Jon Peterson: Thank you for attending, everyone, and I appreciate that Dr. Burkhoff’s talk went before mine because, of course, he set up a lot of what I will be talking about, as well. Dan did not mention that he was actually in Dr. Sagawa’s lab, as was Joe Alexander, another person in our lab here, and Kenji Sunagawa, who you’ll hear later on this afternoon, as well. And I was there, too.

So here we have Harvi at the top and Otto at the bottom, and as I alluded to yesterday, the topologies are very, very similar. They’re largely identical at this point. This is the baseline that we’re working from here. And while Dan has had about 30 years to develop his, our Otto here is a relative newborn.

And so what I’m going to show you is some of the experiments that we’ve done to start to bring Otto up to snuff, as well as some ways that I think are differentiating us from Dan’s approach. Dan’s has been phenomenal in terms of a teaching tool and that sort of thing. But in our case, I think for today, we’re going to first look at one of the things that does differentiate us between how Dr. Burkhoff is approaching parameter estimation and how we hope to approach it. Then I’d like to talk a little bit about how we are extending the basic topology of the model. And finally, yesterday I mentioned that one of the things we would like to do is some closed-loop therapeutics. So I’ll touch a little bit on how we’re doing with that.

So, a reminder of our model here.  Recall that what we’re doing with our model is, we have a parameter space that recognizes all of the ranges of parameters that can go into the model. We have a scenario space that might recognize what’s happening to the patient, and through simulation, we arrive at a measurement space.

But as we talked about yesterday, that’s not where we start, right? We start with measurements from the patient, and possibly a scenario related to the patient, and we want to back out what the values should be in order to get our digital twin.

Well, I want to show you a little bit of preliminary work that is extending towards being able to get that accomplished. And so here’s Otto again, but Otto is a little complicated for our needs for this, and so we looked to have a proof of concept to be able to run through the Bayesian Inference Engine that I referred to yesterday.

And so, rather than having four different chambers and valves and vessels and everything all over the place, we simplified it down to a very simple single chamber with valves going in and out of it, and a much simpler circulatory system. So that only is going to give us about 10 parameters to change, instead of more like 30 in the Otto scenario.

So again, it’s the same kind of thing. We start with our model, we go down in here, we run the simulation, and we arrive at a higher-dimensional measurement space, where we have all sorts of interesting readouts that we can compare against physiology.

So the proof of concept that we did is, we started with a public domain database, which is our measurement space. This is PhysioNet, for those who know. And our scenario space at this point is that we have a patient. We then applied the Embody Bayesian Inference Engine and got our bio digital twin.

And so I’m going to walk through one particular waveform through the PhysioNet system. So here is a pressure waveform, and you could notice that over time, it varies a little bit from beat to beat. And so we represent that a little bit better here by lining up all the beats, all the contractions for this particular waveform. Pay no attention to the blue. I’ll get to that in a minute.

So, we can take this variance… You can see, at each given time in the measurement, there’s a range of pressure values that the patient managed to obtain at that. You can take that variance, run it through the Embody Inference Engine, and you can arrive at a distribution of appropriate parameters for all of the parameters that go into this model.

Now, once you have that distribution of parameters, you can then run across that parameter space, and arrive at, by applying all those different sets of parameters, a range of pressure values. So this allows us to, for instance, it gives you a range of notions of what might happen. And so now we load on top of that, other scenarios, and you can imagine in what Dan said, certain scenarios come out with very different results, depending on who you apply it to and the circumstances.

Well, it’s very possible that this technique will allow us, not through machine learning, but actually just through inference in the analytics that we have here, to very possibly predict what a certain clinical scenario might hold. So it will be interesting to compare the different techniques and see what happens with that.

But at this point, this was a very simple one. And those of you who are observant, which is probably all of you, will note that the blue does not necessarily correspond well to the orange. We knew that going into it, and in point of fact, the Inference Engine was only applied to the latter part of the waveform in each contraction, knowing full well that the former part was not going to fit.

But what we can do is, now that we have arrived at a set of parameters that are appropriate for this simple model, we can take that real pressure measurement, which is here, we can apply it as an input to that part of the model, and we can then predict what the measurement is going to be elsewhere. And in the simple case here, we’re just predicting the flow that would go through here.

But Dan, again, mentioned in the topology that is available to us here, we can look at and predict measurements that are not clinically available, which may well be handy when looking at device and potentially even drug scenarios there. So this represents the first proof of concept of using Bayesian inference and then Bayesian analytics to predict what might be going on. So obviously, our next step is going to be to take our much more complicated Otto model and do that. And we’re working towards that right now, actually. So in the months ahead, hopefully we’ll have something to show for that.

