December 9, 2025 • 54 mins
Can a digital replica of your heart save your life? In CXOTalk episode 902, Michael Krigsman talks with Dr. Joe Alexander, Director of the Medical and Health Informatics Lab at NTT Research, to explore the revolutionary world of Bio-Digital Twins.Discover how researchers are using mathematical modeling to build "computational replicas" of the human cardiovascular system. Dr. Alexander explains how these digital twins can predict heart failure, automate critical care in the ICU through closed-loop intervention systems, and pave the way for a future where personalized medicine is accessible to everyone.We dive deep into the science of treating the heart as an electrical circuit, the ethics of AI in medicine, and the "moonshot" goal of eliminating cardiovascular disease..🔷 Show notes: https://www.cxotalk.com/episode/cardiac-digital-twins-inside-the-research-lab🔷 Newsletter: www.cxotalk.com/subscribe🔷 LinkedIn: www.linkedin.com/company/cxotalk🔷 Twitter: twitter.com/cxotalk💡 KEY TOPICS DISCUSSED:Bio-Digital Twins: How digital replicas of organs allow doctors to test treatments virtually before applying them physically.Closed-Loop Systems: The development of autonomous systems that can regulate heart medication beat-by-beat to minimize oxygen consumption and improve recovery.The Physics of the Heart: How Dr. Alexander uses resistors, capacitors, and diodes to model blood flow and pressure.The Future of MedTech: Moving from acute care (ICU) to managing chronic heart failure with wearable technology.👥 ABOUT THE GUEST:
Dr. Joe Alexander is the Director of the Medical and Health Informatics Lab at NTT Research. With a unique background as a chemical engineer and a burn survivor who later attended Johns Hopkins for his MD/PhD, Dr. Alexander is leading "moonshot" research into bio-digital twins to transform cardiovascular care.📍 CHAPTERS:0:00 - Introduction: What is a Cardiac Digital Twin? 2:20 - The Aviation Analogy: Predictive Maintenance for the Body 4:54 - The "Electrical Analog" Model: Viewing the Heart as a Circuit 6:42 - Autonomous Closed-Loop Intervention: Automating ICU Care 12:04 - How a Digital Twin is Built (Population vs. Patient Data) 17:09 - Animal Trials & Validating the Technology 22:11 - Ethics & Safety: Can We Trust AI with Life-and-Death Decisions? 29:03 - Causal vs. Black Box AI: Why Physiology Matters 44:58 - The "Bionic" Approach: Interdisciplinary Teams 50:32 - Health Equity & The Quintuple Aims of Healthcare#DigitalTwins #Cardiology #MedTech #AIinHealthcare #NTTResearch #FutureOfMedicine #HeartHealth #CXOTalk
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Today we go inside a research lab building cardiac digital
twins, a computational replica of your heart that could
transform personalized medicine and how doctors treat heart
failure. Our guest is Doctor Joe
Alexander, director of the Medical and Health Informatics
(00:21):
lab from NTT Research. I'm Michael Krigsman, welcome to
a fascinating episode #902 of CXO Talk.
The medical and health informatics lab belongs to NTT
Japan and that company put a Bell Labs type innovation
company in Silicon Valley calledNTT Research.
(00:44):
The original 3 labs for entity research are physics lab, a
cryptography group and then my group which is medical and
health informatics. We work on kind of moon shot
ideas. The physics group is working on
quantum computing, cryptography is working on even post quantum
cryptography and my group is working on bio digital twins.
(01:08):
How did you end up where you arenow?
Despite growing up poor in Alabama during the civil rights
movement, I found myself in the chemical engineering program at
Aubrey University. One once at a summer job, I was
on a hard, hard oil explosion, and as a burn patient, that's
what introduced me to medicine. The physicians who took care of
(01:31):
me were from Johns Hopkins, as it turned out.
So I went back to Auburn, said Iwant to go to medical school, I
want to go to Johns Hopkins. And somehow that worked out.
And Johns Hopkins is where I learned about biomedical
engineering and cardiovascular dynamics.
And so I did a combined MDPHD program at Johns Hopkins.
(01:52):
You just came out with, hey, I want to go to Johns Hopkins.
And you know, what was their reaction?
It's like not a common request. So I had been purely in the
chemical engineering program. So I went to the pre
professional advisory committee and gave that very specific
request and they laughed at me and said we'd never send anyone
(02:13):
to Johns Hopkins. But as it turned out, things
worked out for me. Tell us about the work that
you're doing. What is a biological digital
twin? We take our examples mostly from
the aviation, commercial aviation industry where, you
(02:35):
know, they build a mathematical representation of every jet
engine. When that engine is flying in
the air, there's a replica on the ground that's taking note of
all kinds of sensors that are recording information in flight
and mathematically incorporatingthose into the model.
And what that does for commercial aviation is that it
(02:56):
makes it possible to do predictive maintenance of the
engines. You don't have to have a
scheduled maintenance. The computer simulation tells
you when it's time to maintain what.
And so analogously, you know, wethought that if patients are
appropriately instrumented, we had enough data from wearable
technologies or elsewhere, we could predict when a patient
(03:17):
needs some kind of intervention.And that kind of set us on the
road to predictive health maintenance and the building of
a biological bio digital twin. So the inspiration then came
from aerospace. That's correct.
Can you talk about the clinical problems that you're actually
(03:39):
trying to solve? And then we'll get into the
discussion, the complexities around building a biological
digital twin. The core problem is that we want
to improve cardiovascular outcomes because cardiovascular
disease is #1 killer globally, accounting for about 20 million
lives in the year 2022. And the way we want to go about
(04:06):
it is to make use of mathematical modelling and
simulation to tailor a model that's specific to an individual
so that that patient can have the very best outcomes.
Currently we rely on clinical trials that are population level
data and usually there are many populations that don't get
(04:28):
included in the clinical trials and we're not totally sure that
the results apply. When you talk about tailoring
specific results to a specific patient, generally as you
indicated, you do large population studies and then take
a solution like a like a medication and apply it to a
(04:50):
large group, but you're drillingdown.
Can you talk about that? We use models that are causal,
that we understand mechanisms for, that are physiological, so
that we can have insights into how those models behave.
