[00:00:09]
>> All. Right. That was right thank you thank you but see there's a yeah I think we have audio that was a while ago so the athlete part so we won't get into that at all so you guys are all I guess like literally in the back row or the 2nd the back row.

[00:01:15]
I don't know if you want to come up or at the very least definitely Please ask questions throughout and kind of engage so as more of a discussion in fact unlike most probably faculty members talks you've been to I have not loaded the powerpoint with a 100 slides in 15 minutes in fact I have less slides.

[00:01:36]
Then the number of minutes that I'm presenting for so by doing so I encourage hopefully some interaction and it would be a lot more interesting and fun your questions a lot of times or some of the situations from students and. Adjacent groups are usually some of the best questions I get on our research and sometimes they lead to collaboration you'll see that a lot of our projects are not just siloed off in my group but highly collaborative among many different groups at Georgia Tech and beyond so please do that's your charge I guess for today is to please jump in and ask questions.

[00:02:15]
I'm going to actually just tell you about 2 of our projects and really it's each of these is more than a project it's really 2 areas that we're working on the lab each of these areas has multiple projects in the lab going on that kind of synergistically build on each other and I won't get into some of the other things we're working on but not because I'm not as excited about them just because I have to choose something to talk about in the 15 minute time slot where I can still get into some depth so you get a feel for the kind of work that we do we're definitely not the typical I guess nano tech sort of lab everything we do is at the level of building systems that interface to the body a lot of the work that we do is often hospitals or clinics children's health care kind of facilities where we're working directly with doctors and patients a lot of times patients families as well and the goal is really to translate work from the lab setting typically sensing electronics actuation from lab setting into a clinically impactful and important sort of domain I'm going to talk about 2 of these areas the 1st will be wearable joint health monitoring joints meaning we need elbows that kind of thing and the 2nd area will be some of the work that we do in cardio mechanical sensing So sensing the mechanical aspects of cardiovascular function and I'll start with the vision I guess for our joint health monitoring project and not so much to give you the details of this project but to give you an idea of our philosophy as a lab and what our process is going about some of these projects so most of health care and health related sensing diagnostics and monitoring today happens I'll ask you does most of it happen at home or in the doctor's office so how many of you think it happens at the clinic or the hospital and how many people think most of it happens at home.

[00:04:18]
That's cool that's interesting I'd love to hear more about that later but most would think that I guess is not the case that's not always going to be the truth but so maybe you're thinking about for example the abbot freestyle glucose monitoring device that you put on or maybe glucose sensing overall which has to happen at home for diabetes care but pretty much beyond a couple of examples most of health related sensing is primarily happening in the doctor's office in the hospital or clinic So our philosophy is that we have all these miniature eyes sensors powerful computing.

[00:04:57]
Low power electronics micro power nano power electronics combined with materials advances that allow sensors and electronics to be put on the body in ways that aren't necessarily obtrusive and encumbering for the person using them and we have all these advances in the machine learning world that not only teach us about big data which is talked about a lot but more also about how to discover meaningful information within signals that you're measuring maybe that you don't understand well because nobody has studied them before and if you combine all of these things together then you might actually have an ecosystem where you're measuring things from people at home outside of the hospital outside of the clinic in their normal daily living activities you're deriving some information from the sensors on the body that's clinically relevant and meaningful for actually changing that person's health and hopefully improving it.

[00:05:54]
And then through some sort of system whether it's through the patient themselves through the physician through the caregiver you're providing some feedback now that feedback might at 1st just be to change medication types or medication doses eventually it might be in more automated closed loop sorts of systems where you modulator the physiology to get the body to a more healthy state but that's the sort of thing that we're working on and the 1st example of that project that sort of philosophy I'll tell you about is for a joint health monitoring project so we had I had this idea when I 1st got here that you could measure the sounds from the joints I really felt like.