Well, that’s a first step on parameter estimation. Now I want to talk a little bit about how we’ve started to expand on Otto, and what I’ll be able to do here is show some of the steps that Dr. Burkhoff scooted over, because his model right now is quite complex. What we’re doing is we’re taking slower steps and looking at each one. So hopefully this will also serve as an illustration of how these types of frameworks can be used to predict things.

So before I move into any of the scenarios that we’re going to look at here, I thought I would just show which parameters are we actually changing in this. And there are not that many, at this point. We’re changing heart rate, but heart rate actually affects all four of the chambers. It turns out that, at least in normal physiology, they’re beating at the same rate, right? And so that one change will affect all four of those time-varying capacitors.

Well, we also can change stressed blood volume, which is the amount of fluid or charge that is distributed across this entire system here. And that will actually control pressure, which could cause other things to happen, too. We also look at systemic arterial resistance, and finally, in the left ventricle, we can look at left ventricular end systolic elastance, which, again, Dr. Burkhoff talked about.  And essentially, that’s how hard the heart is going to pump. So we’re going to look at how fast it pumps, how hard it pumps, how much fluid it’s pumping, and how hard it is to pump through, the resistance that it is to pump through the systemic system.

So we’ll start out, again, with just a couple of scenarios, and these are very simple scenarios that are more intended to teach us as we walk through things. I do not represent that these are physiological yet.

So we’ll say that we have a healthy person and he’s going to exercise. So in our model, which is designed for teaching us, not for teaching anyone else, so we have, clearly, a very different interface than Dan does. You can see in the lower right, we’ve enabled our autonomic nervous system, and we’ve also enabled age-based calculations, which is largely just deciding on what the intrinsic heart rate is going to be for the patient. This particular patient is 30 years old.

So we’ve run the simulation to the point where all the boring startup stuff that I did yesterday is now over, and now we’re ready to exercise the patient. So, heart rate is around 80, blood pressure is 120/60, and cardiac output is around 4.

So if we start it, you can see that heart rate, unlike my one yesterday, the heart rate is actually increasing at this point, as is blood pressure, systolic over diastolic, and cardiac output. So our simulated person is actually exercising here. You can also see changes in the pressure volume loop. On an individual beat basis, the heart is actually pushing out less volume, but by increasing heart rate, we actually get an increase in cardiac output. And you can see that as we sort of expand out to a bigger thing. We only let our simulated guy exercise for a minute or so. But you can see the heart rate going up, you can see cardiac output going up, and as a result of the shifting in the volume, the pressure volume loop down a bit, ejection fraction actually goes down a little bit and then starts to come back up, again, in our model, at this point. So it’s a nice illustration of how changing just one or two things can start to affect quite a few waveforms.

So our next one here is we’re going to have what we’re going to call a hypertensive emergency. And this is a 50-year-old man. One of my group people, Yuki Sensei, when she came up with these scenarios, for some reason, she decided to beat on men, so these are all men this time. But so he presents to the emergency department complaining of a severe headache, and on examination, his blood pressure was found to be high. Okay, so the initial therapy is to reduce blood pressure, using sodium nitroprusside, but in this case, what we want to try to do is not reduce it too quick, because if you do that, then you affect perfusion of vital organs, like the brain and the heart.

So here’s our scenario. We have our patient coming into the ED, and we’re going to effect treatment. So you can see that by putting on our sodium nitroprusside, which in this case is only acting on systemic resistance, okay? So again, for those of you with electrical analogs, if you don’t change anything else, when you decrease the resistance, you’re going to increase flow and you’re going to drop pressure, and that’s exactly what we see. Now, as you look down a little bit, you can see cardiac output is going up, and so all the things you would like to have happen in the context of a hypertensive emergency has happened here. Now, we decided to look at this in another way.

I mentioned earlier, and I’ve intermixed this in the talk, we’re also looking at the notion of closed-loop therapy. And so I’ll just take a second to explain what we’ve done here. This is, again, all in simulation, it’s all in Simulink. But one of our other team members, Yasu-san, had put together a PI, a proportional integrator controller, just to study the notion of, can we apply some of the things that a clinician does manually in this kind of a system as a proof of concept, to see how well it works? And so here we are, once again, and this time the PI controller acts, and the dotted line on the top waveform there shows the set point of the controller for a mean arterial pressure, and you can see, it pretty much hits it.