We found that for us, rather than go to sophisticated models
(05:11):
that have to do with computational fluid dynamics and
finite element type analysis, which are computationally
expensive, we develop simple models that allow insights.
And our particular structure of model is called an electrical
analogue model, meaning that theflow of blood is analogous to
(05:31):
the flow of ions. Pressures are analogous to
voltages, valves are analogous to diode, and we put together a
structure representing the entire circulatory system and
try to match the parameters in that model to an individual
patient. So it's not just a model of the
(05:55):
heart, but it's an entire cardiovascular model.
Is that a correct way to put it?That's correct.
It's an entire cardiovascular model where we have more
specific models for the heart and the coronary circulation.
And that cardiovascular model also has the necessary
components to do what we want todo at the moment, which is to
(06:17):
manage acute myocardial infarction and acute heart
failure. And what that means is that that
same simple cardiovascular system has to have lungs, it has
to have kidneys, it has to have neural regulation.
So it's a very broad based modelthat you're creating so that so
it's far more complex than simply creating a simulation of
(06:40):
the heart, for example. That's correct.
Can you drill down into your vision of the digital twin and
of the the way that you're designing the solution?
Give us some insight into what you're doing, how you're doing
it, that the challenges and so forth.
One of the easiest ways to consider the kind of advantages
(07:04):
of this sort of approach is a use case, which is autonomous
closed loop intervention system,which you mentioned in the
introduction. This is a situation where we're
trying to introduce a system that manages acute myocardial
infarction and acute heart failure.
After the patient comes to the emergency room, goes to the
(07:26):
cardiac catheterization lab, andthe culprit lesion is relieved,
then they're brought to the CCU or ICU and they're treated with
certain drugs to stabilize the patient and have the heart rest
and recover from having had that, let's say, heart attack.
Those drugs consist usually of something we call catecholamines
(07:49):
to help the heart contract, nitrates to release or to reduce
blood pressure, diuretics to remove blood volume.
These things are given serially by a physician who monitors
patients response. What we're doing with the
autonomous close loop intervention system of what we
aim to do is to have a system where you just type into the
(08:12):
keyboard the desired left atrialpressure, cardiac output,
systemic arterial pressure, and then a system of syringes filled
with each of the necessary agents automatically delivers
the amount of drug that's necessary to achieve those
desired endpoints. Then we monitor the patient's
response, feed that back to our system, and depending on the
(08:34):
error between what we told the machine, the syringes to give,
and the patient's response, thenwe make an adjustment.
This is a feedback control system for delivering care
that's specified at the keyboardin terms of desired left atrial
pressure, cardiac output and arterial pressure.
And how is this different from existing models, right?
(09:01):
I mean, there are machines that deliver drugs, so how?
What makes this unique? Anesthesiology is a field where
there's some level of automationin the delivery of drugs to
maintain blood pressure in patients who are undergoing
operations. The thing that we're doing that
(09:23):
no one else has done and it's very challenging thing, is to
control not just blood pressure,but cardiac output and left
atrial pressure and to do these in such a way as to minimize
myocardial oxygen consumption. So the whole point is that the
(09:44):
heart needs to relax and recover, yet the periphery needs
to be perfused. And so if we can let the heart
work the least possible while achieving all these things, then
that's the best situation for myocardial recovery.
So minimization of myocardial oxygen consumption is the
(10:07):
special thing that no one has been able to achieve.
And what are you doing that you hope will allow you to achieve
this? We have the benefit of some
consultants and some work that gives us a way to monitor and
quickly measure myocardial oxygen consumption live.
(10:32):
And we also have some physiological construct based on
cardiac output venous return surface that lets us do a good
job of maintaining left atrial pressure at a specific range.
At same time, we regulate right atrial pressure.
So the combination of the cardiac venous return surface
(10:54):
and technologies to instantaneously track or
periodically track monocardial auto consumption are some of the
benefits. But there's some other
techniques that we have as well.Folks, you can ask questions.
If you're watching on LinkedIn, pop your questions into the
LinkedIn chat. If you're watching on Twitter X,
(11:17):
use the hashtag CXO Talk and askDoctor Alexander pretty much
whatever you want. Now is a special opportunity to
to do that. Joe, how does what you were just
describing then relate back to the digital twin?
(11:37):
My lab is interested in the human digital twin.
We started with cardiovascular system because we have expertise
and it's #1 killer global and wethought it's a good place to
start. As well as the fact that the
cardiovascular system even attacking something like keep
myocardial infarction, it makes it necessary to consider other
(11:59):
organ systems like the lungs andthe kidneys and and
neuroregulation, this autonomousclosed loop intervention system.
The first guess is given by our electrical analogue model of the
human cardiovascular system thatbest fits a particular patient
who walks in. Now how do we determine the best
(12:21):
electrical model for the patientthat we see lying in front of
us? That relies on huge amounts of
population level data to start with, where we build a virtual
population of patients that are similar to the patient that
walks in, in terms of you know, age and conditions,
(12:43):
comorbidities, etcetera. We did we build a population of
virtual population similar to that patient.
Then based on a specific characteristic we narrow in on
the best guess and that might beseveral 100 patients that are
close to this patient. And we use that to get the
initial first guess on this closed loop, autonomous closed
(13:05):
loop intervention system. Then because it's a feedback
system, it will automatically adjust those parameters to give
the specified low Tatro pressure, cardiac output,
arterial pressure. Now mind you, that information
in the error signal, the correction signal that's being
fed back, that teaches the modelof that patient what those
(13:29):
parameters should be. So the model becomes more and
more specific for that patient as therapy is administered
through this autonomous close dependent pension system.
So it both zeroes in on the bestparameter choices and at the
same time teaches us what the best parameter choice choices
(13:49):
are for that patient. So we will create a registry
including that patient and thosechoices.
So you are gathering broad population data and then
isolating out of that broad group for each individual
patient the right subset of datathat most closely corresponds to
(14:14):
that particular patient's situation.
And then from there you're building the digital twin of
that specific patient. Is is that a correct way to say
it? That's very correct.
We think of the digital twin as a system of characteristics that
represent a patient, characteristics like resistors,
(14:37):
capacitors, dialogues and and soforth.