[00:06:33]
In the end your movements are sort of mechanical right and then it's biomechanics and any time you have surfaces moving and rubbing and bumping into each other there's going to be sounds emitted from that and I had studied before I got here basically I had worked at a pro audio company building miniature microphones kind of like this one but smaller and better for listening to sound from people and so I had this thought that we could actually probably do this for the knee then we could have these maybe in a brace and ultimately derive some information that matters about joint health which is very difficult right now unfortunately not too much is understood about these sounds for a basic level so some of my labs work has been in a cadaver model and this is in collaboration with some folks here from School of Biological Sciences where we have fresh Rove frozen cadaver limbs and we place our sensors which are these small in this case Piers electric accelerometers wideband Piers electric motors which are what you would call contact microphones on the knee immediately and laterally to the kneecap and we actually move the leg up and down as your leg would be flexed an extended stay in a doctor's office were doing an unloaded flexion extension exercise.

[00:07:51]
And as we do that we record these sounds using a data acquisition unit and at the same time we also record with the inertial measurement unit the angle that the joint is moving the trajectory that it's taking and the speed and we try to control that with this sort of experiment we want to do things like understand fundamentally where these sounds originate from and what do they propagate through so which tissues that they primarily propagate through to get to the surface of the skin but to even do that in the 1st place we had to validate this model as being relevant from a clinical standpoint so what we did was and this is a.

[00:08:29]
Ph D. student in my lab who's doing these surgeries but we actually induced a miniscule tear in the knee. In the cadaver knee and then we examine the way that that miniscule stared change the sounds in a very quantitative standpoint so what features of these sounds change when there's a miniscule stare versus a healthy me and of course to do that we had to also do a sham surgery which is kind of the control and then we did a minute 2nd me so so that's basically where after There's a minister you cut off the flap of the minutes because that's remaining But essentially this is a very basic kind of biological experiment to understand how these sounds change with a particular type of acute injury and the primary feature we found that was relevant was something called the Bee value this is actually used in acoustic emissions for structural health monitoring so meaning you know beams and those can sorts of large structures that are examined from a civil engineering standpoint so there's an algorithm that's used there and what we found was that for the terror you get actually quite a bit of a difference in the bee value compared to baseline and sham cases so I have some of these recordings here that you can listen to and see what you think I'm going to turn it up to 1st turn this down.

[00:09:55]
Leave it on a little. I just say quietly so this is for a leg that's being flexed and extended this in our healthy subjects so it's not from a cadaver model. Now for injured subjects so somebody with a A.C.L. tear so one of the 1st sets of recordings we took were in athletes with injury immediately afterwards.

[00:10:58]
So what you can notice hopefully from these recordings is that there are definitely differences and of course person to person there's also a difference in the sound so it's very difficult to just say that there's something simple like in this case it sounds like maybe the injured knee has a louder need sounds that's not necessarily always the case so we had to quantitatively examine what sort of feature may capture the differences that are there so one of the things we found that's most meaningful by the way can you hear me OK Is that.

[00:11:29]
I turn it back on correctly for the video also OK good I had to do a little audio song and dance that I've done for a few years so all right. So what we did find and we publish is very recently actually because it was really cool again this be value algorithm is based on looking at the distribution of amplitudes in these click events so you noticed from that recording that as the person is flexing a star extending in flexing the leg you hear these sort of.

[00:11:59]
Kind of events occurring basically these acoustic emission events they're called. If you then record all of the amplitudes of those events essentially think in a vector or matrix form and then if you look at the log of the number of events occurring on one axis versus the amplitude of the events on the other axis and you look at the slope of that that's what this B. value is that's exactly what they use for structures so and injured knee has a lower B. value meaning a lower slope here which means just that there.

[00:12:39]
Is a weighted more towards larger louder amplitude sounds that is quieter sounds which is what you heard so it's not that the overall R.M.S. power of the audio recording is louder it's that there are more loud occurrences of these clicks than quieter occurrences and that's the same with a structure for example a beam has fractures or other defects then you get more loud sound the Kurds is from that structure because of those defects so it essentially becomes a different type of acoustic signal generator so for the athletes that we studied as well so this is now the same metric we used and learned from the cadaver model that were then applying in an actual set of recordings from human subjects and you might ask if we have this recording Why not just analyze this directly and we did that as well but when we did we had to go with some really heavy hitting graph mining kind of algorithms to get any information out of signals because of the fact that they're noisy when you measure signals from humans there's also the mike rubbing on the skin there's issues with the person maybe not doing the movement exactly the same every time.