The other nice thing about this is, admittedly, partly because we’ve detuned the controller to make sure that it works, but the input of the drug was actually quite nice and slow here. So it had a really nice sort of clinical effect, from our point of view.

So our next one is, we’ll call this one a hemorrhagic shock. And this is a healthy young male, presents to the emergency department with a profusely bleeding wound. And so obviously what we want to do is we want to stabilize, we want to stop the bleeding, and then we want to add volume to compensate. So the first thing we’ll do is we will add volume, but then we’ll also consider adding catecholamines, since the patient in our case is… We’re going to call him nonresponsive to the fluid resuscitation.

So again, first I want to show you what happens if we don’t do anything. So the patient starts to bleed. At some point, their arterial pressure gets low enough that they can’t perfuse organs anymore, okay? So let’s see if we can do a little bit better for our patient here. We’re bleeding, we’re adding some fluid, and you can see that the pressure in the top waveform, the decline has slowed down, but it’s still going down. And now we’ve added on a catecholamine, Norepinephrine, and we’re showing here, actually, all of what we’re doing, and so this drug within the model is, we’re increasing systemic resistance, we’re increasing how hard the heart is pumping, and we’re also increasing heart rate. So we’re trying to get the pressure back up on this patient. And you can see that that’s working pretty well. At some point, we’ve now stabilized the bleeding, the pressure has come up enough, and we cease administration of the drug.

So if we try that on the PI controller side, again, we’re bleeding, and now we do both of them at once, and the cool thing is that you can see that rather than doing this in a stepwise fashion, the PI controller takes both of the things that it has to work with, which is fluid and the drug, and it applies them simultaneously.

So again, is it exactly the way you would do it in a clinical practice? That’s not necessarily the point just yet. The point right now is, we’ve demonstrated, at least within this simulation, that we’re able to have closed-loop therapy delivered to a patient.

Now, finally, our sort of most challenging one, (chuckles) which is, the patient has got a low blood pressure and a very low ejection fraction. Basically, his problem is that his contractility is really, really low. And so our task then is to increase the contractility. And so we’ve chosen dobutamine to do that. So, we’re going to, actually, what do we… Do we add fluid here? We’ll have to see.

Let’s look to see what we do. So once again, we’ve got a low EF, a low blood pressure. We add dobutamine here, and so that’s increasing the pumping ability of the heart. And then, ah, yeah, we added a short-acting beta-blocker here. And so what that’s going to do is, it’s actually going to pull the heart rate down. So we try to decrease the metabolic needs of the heart. We try to make it so that it is appropriately perfused, and so the heart can survive this.

So with our PI controller one, at this point, we haven’t implemented a beta-blocker on it yet, but at least on the administration of dobutamine, it is successful in increasing the pressure some here.

Now, as we have gone through some of these, behind the scenes, there are several things that we’re trying to do that were not necessarily evident here. And this really comes into play when you’re designing a controller for it. If you’re, let’s say, doing an intravenous injection of something, there’s a dead time. There’s a time where the drug gets in, and it has to circulate through until it reaches the site of action, whether that’s the heart, or the vasculature, or whatever. And so, on one of the drugs here, not on all of them, because we’re still getting it implemented, but we have actually implemented that dead time. And then once the drug reaches the site of action, there are kinetics that have to happen. We’re modeling them as first-order kinetics right now, but you have some time constant that’s associated with the drug starting to act on the system.

So those are some examples of some of the things that we’re building into the model itself, so that then an external control system will be able to work with that information.

And so, again, right now, this is all very generically set up right now, but as we move forward, we’ll have more appropriate parameters to put in, so that this looks more like a real patient, instead of just being Otto. And as we move forward, we’ll have better control systems in place to be able to change those values within Otto. And of course, there are many other parts of the model that will be affected by drugs.

One of the things, also, that I did not show here, although it’s one of the first things that I did with the model is, Dr. Burkhoff had mentioned a left ventricular assist device. Well, in this kind of a simulation, it’s really easy to throw in an LVAD, and so these truly are the sorts of things where a very complex, real world mechanical device can be easily simulated in this kind of an environment.

So, we’ve covered a little bit about moving forward from the basic Otto into sort of the next step of how does Otto change over time, and with different drugs on board? And then finally, we’ve at least done a little bit of a proof of concept of using a controller, and in this case, a PI controller. We have much more sophisticated plans in mind, which at some point we will publish. But for now, this was just starting to be very, very simple and looking at proof of concept.