So we want the those parameters,the best choices that most
represent that patient and that patient's characteristics.
And so we begin with population level data that's in the
literature. And from there, we build a
virtual population that is relevant for this particular
(14:57):
patient. And from within that virtual
population, we find the closest patients for that patient.
Then we begin this process of driving the autonomous closest
intervention system. In the course of that feedback
regulation and adjustment, then we learn to titrate or tune
(15:17):
those parameters to match that particular patient.
What kinds of data are you usingas inputs into all of this?
In the world of cardiovascular about digital twins, there are
many places we could have started.
We started with acute myocardialin parks and acute heart
(15:38):
failure, mostly because it's acute and mostly because the
patients will be in the ICU or CCU and heavily monitored and
heavily instrumented where we have readings of dynamic signals
like dynamic pressure pressures and volumes and we have
(15:59):
echocardiography. So we have a lot of time
dependent signals that we can use to train our models and
patient and set those parametersand and gather data.
So the inputs in the ICUCCU are,you know, measurements of left
(16:20):
atrial pressure, right atrial pressure, cardiac output.
We can do SWAN catheters, we canmonitor pressure over periods of
time. And usually the admission for
that patient is about four or five days.
And so we can have four or five days of acute information to
(16:40):
tune to that particular patient.So we chose the easiest option
to be honest as a place to begin.
When you go to chronic heart failure, you have to rely on
data that you know from patientsthat are not in the hospital,
and they're static measurements and they're based on wearable
(17:01):
technologies, not nearly invasive enough to parameterize
such a system. But we'll go to chronic heart
failure next. Where are you in the process of
this set of developments? We have a group of
collaborators. The ones most relevant to what
we're talking about today are a group at National Cardiovascular
(17:24):
Center in Japan, where they havecardiologist, anesthesiologist,
experimentalist and are able to do animal experiments with
animals ranging in size from from mice all the way up to
cows. And so we're doing animal
experimentation to validate these approaches and test out
(17:47):
these technologies. And last year, for the first
time ever, we demonstrated the proper functioning of an
autonomous closed loop intervention system to treat
heart failure in an experimentalanimal, a dog that that had
experimentally induced heart failure and the system
(18:08):
compensated and automatically treated that dog while
minimizing myocardial oxygen consumption.
We're at a stage now where we'recollecting a number of
experiments from animals where we'll prove, we intend to prove
that the autonomous close up intervention system outperforms
(18:29):
the standard of care. By showing that these
cardiologists and their treatment compared to the
autonomous system cannot yield the such a small myocardial
infarct size as what we can get from the autonomous close up
intervention system. So reduction in infarct size
(18:49):
would be physical evidence of how this system could outperform
clinicians and that's what we intend to do.
The autonomous system then is relying on this general
population data plus very detailed data coming from the
(19:09):
particular patient who is in theICU or the CCU, because there's
lots of data and you're finding so far are a greater degree of
accuracy and accuracy in what specifically that is producing
the result. We are still in the very early
(19:32):
stages and developing you know what we can do in the real life
clinical situation. So we're all of what I'm saying
is based on hopes that depend onwhat we've been able to validate
then verify in animal experiment.
So we're still at the animal experimentation phase and so we
(19:53):
are not looking at population level data in animals.
We're even beginning without that kind of referenced model.
Of parameters for animals, except what we can get from the
literature very simply and stillthis autonomous system is able
(20:14):
to perform. We demonstrated last year.
So, but your description of the process is correct.
That's what we hope to achieve and there are many challenges as
you can imagine. Again, we're at the early stages
achieving this in animals and before we get anywhere near
(20:35):
regulatory approval, we have to go through and I'm sure you're
aware of these types of hurdles.We have to use this kind of
system to do clinical decision support before we can go to
complete autonomy. And so that means the feedback
would be going through a physician, the physician would
be in this closed loop and we advised the physician and the
(20:57):
physician would make a decision whether to do what the software
says or not. So we have many things to
achieve before we before I can really answer your question
well. I guess similar to autonomous
vehicles, before you let a fleetof autonomous taxis on the road,
(21:18):
you have a driver in the car because you really don't know
what's going to happen until youdie.
It try it because it's new. So it's basically the same
concept. That's exactly right.
So some people have suggested that we not call it autonomous
intervention, but like a copilotsystem.
But we're ambitious in thinking that this kind of a system can
(21:40):
replace the need for a specialist to be at the
patient's bedside. It could be a system that's
monitored by, you know, another clinical professional, but may
not need, probably will not needa specialist, which makes it
accessible to kind of remote hospitals with fewer resources.
(22:04):
And so that's part of our visionfor this, this HealthEquity.
We have an important question from Twitter, from Arsalan Khan,
who's a regular listener. He always asks excellent
questions and he says this. It's this is a very good way
into getting to to dive into theethical and equity aspects of
(22:30):
this. He says this How do you make
sure data from these medical instruments is correct and the
doctor makes the right decision based on the data versus his or
her personal experience? My bias to be honest, and one
clinician made this comment. He said that he'd rather have an
(22:53):
autonomous system that focusing beat to beat on adjusting meds
for a patient than to have an exhausted resident standing in
the hallway checking his phone and doesn't have enough time to
look in on the patient. We had the bias that if the
system is built correctly, then it will correct even for errors
(23:17):
that the position might make. But it's a, it's necessary like
you described for levels of autonomy in vehicles to have
that supervision. FDA is not going to approve an
autonomous device no matter, no matter how well it's tested
without that step of clinical decision support.
(23:40):
And so we we have to do that. We can't take the position out
of the loop. And then we have another related
question from Elizabeth Shaw, who says how do you know you can
trust the model and then the decisions in healthcare, the
standard is let AI guide you, but the doctor should make the
decision. Yes, it's related.
(24:01):
Our first pass is to include thedoctor in loop so the doctor can
judge, you know, what the the system is telling him, him or
her. And let me just say that the
thing about our system that's different from AI is that it's
causally based. Even though cardiologists don't
(24:23):
do as much Physiology as they once did, the decisions that
would be coming out of our system will be based on clinical
practice, cardiovascular Physiology, cardiovascular
dynamics, and it won't be a black box deep learning type
(24:45):
output that cannot be reviewed and and criticized.