[00:13:55]
There's issues with human to human variability and even in the athlete population how do you know that they injured their left knee how do you know the right knee is not slightly injured as well right most people for example who are doing soccer basketball football or other sports that are really impactful and stressful on the joints they may have an injury on one side but they may also have quite a bit of wear and tear and possibly developing injury on the other side so those are some issues that make it difficult to directly work in the human population so in this case it was very helpful for us to 1st get the information from the basic sort of biological testing and then apply that in this domain when we did what we noticed was these are the same group of athletes 7 days before the injury so it's a taste after the injury to 4 to 6 months after the injury and after reconstructive surgery so again to be value is lower here than it is here there is one subject where we saw the opposite result we saw that with our machine learning approach as well that subject had a noisy initial dataset.

[00:15:01]
But this lower B. value in the injured case is exactly what we would expect when you look at the Contra lateral knee again you see variability each of the athletes has a different trajectory from one time to the other sometimes it goes up sometimes goes down but what you don't see is the same sort of average group trend in one direction and in almost all of the cases the contra lateral knee has higher B. values indicating that it's healthier in the compared to the injured side so this is some of the exciting results we got from these studies we're also really interested in using sounds to better understand not just the structural integrity of the joint.

[00:15:41]
Relative to some sort of injury or injury recovery in Riyadh but also for arthritis where you have disruption to the cartilage you have actually the bones themselves and potentially degrading and there would be different acoustic signatures associated with that arthritis potentially if there's an inflamed joint versus not so we've done some work in collaboration with children's healthcare this is with Dr Prahlad there where we compared for a group of about 7 control subject and 4 subjects with human arthritis that dataset has grown since then that we see differences between healthy controls and patients with arthritis and maybe what's more exciting as the patients with arthritis pretreatment to post treatment we see quite a bit of difference in the signatures of these sounds meaning that you might be able to use sounds as a way to understand early on if a treatment is effective for a patient or not which is really difficult and kids without right it's a very head erogenous condition where they have multiple different therapies and it's a little bit of a guess and check hoping that something will work waiting 6 months 8 months 9 months to see if the patient's symptoms and pain have been alleviated and then deciding kind of where do we go from there so quantitative measure that could give quicker feedback is a really attractive for that clinical community.

[00:17:05]
At the 1st slide I gave you an indication that where we want to go with this is to derive a joint health score from these sounds so some of the work that we've done for juvenile arthritis has been exactly that we're essentially we use a soft classifier where the output doesn't necessarily have to be just one or 0 but provides something that can be in-between so a little bit more like a regression kind of approach and when we do that when we train it with leave one subject out cross validation for the subjects which I am the control subjects we see a big difference between the sound characteristics from the joints in patients with G.I. compared to controls that much maybe you'd expect this sort of initial verification but what was more important for us was that from pretreatment to post treatment we saw a pretty big difference in most of the subjects but in some of the subjects actually you saw that the average scores did not improve that much and that's common in jail in a where those subjects happened to be the same ones where the physical exam was also showing that maybe their joint wasn't actually doing much better so we're now validating this at a larger scale as part of an S.F. bonded project to better understand the sounds in the context of an arthritis I want to stop there for 2 minutes just ask if there are questions about the joint health monitoring work and then I'm going to totally move on to a completely different area yeah.

[00:18:39]
Yeah. It's. A great question that is one of the most difficult things about the contra lateral side if you're injured on your left side you're going to put a lot more stress on the right side as you're recovering you know you're going to any activities that you do if you're walking up stairs are going to push off of the help healthy leg more often and certainly see differences in the contralto side there's no gold standard to compare against That's the issue there so yes definitely I would imagine that some of those athletes who are especially 4 to 6 months later they're starting to get into sports related sorts of even just plyometrics and other sorts of jumping type activities.

