bit more accessible to the world.
So, those are the two areas that I’ve seen generally disparate, but I think over time we’ll be able to solve these in an end-to-end manner. Instead of saying my image creation is one algorithm and my image analysis is one algorithm, we’ll be able to do this jointly. I think that’s where I’m slowly but surely seeing the field go.
Jeremiah: Absolutely. You talk about these disparate data sources and we’ve seen in the world of cognitive behavioral therapy and psychology, this use of online data sources: what is their activity within social networks to influence and understand the progression of care for them? We (CIONIC) are in a space where we are sensorising the human body and having body-worn sensors and an increase in that amount of data. Then you have the patient health record and the electronic health record. Again, we’re talking about a lot of disparate data sources and ones where when we talk to certain clinicians they say “Look, I don’t even want to be responsible for that. I don’t know what to do with it.” How can we integrate those in a way that is actionable for those clinicians?
Akshay: I wish you had an answer for that because I think that’s a question that everyone really has. I think that’s another area where some of these machine learning based methods can actually help us combine these multiple different sources. Let’s say we are trying to combine some Imaging with data from a wearable device that we may have. Well, an imaging data set is maybe two or three dimensional images, but it’s a single snapshot — there’s only one time point. Now, if we compare that to trying to analyze my heart rate over time from my wearable device, that’s a one-dimensional signal, but it can last for days. So if I had to approach my clinician and just show them the raw data, they have no idea how to interpret it. And for good reason. We don’t know how much signal exists within that data, what is noise and what that signal can actually contribute to in terms of physiological outcomes.
If we can come up with different techniques where we can take our long term data streams and be able to build salient insights from that. Perhaps instead of just having a long-term timeline of my heart rate data, maybe the first level of my data analysis algorithm could be to classify when I’m asleep and classify when I’m awake. Then when I’m asleep, we can look at my different sleeping levels. When I’m awake, can we classify my different activity levels? Was I just sitting on the couch all day? Did I go for a walk? Did I go for a run? And be able to look at these long-term longitudinal data sets from a more granular manner and to be able to build up these granular insights, and be able to combine these insights along with medical record data or Imaging based data. I think that is really hard to do for a human. Because even though that task may be easy, doing it for a long term time scale is challenging. So I think machine learning has the ability to extract those insights, but extracting those insights in an efficient manner.
Jeremiah: I’ve always been a big believer that the future of healthcare really does center around these continuous streams of data that are really guiding the decision process on when to take deeper looks at higher precision, higher granularity of information. This idea really leading to integration of those streams in a way that you can provide precise, individualized health care. What we talk about typically as “precision medicine”.
Beliefs on our ability to deliver precision medicine — anywhere from pharmacology to physical therapy to integrated behavioral therapy as part of that whole entire stack up?
Akshay: Certainly. I think we can maybe even go beyond that and this is something that we’ve started really exploring at Stanford. Precision medicine is a term that we’ve been using for quite a while, and precision medicine really entails being able to treat a patient once we’ve figured out what may be wrong with that patient. But can we go even earlier along in the disease process? And the term that we’ve coined for that at Stanford is this notion of Precision Health. Can we actually make sure that no one gets sick in the first place?
A lot of these ideas have been inspired by our past Radiology Chair at Stanford, Sam Gambhir, who unfortunately passed away a few months ago. He really espoused this idea that a successful healthcare system is where our hospitals are empty. That there is no need to have a hospital because if we have sensors at home, we can ambiently monitor patients and try to come up with warning signs or have some of these triggers to then say “Hey, I think that you’re on a negative trajectory; why don’t you go to your doctor right now before it gets worse?”
I think that’s where the pervasiveness of a lot of new hardware that we’re building can really come in handy. There’s these notions of smart mirrors, there is wearable devices that we can have at home, there’s audio and video sensors. It brings up other questions about privacy, obviously, but at least from a technology standpoint, there’s a lot of room that we can explore. I really do think that is what the future of healthcare will be: trying to minimize having to treat people, because we have enough signal available to understand that they’re getting sick and try to nip the problem in the bud.
But I think that’s a really long term vision. So the more we can do right now, with trying to understand and acquire new data sets and try to interpret that into actionable insights. I feel like every little step is needed before we can really reach that precision health framework.
Jeremiah: I love that term precision health, and I also think about some of the work that we are doing to close the loop in that precision health system. So, not only can we sense what is happening in the patient’s life, drive deeper care decisions, but also start to provide care based on that closed loop system.