I'm a technologist and I absolutely support AI and
autonomous systems when they're used properly with the right
type of governance. But if we look for example, at
the Boeing 737 Max crashes that were caused by faulty sensor
(25:10):
algorithms and then incorrectly documented or incompletely
documented recovery procedures, it's an example of an autonomous
life and death system. And, and so the idea of of
having this kind of autonomous system relating to the to
(25:31):
medical medicine is kind of scary.
There's certainly lessons to be be learned, and some of the some
of the lessons are avoiding shortcuts and some of the
lessons have to do with hubris. Certainly fail safe mechanisms
have to be built in and certainly you don't want to put
critical components in the hands, at least I don't of AI.
(25:53):
The short answer is the work we do is a beautiful dance between
creativity and skepticism and we're skeptical always.
And we really don't trust and certainly don't fully rely on
systems like AI, or I would say we depend on AI for particular
and certain algorithms for particular applications or
(26:15):
particular aspects. That's why our approach is
crucially causal and mechanisticand based on Physiology so that
we can see better than what is being called, I guess,
explainable AI. Explainable AI doesn't give the
kinds of explanations that cardiologists understand in
(26:37):
terms of Physiology and mechanisms and we do provide
those. There will be rationales for
what our systems would advise the physician when the
physician's in the loop. Now when the physician done in
the in the loop, then we can't say that the system will be
(26:57):
perfect all at once, but we'll suddenly be designing with with
safety in mind and risk. I should also just just mention
that some of the, the, the deeper questions that have to do
with testing implementation, they will involve large scale
(27:22):
clinical trials just like drugs.So the same questions that are
being raised about algorithms and how do you know if they're
safe or not, you know, can be raised about drug development.
And we accept the results of clinical trials even when we see
these commercials saying that there's a risk of death with,
(27:47):
you know, some of the with some of the the medications under
certain conditions and that wereproven in the trial.
And so there is no perfect solution without some level of
risk. And the goal would be to
minimize those risks. Now as entity research is Bell
Labs type moonshine ideas place,we're looking to develop proof
(28:09):
of concept to develop a regulatory strategy and to hand
off to those those companies to industry that knows how to go
into that risk safety environment like pharmaceutical
companies, like medical device companies.
And we haven't touched on medical devices, but everything
I'm describing about drug treatment also applies to the
(28:33):
management of devices controlling patients who are
more severely I'll. So we will be looking to partner
with pharmaceutical companies, medical device companies that
are able to afford that have expertise and ethical aspects,
safety aspects and conducting those, affording those large
scale clinical trials. Then we'll move on to the next
(28:55):
place so we can paint beautiful pictures and brilliant
landscapes on our small prison walls.
You mentioned several times thatyour work is causal and
mechanistic. Can you dive into that and
explain what you mean with in comparison to things that that
(29:16):
are not causal or mechanistic? If we think about a patient who
has heart failure, heart failureusually in patients end up with
some degree of renal failure. The renal failure causes an
increase in volume retention. The volume retention means that
(29:37):
you get fluid in the circulatorysystem that flows back to the
lungs and you get fluid in the lungs, fluid in the lungs.
Elevation of left atrial pressure means that you have
poor oxygen diffusion capacity, so you don't get as much oxygen
across the lungs into the circulatory system because you
don't have as much oxygen. The heart has to pump harder and
(29:59):
faster because in order to deliver enough oxygen to confuse
the tissues, in order to do that, the heart has to work
harder. This patient with heart failure,
that patient's heart has to workharder and consume more
myocardial oxygen. It needs more oxygen, but it
can't get more oxygen. So the heart is feeding itself,
(30:20):
but it can't get enough oxygen to feed itself, and you quickly
deteriorate. That is a sequence of events set
in place by various causes that originate with heart failure.
That's different from, for example, having some speckled
(30:40):
tracking images that are fed into AI and making some
hypotheses from that alone aboutwhy a patient is not ventilating
well. That for me is pattern
recognition with an attempt to extrapolate to causality.
(31:01):
But this sequence of events I just described, which also by
the way includes anemia and which exacerbates the situation
more, these are all things knownphysiologically and
mechanistically to play a significant roles in heart
(31:21):
failure. And all of the drugs that have
been developed are mostly addressing these various
aspects. Diuretics to relieve, remove
volume, contract catecholamines to improve contractile function.
But at some point, the cost in myocardial oxygen consumption is
(31:42):
just really the last draw for a patient with chronic heart
failure. It's very, very serious.
And the intervention that your system is doing, correct me if
I'm wrong, is injecting certain types of medications and
precisely the right dosages and the right times.
(32:03):
Is that correct or is there moreto it?
That's correct. For our particular use case of
acute myocardial infarctum, acute heart failure where we
have control over everything andwe're measuring everything.
It does those injections to manage the acute case.
But the best we can expect for that autonomous close up
(32:25):
intervention system is the acutecase.
And so four or five day hospitalstay it doesn't.
It gives us some knowledge that's helpful in the chronic
case when the patient's not in the hospital.
We have some starting places, but we need much more data, we
need much more laboratory chemistry.
(32:45):
We need to be monitoring patients with wearable
technologies and so forth over longer periods of time.
However, we expect that the excellent care in the acute
situation will have benefits forthe patient when they're
released from the hospital. They won't be readmitted.
(33:07):
For example, the readmission rate we expect would be less or
would be over a longer period oftime compared to the standard of
care. You described earlier the
analogy of an electrical system with diodes and resistors, and
(33:30):
as I was reading through some ofyour papers and presentations, I
was struck by these electrical diagrams.
I mean literally like schematic diagrams that you'd see for
electrical devices. Can you tell us more about that
that aspect? When we naturally, when we look
(33:51):
at that diagram, we see what looks like the electrical
circuit and the circulatory system is a circuit.
And so that part, it kind of works for us naturally.
And actually, elements in that electrical analogue diagram
represent left ventricular contraction, right atrial, right
ventricular contraction, atrial contractions.
(34:13):
The person who is familiar with those symbols can see the entire
circulatory system represented there, but they're represented
very simply. For example, just to give you a
notion, the left ventricle ejects into the systemic
arterial system across the aortic valve.