[00:19:24]
They can weigh differently to one side versus the other it happens I mean I tore my calf about a month and a half ago on my left side playing tennis and so I've been rehabbing and now my right calf is super sore and and tight so you know definitely I've started playing again and I can feel it so that and that's common of this commonly seen literature so it's absolutely possible hard to tell though because the lack of Gold Standard one of the tough things and the sort of work.

[00:20:02]
Yeah for sure so this I'm sorry this is definitely not amplitude So this is actually in fact so one of the things we did in this paper was we examine time domain frequency domain amplitude based features and even. Features which is more used for speech but still we wanted to look at if those sort of more acoustically relevant features would make sense if I remember right it was the frequency domain information that's most valuable for arthritis.

[00:20:32]
For the injury case I think it's more of a structural disruption so it's more events that are occurring almost like impulses to some extent right so in terms of frequency very broad in terms of time very narrow and it's more of the loudness of those events and how many loud events that are occurring I think that's meaningful in that case but again that level of diving into it we need to get into with the cadaver model and we have other experiments coming up where we'll be inducing the sound from inside of the knee measuring how it propagates the outside along different tissues and surfaces and I think that will really get at more of these questions in more depth.

[00:21:18]
But yeah that's a great question. In fact amplitude usually is not a robust measure compared to some of these other things like frequency but but in this case because it's the distribution of amplitudes rather than the absolute Apple to value that's what makes it more robust OK well I'll get into some of this work now that we've done so you're going to hear now how many of you work actually on accelerometer design or gyroscope design in your research a while really it's surprising how many of you have used an accelerometer gyroscope that's also surprising how many of you know what an accelerometer it's OK it's getting better so that's really actually helpful for me to know because then I can at least give you the one sentence so and accelerometers a device that measures linear accelerations a gyroscope which I'm also going to talk about in this next section of the talk is a device that measures rotational velocity.

[00:22:20]
OK So the goal of this work them out to talk about is to measure the mechanical vibrations of the body associated with the heartbeat and to use those to try to quantify cardiovascular health parameters so every time your heart beats inside your chest. It's actually kind of performing this he little torquing motion and the blood is coming out into the aorta which is the main vessel out of the heart and that's also it's not just vertical sort of head to foot it also has kind of this wrapping around type direction to it where the blood is moving out and around the body like that you could almost think of like a counter-clockwise if you're looking down at the body motion that's occurring internally so there's a lot going on there right in how hard our heart is beating and how much blood spurting out with each of the beats in the timings of when relative to the point where it gets the signal electrically to say start beating to the point where it actually physically ejects blood all of that has a lot of information that can be unpacked and used and exploited by medical monitoring purposes so I've been fascinated with these signals actually since I was a Ph D. student when I was a student I was hacking weighing scales to try to measure these vibrations as a person stands on a scale which you can but since I've gotten here we've been more interested in these wearable measurements using mems basic Celeron meters and gyroscopes.

[00:23:58]
So we have a device that we built this is in collaboration with one of my really good friends who is also a grad student at same sorry in undergrad when I was a grad student at Stanford and now he's an M.D. Ph D. at Northwestern but he and I have worked together for years to build these small patches that go on the chest that measure both electrocardiogram through the electrodes so it's mounted on the chest by 3 electrodes their silver silver cord gel lecture exists like you use for an E.C.G. and the patch also has a small very low noise accelerometer on Born on board 3 axis that measures the chest wall vibrations in response to the heart beat due to both the heart motion and also the blood motion in the chest cavity and what we've been working on is a National Institutes of Health funded project where we're really excited about trying to get this device and the weighing scale in the homes of patients with heart failure to see if we can potentially detect an exacerbation before it occurs there's some precedent to this where there's invasive implantable devices specifically was devised called Cardio memes which was actually founded by Mark Allen and some others Mark when he was here and that device is shown really good results in terms of being able to detect something called congestion so in heart failure the heart is really there's usually been a heart attack some component of some portion of the issue is dead the heart is not able to pump affectively so not enough blood is making it to the organs and tissues that's required that's kind of the definition of heart failure and unfortunately what happens is because the body is trying so hard to compensate for this deficit in a heart function through other sort of means that it has it's very hard to detect when that patient is really in danger of what's called a D. compensation event or an exacerbation where they need to be hospitalized.