I think there’s a really interesting future — and we can look to some of the things that we’ve seen in the past — — diabetes care for instance. It started with a finger prick and then separately administered insulin, and now you have systems that can do the whole cycle. It’s advised and there is a clinician in the middle, but over time those things can get smarter, more personalized, more individualized and more aware of context as you described; which is one of the key abilities to provide actionable insights to that data.
I want to switch gears for a second and talk about some of the work that you’ve been doing in sport. And so we talk about health and oftentimes, we think of someone being sick or injured, but really some of the work that you’ve been doing to prevent injury, get players back to the field and also to optimize performance.
Can you tell us a little bit about that work and how you’ve been engaged with the elite athlete community there at Stanford?
Akshay: There’s a couple of different studies that we have going on.
I think primarily, the one question that we’re trying to answer is “Can we prevent injuries in the first place?” And then the second question revolves around “If someone is injured with a sports-related injury, how can we get them back on the field or how can we get them back to the same quality of life that they want?”
In terms of the research studies that we’ve been doing to prevent injury, we’ve been working with some basketball players and we’ve been working with swimmers to try to understand how the exercises and the season they have of their play affects physiological changes within their knees.
So we’re really interested in “Jumpers knee syndrome” where basketball players typically have a lot of pain in their knees. Pain in the knee is very nonspecific, but it can be really debilitating if you’re an athlete and if you’re really trying to make it to the first team or if you’re really trying to perform at a high level. So, we’re trying to understand how some jumpers load their knees differently than let’s say some swimmers — who are still working out, who still have a very active training regimen — but their sports are a little bit different. How do the hips in swimmers work differently if they’re doing a back stroke versus a freestyle? We’re really trying to come up with ways in which we can look at a cause and effect relationship.
Because causal relationships in healthcare are always challenging to figure out.
By trying to look at different sports, we’re just trying to understand how do different athletes work? How are they actually using their body? How are they using different muscles? And what are some measurable impacts that we can have? Maybe that is through any sort of qualitative pain scores. Maybe that is through some quantitative MRI imaging, and once we can have these multiple snapshots over time, can we go back and if someone happened to injure themselves, can we now correlate some of the changes that we’ve been able to see — and hopefully, by being able to correlate some of those retrospective changes, maybe that will guide us in being able to prevent those injuries from happening in the first place.
So I think that’s kind of one perspective for research, and the other one is we’ve been really focusing on athletes who’ve had a tear to the anterior cruciate ligament or the ACL. Very common injury in sports like soccer,basketball and skiing. Right now, it’s really challenging to figure out after you’ve been injured, when can you return back to sports?
If you are a football club, if you’re a soccer team, you obviously want your athletes to be back as soon as they can. But once you’ve injured yourself once there’s a higher likelihood that you’ll probably enjoy yourself again. So this notion of return to play after injuries is a very qualitative science right now. Typically, you may go to your physical therapist — they’ll try to just come up with a rough heuristic measure. “Hey, I think your muscles on your injured leg are more or less the same on your muscles on your contralateral leg. I think you’re ready to play”.
This whole field of going from these qualitative measurements to quantitative measurements using Imaging, using patient-reported outcomes and using some wearable devices is probably the next frontier where we can actually quantitatively figure out whether or not a person may be more susceptible to injuries after they’ve been injured — — and hopefully that will also give us some insights into preventing injuries in the first place if we see a repetitive patterns in some physiological degradation of different tissues.
Jeremiah: It’s really interesting to think about. We talk about it at CIONIC, we have these populations: we have folks who are suffering from a neuromuscular disease like Parkinson’s or MS and then we have people who have an orthopedic injury that they’re recovering from — who are elite athletes optimizing performance — and we really talked about this as a continuum of abilities. Not being disabled or abled but really on that continuum. What have you learned from working with each of those populations that has benefited others on that spectrum?
Akshay: That continuum can be challenging because oftentimes when we try to interrograte a person physiologically, be it through Imaging or maybe through an appointment with their orthopedic surgeon, you can have different trajectories of diseases. I’ll give you an example of someone who’s had an ACL tear and someone who’s just had chronic osteoarthritis for their whole life. If you just look at their knees or if you look at their hips, they will look identical. If we only limit ourselves to that one snapshot of health, we won’t really be able to appreciate that the trajectories that they’re coming to that same snapshot are very different. Since they’ve had different trajectories: one person is a young athlete who has injured themselves, another person with osteoarthritis may be older and just has more degenerative changes — they end up at the same level, but they have very different trajectories or future progression.