(34:33):
So for us the left ventricle is a capacitor where for a certain
volume, you know it generates a certain force and we call it a
time varying capacitor because we let the capacitor level
change or electric ventricle become stiffer to generate
pressure. That pressure in the electrical
allogue sense is a voltage, and that voltage in the heart in
(34:56):
that left ventricle gets higher than what's downstream across
the valve. Then that valve will open or
that diode will open and flow will go out of the diode or
across the aortic valve into theanother set of resistors and
capacitors which we say represent the systemic arterial
system. So resistance in the electrical
(35:20):
analogue network is the relationship between flow and
pressure in the real system capacitor.
A capacitor in that electrical analogue network is
representation between volume that goes into an elastic
artery, it stretches the wall ofthe artery and that generates a
(35:43):
certain pressure. So by a series or a system of
resistors and capacitors and diodes representing different
compartments around the arterialsystemic arterial system, the
venous systemic venous system, the pulmonary arterial system,
the pulmonary venous system, we can represent the entire, simply
(36:07):
represent and I know simple, simple is a relative term, but
simply represent the entire circulatory loop in the
essential measurements or parameters or variables that are
important for our application toacute myocardial infarction and
acute heart failure. And how does this type of
representation interplay with the creation of the digital
(36:33):
twin? This representation is our
digital twin. It's a mathematical
representation of the patient inthese terms.
Now what I described to you was the simple circulatory loop.
We have other subsystems that are more specialized that are
not represented in probably the diagrams you looked at that
(36:56):
represent the the coronary circulation.
Now, if you think about the heart, the heart generates a
pressure and it generates a flow.
That flow goes into the systemicarteries across the aorta.
But that same aorta is has a coronary circulation that
supplies blood to the heart itself.
(37:18):
That's why I say the heart feedsitself in supplying its own
blood supply. That coronary, that coronary
circulation is has a little moresophistication that we need to
account for. If there's auto regulation of
coronary flow, we need to account for that.
When the heart beats, it squeezes the vessel surrounding
(37:38):
it, so it squeezes its own coronary vessels.
And so we need to account for that in describing what flow
gets to an ischemic heart. So we have to do a more
sophisticated model drilling in on the heart and this
metabolism. Now it turns out that myocardial
oxygen consumption, cardiac metabolism has been thoroughly
(38:02):
described in simple terms that relate to pressure and volume
mechanics. Again, those signals that we're
measuring all the time, we're interested in pressures and
volumes you can measure. You can use a pressure volume
loop to represent the mechanics of a Chamber of the heart, where
the slope of the pressure volumerelationship is an index of the
(38:24):
ability of the heart to contract.
So a simple index of the abilityof the heart to contract can be
derived by pressure volume measurements as well.
Those same simple signals and the areas in various aspects for
the pressure volume can show potential energy of the heart,
the work that the heart does, and they correlate with
(38:45):
myocardial oxygen consumption. So there's a tremendous amount
of data about the heart that's implicit in the way that we
study it using pressures, flows and volumes.
And Arsalan Khan on Twitter comes back again and he asks,
with your research, what did youfind or did you find that the
(39:06):
systems and processes, traditional ways of describing
the body were either not useful or overly complicated?
How did you how did you end up with this electrical analog?
In the world of cardiology, although the electrical analog
(39:28):
model is not so thoroughly represented as what we're using,
it has become a traditional approach to characterizing
hemodynamics. And so pressure volume area in
systolic, pressure volume relationship, the representation
(39:51):
of the dynamics of the circulatory system and neural
control of the circulation usingbaroreceptors and so forth.
This way electrical analogue wayis traditional, It's just that
many practitioners don't understand it on this level.
(40:13):
They're using volume conductancecatheters in the Cath lab to
measure volume and pressure and generate pressure volume loops
and generating in indices of or indexes of contractility.
So it's not non traditional, it's just that we're taking it
on a whole nother level and extending it and using some non
(40:33):
traditional or, or some elementsthat we discover and that have
been published but have not beentraditionally used because of,
I'll just say, complexity of biomedical engineering compared
to clinical medicine. So a lot of clinical
(40:54):
cardiologists are using the principles but don't understand
it so thoroughly. Now the I mentioned, I may have
mentioned a consultant that we have on the project, his name is
Kenji Sonogawa. He discovered and prove many of
the principles that are being used in clinical cardiology
(41:15):
today that are based on pressurevolume characteristics and
electrical analogue models. There's an electrical analogue
model of the systemic arterial system called impedance that not
only represent is not only represented representing the
resistances, but the frequency dependent nature of pressure and
flow in the dynamics of the arterial system.
(41:36):
People are using results from impedance characteristics
without being able to measure impedance, which is another
electrical term because it's just a little bit more
complicated than what people areare ready for.
And so all of these insights, for example, another example is
ventricular vascular coupling. Cardiologists talk about
(41:58):
ventricular vascular coupling and one concept is to think of
the heart that's a as a, if you will, a battery with a charge
and the arterial system represented by pressure
installed. Pressure and stroke volume
relationship where you have two capacitors and opening the
(42:20):
aortic valve is the closing of aswitch between the capacitors
and volume will move from one capacitor to the other.
That notion of ventricular vascular coupling is something
also derived by Doctor Kenji Senegal, who's a consultant.
And I'll just add that a lot of the work I did together with
(42:40):
Kenji Sunugawa when I was a graduate student at Johns
Hopkins. And so we have a long history of
work using electrical analog models dating back to the 70s.
How much of the math did you need to develop versus relying
(43:02):
on traditional or existing understanding of these various
relationships? As a biomedical engineer, we've
used, you know, whatever popularmathematics was available at the
time that might render some new piece of information about
biology to be understandable or to make sense.
(43:23):
So I, I started way back in the 70s doing neural networks, you
know, when it was a fashionable thing.
We did wavelet analysis of waveforms and reflection
coefficients reflected waves. So we as biomedical engineers,
we love complexity. And you know, we'll use
(43:45):
homeomorphic deconvolution mathematics of, you know,
measuring earthquakes, anything that's available, we'll learn
and use how much, as much as necessary and practical.