[00:25:58]
So that's very difficult to sense but the issue is that because of the fact that those hospitals ations occur frequently there's about $30000000000.00 spent per year on heart failure of which about $12.00 to $15000000000.00 is on hospitalizations alone they're very expensive each time hospitalization occurs about $10.00 to $20000.00 for one of these patients and there are $6000000.00 patients that hard for the U.S. So heart failure is typically classified by looking at physical activity tolerance and the Bill of the of the person to modulating their cardiovascular function cardio pulmonary function really in response to an exercise stress or so because of this actually one of the primary tests used in the hospital to assess heart failure patient status is something called the 6 minute walk test so where what for one of us walking for 6 minutes is probably pretty easy right except maybe in this weather.

[00:26:54]
For a patient with heart failure it's really difficult to walk for 6 minutes and sometimes they can't even finish the 6 minutes that walk maybe for 3 minutes and stop so the distance they can walk is one of the measurements that people look at so what we're interested in doing though is specifically to look at these noninvasive wearable measures of cardiovascular mechanics in response to that 6 minute walk test so where we know we will not have the same level of sensitivity as a pressure sensor that goes in the Palminteri artery that directly senses congestion through elevated prominent artery pressures what we do believe is although our devices noninvasive we can measure the response to activities and perturbations in normal daily life and use those to extrapolate what we think the congestion status of the patient is in other words are they compensated are the compensated so we've been working with our colleagues.

[00:27:51]
At U.C.S.F. for a few years now on this project we've had about 70 patients so far enroll in our studies some of these patients are. Having measurements taken in the hospital setting in patients some of them are being taken in the clinic so outpatient settings and some of them are actually taking some advices home so we're going home measurements as well what I'll show you here is some of the results we've gotten from these patients in specifically in response to the 6 minute walk test so 6 m w t a 6 minute walk test so for a compensated patient which means a patient that has heart failure but the body is able to compensate for it at that point through its mechanisms and so they're usually out patients while of course they're sick they have heart failure of course you know they're not able to pump enough blood to the organs and tissues their body is able to compensate for it somewhat and remain in a state where they're outside the hospital and then we have these deep compensated patients that are of course hospitalized in patients.

[00:28:55]
And what we notice is that for the compensated patients this is a basically just. K.N. graph you can think it's OK nearest neighbor graph that captures essentially think of the feature space that's multi-dimensional from these signals that were that were measuring from the chest these are accelerations from the chest and what we've done is taken that N. dimensional high dimensional feature space and collapse it down on to 2 dimensions with this that's what the standard tool can and allows you to do but what you'll notice is if you go from rest to after 6 we have walked us you see a big difference in the features of the patients who have compensated heart failure whereas if you look at the compensated patients they look a lot more similar before and after this exercise stressor and that's we believe because of the fact that their body is not able to compensate for the exercise and so it's not able to.

[00:29:50]
Their cardiovascular function in response to that exercise as one of us would be able to for example and when we compare those 2 groups we do see a big statistically significant difference and more importantly again like with the audio data Although comparing between groups is useful as a 1st verification what we care about most is longitudinally tracking the same patients over time and seeing if their condition changing one way or the other reflects in the measurements that we're taking So in this case you see a group of 6 patients that have gone from Adam it to the hospital to discharge.

[00:30:29]
And typically what happens is they're diarists meaning that they have to your innate out a lot of their fluid they could be in the hospital for 2 weeks in some cases one week in some cases but they're doing quite a bit better and at that point they're compensated they're ready to be discharged and to become basically operations to go home.