So, that is it, but I did want to just show you one thing, and this was put in, honestly, so that it would help me just in case I need it, but the thing that I like about this is, on the left, we mention a bunch of drugs that we’ve implemented in the model, and we mention some of the real-world effects that they have. And then on the right, we show how they have been implemented in the model thus far. And so you can see that there’s a lot more writing on the left than the right. So we have a little ways to go in order to take all of these different components, even with this small list, and make it highly robust.

But on top of all of these drug interactions, we actually have the systemic autonomic nervous system interactions, where as pressure changes, we have implemented a baroreceptor loop in there. So the model will try to compensate for changes in pressure, by changing, actually, heart rate and left ventricular elastance, so pumping ability, and systemic resistance, at this point.

And that list will also increase over time. So that’s where we’re at. Hopefully by next year, I’ll be able to show you some advances beyond there. So thank you very much.

Moderator: I think we have a couple minutes for questions. It looks like we have one already.

Attendee: Am I the only one here asking questions? (people laughing) Jon, great talk. So I look at this through the lens of someone who’s spent some time thinking about S&T for combat casualty care and for taking care of casualties where people are injured in austere environments. Where in the military, we have not physicians in ICUs, but we have physician assistants and advanced combat medics in an austere environment. So managing these kinds of casualties, also during the aeromedical evacuation, is a big challenge within the Department of Defense, so I want to take a look at this. And I know that they have some of these types of software packages for training some of our medical personnel, but what’s the timeline for product development for this? Because this is an ongoing quest by DoD to get this type of training tools available, with using this as a training tool for our combat medics. I’m just curious as to what your product vision is, what the horizon is; thanks.

Jon Peterson: Sure, and so I can sort of answer that in two ways, and one of them is, we’re R&D, kid! What do you want for… You know, we’re basic research, what do you want from us? But no, we do actually have plans to look at this, and our hope is that over the next five-ish years, we’ll have a prototype closed-loop therapeutic system in place.

Now, you raise a really good point in terms of battlefield conditions and that sort of thing, but I think that having, of course, a robust system that would work in that kind of an environment, we would go outside here and employ a contract manufacturer of some sort. But those are the kinds of things that we’re thinking about, and that’s a great example of how this might be very useful.

Moderator: Okay, one more from the audience. 

Attendee: Hey, Jon. Jon, really great talk. There were a couple of things in the last two scenarios where you were simulating the consequence of significantly reduced perfusion, both sort of peripherally, say, perhaps, in a multiple-organ dysfunction scenario, if that could come up, or certainly in the hemorrhagic, in the blood loss.

Jon Peterson: Mm-hmm.

Attendee: So you’re replacing the fluids, but you’re not replacing red cells.

Jon Peterson: Yes.

Attendee: So is there a biology existing or planned for Otto, where the consequences of poor perfusion in the heart or the consequences of poor perfusion in the periphery, because you’re replacing with fluid, not red cells, that will give an additional feedback in terms of what Otto and the closed-loop systems will have to deal with?

Jon Peterson: Absolutely. The first thing is, for coronary perfusion, right now, that’s actually part of my brief in the lab, and I’ve been too busy preparing talks and things to have enough time to work on that. But yes, and it is an exquisitely complex scenario, because as soon as the… Which, of course, is why you get people dying! (laughs) As soon as arterial pressure goes down, that decreases the pressure in the coronary vasculature, which decreases flow, which decreases oxygen delivery, which leads to worse performance of the heart. So absolutely, that’s one of the things that’s going on. We actually have, in the model, at least stubs for hematocrit, pO2, and pCO2, and so all of those things we are hoping to be able to implement.

We also have means of infusing different… Well, we don’t yet, but we will, I’m sure, within a week, have different ways of infusing, whether it is just going to go straight into stressed blood volume, or the other thing that I have not talked at all about, is unstressed blood volume. Stress volume represents 20-ish percent, yeah? Yeah, around 20-ish percent of the fluid in the body. And so there is an interchange of fluid between the stressed and the unstressed reservoirs there. So we actually, we didn’t implement it for this, but we actually do have a splanchnic circulatory system in place, where we will be able to start affecting that. And then depending on what we infuse and what drugs we use, that may well affect that, as well. So it’s fascinating, though, how all of these things thus far are still only affecting those few parameters that I’ve mentioned earlier, that we’re actually touching in the model right now.