So we built back in the day we built, we had RadioShack
computers controlling analog computers that we patched in in
(44:08):
order to have real time control of an isolated perfused dog
heart preparation. That was just the way we do
things. So I don't know how much
mathematics, but as much as necessary to be practical.
There's a image near Ian Hunter at MIT who once said to me, he
(44:31):
said scientists like taking things apart and you engineers
like putting things together. So I in retrospect, everything
I've done, I'm more an engineer than a scientist, has been
integrative. How to explain the system when
it's already put together. And sometimes that takes a lot
(44:54):
of mathematics, and we'll do what's necessary about them.
Sounds like you're building the bionic system.
Remember there used to be ATV show called the $6 million man
that was sounds like you're you're building the bionic
approach. What about the the team required
(45:16):
to build this? There are so many different
skills that are involved. One comment to your comment
about the bionic man by like woman, woman is that Kenji and I
once wrote a paper around the bionic baroreceptor reflex with
exactly that notion in mind, where you have a device that
(45:38):
automatically can regulate arterial pressure reproducing a
failed autonomic system. And there's a syndrome called
Scheidrager syndrome where the receptor doesn't doesn't work
and that kind of system has beenproven to work.
So we back in the 70s or 80s, wewere developing A bionic
(46:00):
baroreceptor reflex. Yes, on our team in the May lab,
immediately we have a number of PhDs in biomedical engineers, in
biomedical engineering, we have MD's, we have masters people and
everybody can do programming. Everybody knows Physiology
(46:24):
either from the beginning or they've learned from our
consultants, and we have routineconferences with cardiologists
and anesthesiologists doing experiments together with them.
So we have that level of expertise.
And at the National Cardiovascular Center, I
mentioned earlier, those guys are seeing patients on the one
(46:47):
hand, then doing experiments in animals, then writing programs
and then building electrical analog circuits.
The same person is doing all of those things.
And so we have access to all of their training and their
expertise. And this is an occasion where I
should also mention that I've talked about our main
collaborator and our main projects on the larger scale.
(47:12):
We also have collaborations looking at other aspects of
digital twins with Harvard Disease Biophysics Group, where
we're working on building developing a bio hybrid digital
twin strategy using stem cells to manufacture tissue engineer
(47:34):
hearts from patients own stem cells and then study those
chambers beating and isolation. And we have a goal or they have
a goal and we're trying to facilitate them to build a heart
based on on patients own stem cells and tissues generated in a
bioreactor from those stem cells.
(47:55):
So there were tapping into tissue engineering and lots of
machine learning. We published a paper with them
in Science Robotics on machine learning inspired design of a
ray, where they posited that it may be possible with AI and
(48:17):
machine learning to have biological design evolve at a
more rapid rate than evolution on the basis of what would be
predicted from some first principles that we know about
how tissues function. Not necessarily a trial and
error approach to revolution, but a specific design inspired
(48:37):
by machine learning. Then we have a Group, A
collaboration with Technical University of Munich, where
they're developing different types of biosensors and
different types of technologies for in vivo measurements of
electrical signals. And so we have a lot of
(48:59):
expertise available in our collaboration done, a lot of
expertise local to our small team, which is only about 9
people at the moment. So you really are fusing the
Physiology with medical practice, research with robotics
(49:24):
and the math abstracted to create your models?
I mean, it's so interdisciplinary.
It is. And while we have a strategy,
we're also again we're that BellLabs type group where you get a
bunch of creative people who have lots of new ideas.
So some of this, the work that we're doing is born out of many
(49:47):
great ideas that need to be pursued at the same time that we
try to keep a larger strategy inmind so that everyone is working
around a kind of organizing principle for what we want to
achieve. And one of the main goals is
improved patient outcomes. Some of those goals we expect to
(50:10):
achieve in the longer term, but some we want to demonstrate in
the shorter term. And this autonomous closed loop
intervention system is turning out to be at least in terms of
first in human studies. We expect to see some first in
human studies within five years.You brought up earlier the
(50:35):
HealthEquity vision that you have and how does this research
relate to that and very quickly,because again, we're going to
run out of time. If you think of the quintuple
aims of healthcare and that's improved experience of the
patient, improved experience forthe clinician, improved
population health, decreased cost, improved HealthEquity,
(50:59):
then I think for example, the autonomous flows of an invention
system does all of that. So improves HealthEquity by not
requiring so many specialists inorder to treat a very deadly
cardiovascular disease. The cost, we hope, should be
(51:19):
less. Although the history of medicine
would suggest that no matter howdramatic an improvement we make
in scientific advancement, the entrepreneurs find a way to make
it more costly. But we think that this
technology will be less, less costly and should improve the
(51:41):
physician experience. I won't be so exhausted having
to give so much care to managingparticular patients and the
benefits just strike directly, Ithink with the quintuple aims of
improved healthcare. Where is all this going?
Where is your research headed right now?
In the immediate term, we're headed toward those first in
(52:04):
human studies that have to do with clinical decision support
and we're looking to develop regulatory strategy with a
partner and hand off the expertise, hand off to them and
their expertise and investments in large scale clinical trials
(52:27):
and developing technology devices or pharmaceuticals.
And then we would move to focusing on chronic heart
failure and that's where we would expect to make use more of
the collaborations we have with Harvard Disease Biophysics Group
(52:48):
and with our folks at TUM wearable devices and such.
And again, this is all starting from the notion of a human bio
digital twin or human bio digital twins, not just
cardiovascular. It was a starting place and I'm
personally interested in neurodegenerative diseases and
(53:09):
so if I'm around long enough, I hope we can get to to that
branch of work as well. All right.
And with that, I'm afraid we're out of time.
A huge thank you to Doctor Joe Alexander.
He is director of the medical and health Informatics lab at
(53:31):
NTT Research. Joe, thank you so much for
taking time to be with us. It's been a very fascinating
glimpse into your work. Thank you so much Michael for
having me. Everybody, thank you for
watching. Before you go, subscribe to the
CXO Talk newsletter so we can notify you about upcoming shows.
(53:53):
We have great shows that are coming up.
Thanks so much everybody for watching and we'll see you again
next time. Have a great day.