[00:30:48]
Hospital in all of those cases we saw market improvement in this graph similarity score essentially it became less similar for the same activity meaning that their heart was able to actually change its function appropriately in response that activity and for one of these patients they were readmitted about 2 weeks later with volume overload and in that patient we did see the G.S. us increase again back to the D. compensated level one of the patients received a heart transplant and he saw a dramatic decrease in G S S indicating of course severe improvement in condition so that was pretty exciting and going forward from there also were interested in being able to measure some of these signals during activity so as sensitive as it is to look before and after a 6 minute walk test it would be really nice if we could get information while a person is performing some sort of walking test or other sorts of activities because that's when the information is richest that's when their heart is really being stressed and you want to know how it's responding to that.

[00:31:51]
So we've done some work on using this tool called empirical mode to composition to be able to remove some of the motion artifacts during walking and be able to capture these cardiac timing intervals with good for the ality it works fairly well this is for healthy subjects in this case and this is for the 3 mile per hour walking up to 3 miles per hour walking on the treadmill that we were able to do this this is what got us excited about maybe being able to do more with the patients with heart failure during exercise.

[00:32:23]
But before I even go there I'll tell you a little bit about some of the other sensors that we're using to measure the signal so we have a study of that is submitted now kind of in the revision phase where we collaborated with Dr ideologies group here. And what we were working on is determining if gyroscope based rotational velocity signatures might be able to augment the information that we get from these linear accelerations on the chest and specifically we wanted to focus on one feature which is of importance to us which is the section of aortic valve opening so when the heart has initially released blood into they order from from they are about so we ran a study where we used both of those signals we measured gold standard cardiac timing intervals as well with something called impedance cardiod graffiti and we wanted to compare the 2 of them and also fuse the 2 of them see if we get better results and I want to show you this so this is actually from someone else's paper this from 1909 but this is looking from the right of a patient afford the M.R.I. So in other words 3 D. vs time M.R.I. of the blood flow from the heart out through the area and you can see that the trajectory it takes as I was mentioning at the very beginning has this helix will shape to it.

[00:33:46]
So we hypothesize that a Adding gyroscope signals to these linear accelerations could help in detecting this important cardiac timing interval because of the fact that the blood is taking this sort of trajectory to its movement and the set B. We bought a size that specifically the gyroscope So the gyroscopes 3 axis we measure rotations along this axis this axis and of course this axis right and so we hypothesized that this one so the one going around the head to put the direction of the body would have the most salient information so it's good to have some sort of physical basis for what you expect to see and when we do actually look at the results we see exactly that so the error in G. X. which represents the gyroscope in this kind of around the head to put axis and in the linear direction a subsidy which is in and out of the chest that much we had expected from our knowledge of the signals before had the lowest error and the combination of the features from gyroscope and the accelerometer which is this result right here had the lowest error all together of everything so it did actually improve to include both the gyroscope and the excel or ometer and specifically the gyroscope access around there was pretty rich with information we're definitely studying that aspect of the signal more now going forward so I did mention that we wanted to do more during exercise we're doing exactly that so for the patients with heart failure they do a test called cardio pulmonary stress testing where you can see they'll be walking on a treadmill while wearing this mask which measures their breathing gases and the flow of air in and out of their body and of course of the same time there were blood pressure cuff E.C.G. electrodes and we're having them also where a patch.

[00:35:46]
As they do so we're measuring of course all the signals from the cardio pulmonary exercise testing R.C. pet and at the same time we're measuring our E.C.G. and our 3 dimensions of linear accelerations through the size of a cardiogram patch and you can see they're really clean so this is from a representative patient with heart failure you can see all of the rich details of these signals and they have some similarity from beat to beat So there's a lot of good information here to extract and this is what we're doing with the data so our goal is to be able to take our patch and measure the same kinds of things that you normally would require this cardiopulmonary excise testing set up to measure and specifically what we're trying to do is estimate something called P O 2 which is the flow oxygen into the body but really what we the only reason why we feel like we can correlate the V O 2 is because of you know 2 is derived from cardiac output so the flow of blood through the arteries so it's a cardiovascular parameter that we're measuring that's just what the machine outputs space so the idea is that we take our seismic cardiogram signals we window them we detect these artifact 3 frames using some envelope and sorts of functions we extract heartbeats and then we do some regression to see which of those features might be most relevant in predicting video too so this is an example of one of the subjects where you see the actual value of V o 2 versus time so as the person is exercising harder and harder on the treadmill V O 2 is of course increasing and our predicted value our estimated value of below 2 is also increasing similarly And these are the results for all of the subjects so that's 15 subjects again using leave one subject out cross validation to make sure we're not overfitting with the model and we get an R M A C. which is about 16 percent to be able to estimate this we go to during exercise and we think that's good enough so that.