I think that’s really where the continuum can come into play.
If we can interrogate these people more in terms of what is happening physiologically, then we have more data points on that trajectory and it all leads in towards this notion of personalized health or maybe personalized medicine, where we can really follow the people over time and we can use the insights that we’ve been able to get with longitudinal measurements to predict where they end up on that trajectory. And once we know that, then we can try to tailor effective interventions. Maybe for the athlete in that case, an intervention could be resting for six weeks or not being weight bearing. But maybe the same intervention with the same type of knees for the older person with osteoarthritis could be being a little bit more active and building up a little bit more muscle strength. So, even though your images may look similar or maybe your wearable data may look similar, there’s a history that the patient has, so the more we can incorporate that, I think the better treatments that we can tailor for them in the future.
Jeremiah: Absolutely. Longitudinal data on progression and how interventions can impact that progression — I think for sure is one of the big trends that we’re seeing in healthcare. And with that, give me your prediction is look ahead 30 years from now. Are we all cyborg? are we all mechanized or what does the future look like for us as a species?
Akshay: Can we pause recording when I give you my answer?(laughs)
I think that’s a tough question to answer because if we look back even 30 or 40 years ago, I don’t think any of us could have predicted the sorts of technologies that we have right now, I’m hopeful that we can try to achieve this notion of Precision Health. Where we can try to really stop ourselves from getting sick in the first place. Maybe that is through ambient monitoring — and there’s a lot of techniques in which we can actually do that. But instead of really focusing on treatment, I think focusing on the cause of diseases is the next frontier that a lot of people are really interested in and hopefully intuitively it makes sense, right? So that’s maybe my naive vision as to where we can go in the next few decades, but as someone who has been working in the space, I’d be curious to hear your opinions too.
Jeremiah: We talk about our work in the spectrum of human augmentation and one of the things that’s been very important for me and something that I keep as a North Star
is a world beyond disability. And the word ‘disability’ is an interesting one: it is something taken away or not able to be done.
Well, technology has always provided us an opportunity to do something that was not possible — whether it was the way that we share information … the way that we interrogate that information …the way we communicate with others. How does that marriage of machine and human actually help us to overcome limitations to what is done? At the point where — regardless of what my disease state or my injury state is — if I can do all of the things that everyone else can do, then there’s no word ‘disability’ anymore.
I’m not saying we won’t have diseases. I’m not saying we won’t have these injuries. And in fact, if you look at the predictions of the world, the CDC suggests that by the year 2050, twenty percent of people will suffer from an injury or limitation to their mobility that impacts their quality of life. How do we as a group of scientists, researchers, clinicians and technologists come together to make it so that — regardless of what is happening with the physiology — people are able to do the things that they love?
Akshay: That’s a really interesting point that you bring up, Jeremiah. I feel like there’s been a lot of research into increasing the quantity of life. Coming up with new therapeutics for oncology or any sort of neurological disorders. But as we’re getting better at those, people are getting older. People are not dying of the same reasons they would have 20 or 30 years ago. And, because of that, I think the other problem that brings up is, since people are living longer, now they have quality of life problems.
As we’re getting older, maybe it’s harder for us to move and just to be mobile over time. I think there’s different tracks of research going on in the future. It’s important to be able to focus not just on quantity of life, but really quality. I’ve definitely heard some colleagues and researchers say “What good is quantity, if you can’t have a good quality of life?”
I’m sure there’s a spectrum within that perspective, but I think being able to focus on trying to maintain both, will be important in the next 30–40 years like you mentioned.
Jeremiah: I think it’s absolutely possible and it comes with people like yourself who are dedicating their lives to this work. I’ve been super excited to talk to you today and hear your perspective on things, and will continue to follow your work and look forward to collaborations in the future.
Well, that’s all for this episode of Superpowering the Human Body. Thank you for listening.
And thank you Akshay for joining me today and sharing your perspectives and your work on precision health and the future of human augmentation.
Akshay: Thanks very much, the pleasure was all mine. Thanks for having me.
Jeremiah: Please subscribe to our podcast on Soundcloud or YouTube so you never miss an episode. Until next time, this is Superpowering the Human Body.