Today we go inside a research lab building cardiac digital
twins, a computational replica of your heart that could
transform personalized medicine and how doctors treat heart
failure. Our guest is Doctor Joe
Alexander, director of the Medical and Health Informatics
(00:21):
lab from NTT Research. I'm Michael Krigsman, welcome to
a fascinating episode #902 of CXO Talk.
The medical and health informatics lab belongs to NTT
Japan and that company put a Bell Labs type innovation
company in Silicon Valley calledNTT Research.
(00:44):
The original 3 labs for entity research are physics lab, a
cryptography group and then my group which is medical and
health informatics. We work on kind of moon shot
ideas. The physics group is working on
quantum computing, cryptography is working on even post quantum
cryptography and my group is working on bio digital twins.
(01:08):
How did you end up where you arenow?
Despite growing up poor in Alabama during the civil rights
movement, I found myself in the chemical engineering program at
Aubrey University. One once at a summer job, I was
on a hard, hard oil explosion, and as a burn patient, that's
what introduced me to medicine. The physicians who took care of
(01:31):
me were from Johns Hopkins, as it turned out.
So I went back to Auburn, said Iwant to go to medical school, I
want to go to Johns Hopkins. And somehow that worked out.
And Johns Hopkins is where I learned about biomedical
engineering and cardiovascular dynamics.
And so I did a combined MDPHD program at Johns Hopkins.
(01:52):
You just came out with, hey, I want to go to Johns Hopkins.
And you know, what was their reaction?
It's like not a common request. So I had been purely in the
chemical engineering program. So I went to the pre
professional advisory committee and gave that very specific
request and they laughed at me and said we'd never send anyone
(02:13):
to Johns Hopkins. But as it turned out, things
worked out for me. Tell us about the work that
you're doing. What is a biological digital
twin? We take our examples mostly from
the aviation, commercial aviation industry where, you
(02:35):
know, they build a mathematical representation of every jet
engine. When that engine is flying in
the air, there's a replica on the ground that's taking note of
all kinds of sensors that are recording information in flight
and mathematically incorporatingthose into the model.
And what that does for commercial aviation is that it
(02:56):
makes it possible to do predictive maintenance of the
engines. You don't have to have a
scheduled maintenance. The computer simulation tells
you when it's time to maintain what.
And so analogously, you know, wethought that if patients are
appropriately instrumented, we had enough data from wearable
technologies or elsewhere, we could predict when a patient
(03:17):
needs some kind of intervention.And that kind of set us on the
road to predictive health maintenance and the building of
a biological bio digital twin. So the inspiration then came
from aerospace. That's correct.
Can you talk about the clinical problems that you're actually
(03:39):
trying to solve? And then we'll get into the
discussion, the complexities around building a biological
digital twin. The core problem is that we want
to improve cardiovascular outcomes because cardiovascular
disease is #1 killer globally, accounting for about 20 million
lives in the year 2022. And the way we want to go about
(04:06):
it is to make use of mathematical modelling and
simulation to tailor a model that's specific to an individual
so that that patient can have the very best outcomes.
Currently we rely on clinical trials that are population level
data and usually there are many populations that don't get
(04:28):
included in the clinical trials and we're not totally sure that
the results apply. When you talk about tailoring
specific results to a specific patient, generally as you
indicated, you do large population studies and then take
a solution like a like a medication and apply it to a
(04:50):
large group, but you're drillingdown.
Can you talk about that? We use models that are causal,
that we understand mechanisms for, that are physiological, so
that we can have insights into how those models behave.
We found that for us, rather than go to sophisticated models
(05:11):
that have to do with computational fluid dynamics and
finite element type analysis, which are computationally
expensive, we develop simple models that allow insights.
And our particular structure of model is called an electrical
analogue model, meaning that theflow of blood is analogous to
(05:31):
the flow of ions. Pressures are analogous to
voltages, valves are analogous to diode, and we put together a
structure representing the entire circulatory system and
try to match the parameters in that model to an individual
patient. So it's not just a model of the
(05:55):
heart, but it's an entire cardiovascular model.
Is that a correct way to put it?That's correct.
It's an entire cardiovascular model where we have more
specific models for the heart and the coronary circulation.
And that cardiovascular model also has the necessary
components to do what we want todo at the moment, which is to
(06:17):
manage acute myocardial infarction and acute heart
failure. And what that means is that that
same simple cardiovascular system has to have lungs, it has
to have kidneys, it has to have neural regulation.
So it's a very broad based modelthat you're creating so that so
it's far more complex than simply creating a simulation of
(06:40):
the heart, for example. That's correct.
Can you drill down into your vision of the digital twin and
of the the way that you're designing the solution?
Give us some insight into what you're doing, how you're doing
it, that the challenges and so forth.
One of the easiest ways to consider the kind of advantages
(07:04):
of this sort of approach is a use case, which is autonomous
closed loop intervention system,which you mentioned in the
introduction. This is a situation where we're
trying to introduce a system that manages acute myocardial
infarction and acute heart failure.
After the patient comes to the emergency room, goes to the
(07:26):
cardiac catheterization lab, andthe culprit lesion is relieved,
then they're brought to the CCU or ICU and they're treated with
certain drugs to stabilize the patient and have the heart rest
and recover from having had that, let's say, heart attack.
Those drugs consist usually of something we call catecholamines
(07:49):
to help the heart contract, nitrates to release or to reduce
blood pressure, diuretics to remove blood volume.
These things are given serially by a physician who monitors
patients response. What we're doing with the
autonomous close loop intervention system of what we
aim to do is to have a system where you just type into the
(08:12):
keyboard the desired left atrialpressure, cardiac output,
systemic arterial pressure, and then a system of syringes filled
with each of the necessary agents automatically delivers
the amount of drug that's necessary to achieve those
desired endpoints. Then we monitor the patient's
response, feed that back to our system, and depending on the
(08:34):
error between what we told the machine, the syringes to give,
and the patient's response, thenwe make an adjustment.
This is a feedback control system for delivering care
that's specified at the keyboardin terms of desired left atrial
pressure, cardiac output and arterial pressure.
And how is this different from existing models, right?
(09:01):
I mean, there are machines that deliver drugs, so how?