[00:37:40]
Potentially you could do this exercise test use the wearable device and what they really want to do with this test is predict which of the patients are really not doing well and needed Vance therapies like transplants or left ventricular assist device was and we think that is pretty much at the level where we could do that we're using these wearable devices with cardiac by ration signals for a couple of other projects so I'll just touch on those briefly now one of them is to try to measure blood pressure without the need for a cut so the need for the reason for that is that these cops actually people don't like using them especially if you had to take cough measurements throughout the day for multiple days ambulatory blood pressure measurements called Nobody wants to wear a cough around nobody wants to use it at night while they're sleeping of it turns on every time it turns on a wake up so there's definitely a need that has been identified by and I H. and other sorts of agencies around the world for couples blood pressure measurement at the same time these cops that you see out there the automated ones used at home are actually not that accurate but most people don't know but they can be off by enough errors so that you could make wrong decisions based on them so that's another issue so what we've done is we've built a device that's in kind of a watch form factor but really the form factor doesn't matter the ideas that we want to measure the chest wall vibrations at the same time as optically measuring when the pulse arrives at the wrist and when the pulse arrives at the skin of the chest and by looking at those time intervals there's physiological mechanisms by which you can predict changes in blood pressure if you can calibrate so some of our earliest results with this have actually been pretty strong where we're getting pretty low errors about 3 to 4 millimeters of mercury for estimated mean arterial pressure the systolic blood pressure is higher than that.

[00:39:39]
And there's still a lot of work to be done on this there's still a lot of major challenges around calibration around how the calibration curve holds up in different conditions but we think that there's some good opportunity here for some successes particularly if the goal is for monitoring changes in blood pressure let's say associated with an activity so we're thinking of folding this back into the heart failure project at some point and saying that beyond what we measure with C.G. alone what if we measure this pulse transit time based blood pressure estimate and how that changes with exercise for the same patients does that add value so we're doing some work on that.

[00:40:17]
We also do some work at the intersection between mental stress and cardiovascular measurements and there's some great people at Emory that we collaborate with that have done. World leading research in this area and it's a relationship that most people don't think about but actually maybe nowadays it's become more mainstream so people do but as you get stressed mentally your heart sees an impact there and your cardiovascular system sees an impact stress can cause your vessels to constrict cause your heart rate to go up and cause your blood pressure to increase and more of that occurring on a daily basis throughout the day can have impact on our cardiovascular system and the other way round of somebody has a heart attack and then afterwards are mentally stressed that can be an issue as well so what we've been measuring is the signals called functional near infrared spectroscopy measurements from the forehead which is typically how non-invasively people have examined the ability to assess stress levels but we're interested in combining those with wearable measurements of these cardiac mechanics signals and to see if that can improve for example the classification of whether somebody is stressed or not through some tasks such as mental arithmetic or and back memories asks and even whether we can separate between the tasks that they're performing.

[00:41:39]
So we had some pretty pretty solid results so far showing that the fusion of cardiovascular and mirrors based wearable sensing signals do lead to better accuracy better precision and recall for determining which type of stressor the person is experiencing mental stress or in this case and we've been using some of these cardiovascular features in another study going on at Emory that I didn't get a chance even to get into here but involving patients with P.T.S.D. and looking for new therapies using electrical stimulation peripheral nerve stimulation as maybe a means to do that so I've got obviously a fantastic group of Ph D. students that are doing really hard work in this area were funded by a lot of agencies that have been generous and support our group.

[00:42:34]
And most importantly maybe not most importantly But I think one of the most important takeaways is all the collaborators that we work with that allow this kind of work to happen I think that that's maybe an important takeaway of for you also if you're interested in getting into research that's really addressing these kinds of big problems in health in particular health and engineering it takes large collaboration networks and people working together to do that and so we're excited to be part of that kind of ecosystem so I don't know how I'm doing on time but that's all I have on slides and so we can out of questions.

[00:43:13]
Yeah. Yeah yeah. That's a great question so so the same sensor actually that we use for C.G. If you look at a broader band with can also give you heart sounds so to measure heart sounds you actually don't use a microphone typically use a stethoscope right the drum head picks up mechanical vibrations of the chest wall and transducers them into acoustics that you can hear so an accelerometer on the chest is one of the best ways to measure heart sounds but we haven't gotten much into that area we're interested in that point but we haven't done too much in terms of murmur classification right that there are a lot of other groups working on that but I do think the combination of maybe sounds with some of these vibration signals could be really useful Yeah good.

[00:44:28]
Up. One of the best things we could do if this were possible is to prospect of lead determine if somebody is at higher risk of injury because they've had too much wear and tear on a joint I think that would be one of the most exciting applications and it might be possible with the sounds but we just need to be able to collect data from huge populations right because in that case in our case we can just find someone who's already been injured and ask if they want to participate get their data and then compare that to the controls in this case you'd have to measure from a huge population wait to see if anybody gets injured and then go back and compare their signals to other people signals and then they might get injured for other reasons besides overuse as well but in terms of impact that would be huge I mean that would we want to of course are really excited about that it's just it's harder to do an actual study.

[00:45:31]
If these things are broadly use then maybe we'll accidentally find yes. Yeah so so what we do is there's a accelerometer that you place against the chest that measures the vibrations of the chest wall and those vibrations are approximately vent so as soon as the heart is moving to Geck blood and as soon as they order valve opens you see big vibrations mechanically at the same time if you have optical sensing at the wrist something called photo op with his MO gram which is just what used for S.P.O. to what's the use of any smart watch to measure the pulse then you can measure the pulse arrival at the wrist but you do both from the same location great question yeah.

[00:46:44]
I don't know I mean there's a lot of people from the industry side of things that are interested in that right because then you could just the person does not do any maneuver I have my doubts about whether that will be possible because of the fact that things like the stiffness of the arm right so there's a lot of other confounders and it's an already noisy situation to try to use these signals that arrive blood pressure so it just adds another variable to a very challenging problem but maybe who knows right maybe there's some active sensing solution that might do it and that's it's always so so you'll see a lot of engineering principles applied here right.

[00:47:25]
Like when we talk about the 6 minute walk test before and after that's really saying we don't want to use a large signal model we want to use a small single model. We want to know what the gain is of that system so we're create a small perturbation and look at how the system responds so there's a lot of these basic engineering principles that are being applied at the systems level here active sensing is another one you know if any time you can put in energy and look at what results from that that's a lot stronger than just listening so maybe if there's some kind of mechanical vibration induced and you look at the response that but I don't know I'm just totally brainstorming I'm not sure yet.

[00:48:23]
It's a great question we have a another collaboration ongoing with with Dr Oz's lab where we want to measure the breathing sounds as well in that setting and see see where things go with that it's really early still but yeah there's a lot of so another common theme you'll notice in his work is the use of sounds and vibrations as the sensing.

[00:48:45]
Modalities because I think they're not used often enough in wearable sensing and I think they carry a lot of information because of the way that our body creates sounds our body creates vibrations they can propagate actually through the skin through the tissues of the body sometimes as as well as or better than electrophysiological signals and they kind of represent the outputs of the systems in some ways right electrophysiology is sort of the signaling that tells For example the heart what to do that tells the muscles to contract.

[00:49:18]
But sounds admitted during movement or vibrations emitted when blood is objective from the heart those are kind of in some ways outputs so I feel like there's a lot more information there that can be at least fused with electrophysiology. Well thank you thanks guys thanks a lot.