What makes this unique? Anesthesiology is a field where
there's some level of automationin the delivery of drugs to
maintain blood pressure in patients who are undergoing
operations. The thing that we're doing that
(09:23):
no one else has done and it's very challenging thing, is to
control not just blood pressure,but cardiac output and left
atrial pressure and to do these in such a way as to minimize
myocardial oxygen consumption. So the whole point is that the
(09:44):
heart needs to relax and recover, yet the periphery needs
to be perfused. And so if we can let the heart
work the least possible while achieving all these things, then
that's the best situation for myocardial recovery.
So minimization of myocardial oxygen consumption is the
(10:07):
special thing that no one has been able to achieve.
And what are you doing that you hope will allow you to achieve
this? We have the benefit of some
consultants and some work that gives us a way to monitor and
quickly measure myocardial oxygen consumption live.
(10:32):
And we also have some physiological construct based on
cardiac output venous return surface that lets us do a good
job of maintaining left atrial pressure at a specific range.
At same time, we regulate right atrial pressure.
So the combination of the cardiac venous return surface
(10:54):
and technologies to instantaneously track or
periodically track monocardial auto consumption are some of the
benefits. But there's some other
techniques that we have as well.Folks, you can ask questions.
If you're watching on LinkedIn, pop your questions into the
LinkedIn chat. If you're watching on Twitter X,
(11:17):
use the hashtag CXO Talk and askDoctor Alexander pretty much
whatever you want. Now is a special opportunity to
to do that. Joe, how does what you were just
describing then relate back to the digital twin?
(11:37):
My lab is interested in the human digital twin.
We started with cardiovascular system because we have expertise
and it's #1 killer global and wethought it's a good place to
start. As well as the fact that the
cardiovascular system even attacking something like keep
myocardial infarction, it makes it necessary to consider other
(11:59):
organ systems like the lungs andthe kidneys and and
neuroregulation, this autonomousclosed loop intervention system.
The first guess is given by our electrical analogue model of the
human cardiovascular system thatbest fits a particular patient
who walks in. Now how do we determine the best
(12:21):
electrical model for the patientthat we see lying in front of
us? That relies on huge amounts of
population level data to start with, where we build a virtual
population of patients that are similar to the patient that
walks in, in terms of you know, age and conditions,
(12:43):
comorbidities, etcetera. We did we build a population of
virtual population similar to that patient.
Then based on a specific characteristic we narrow in on
the best guess and that might beseveral 100 patients that are
close to this patient. And we use that to get the
initial first guess on this closed loop, autonomous closed
(13:05):
loop intervention system. Then because it's a feedback
system, it will automatically adjust those parameters to give
the specified low Tatro pressure, cardiac output,
arterial pressure. Now mind you, that information
in the error signal, the correction signal that's being
fed back, that teaches the modelof that patient what those
(13:29):
parameters should be. So the model becomes more and
more specific for that patient as therapy is administered
through this autonomous close dependent pension system.
So it both zeroes in on the bestparameter choices and at the
same time teaches us what the best parameter choice choices
(13:49):
are for that patient. So we will create a registry
including that patient and thosechoices.
So you are gathering broad population data and then
isolating out of that broad group for each individual
patient the right subset of datathat most closely corresponds to
(14:14):
that particular patient's situation.
And then from there you're building the digital twin of
that specific patient. Is is that a correct way to say
it? That's very correct.
We think of the digital twin as a system of characteristics that
represent a patient, characteristics like resistors,
(14:37):
capacitors, dialogues and and soforth.
So we want the those parameters,the best choices that most
represent that patient and that patient's characteristics.
And so we begin with population level data that's in the
literature. And from there, we build a
virtual population that is relevant for this particular
(14:57):
patient. And from within that virtual
population, we find the closest patients for that patient.
Then we begin this process of driving the autonomous closest
intervention system. In the course of that feedback
regulation and adjustment, then we learn to titrate or tune
(15:17):
those parameters to match that particular patient.
What kinds of data are you usingas inputs into all of this?
In the world of cardiovascular about digital twins, there are
many places we could have started.
We started with acute myocardialin parks and acute heart
(15:38):
failure, mostly because it's acute and mostly because the
patients will be in the ICU or CCU and heavily monitored and
heavily instrumented where we have readings of dynamic signals
like dynamic pressure pressures and volumes and we have
(15:59):
echocardiography. So we have a lot of time
dependent signals that we can use to train our models and
patient and set those parametersand and gather data.
So the inputs in the ICUCCU are,you know, measurements of left
(16:20):
atrial pressure, right atrial pressure, cardiac output.
We can do SWAN catheters, we canmonitor pressure over periods of
time. And usually the admission for
that patient is about four or five days.
And so we can have four or five days of acute information to
(16:40):
tune to that particular patient.So we chose the easiest option
to be honest as a place to begin.
When you go to chronic heart failure, you have to rely on
data that you know from patientsthat are not in the hospital,
and they're static measurements and they're based on wearable
(17:01):
technologies, not nearly invasive enough to parameterize
such a system. But we'll go to chronic heart
failure next. Where are you in the process of
this set of developments? We have a group of
collaborators. The ones most relevant to what
we're talking about today are a group at National Cardiovascular
(17:24):
Center in Japan, where they havecardiologist, anesthesiologist,
experimentalist and are able to do animal experiments with
animals ranging in size from from mice all the way up to
cows. And so we're doing animal
experimentation to validate these approaches and test out
(17:47):
these technologies. And last year, for the first
time ever, we demonstrated the proper functioning of an
autonomous closed loop intervention system to treat
heart failure in an experimentalanimal, a dog that that had
experimentally induced heart failure and the system
(18:08):
compensated and automatically treated that dog while
minimizing myocardial oxygen consumption.
We're at a stage now where we'recollecting a number of
experiments from animals where we'll prove, we intend to prove
that the autonomous close up intervention system outperforms
(18:29):
the standard of care. By showing that these
cardiologists and their treatment compared to the
autonomous system cannot yield the such a small myocardial
infarct size as what we can get from the autonomous close up
intervention system. So reduction in infarct size
(18:49):
would be physical evidence of how this system could outperform
clinicians and that's what we intend to do.
The autonomous system then is relying on this general
population data plus very detailed data coming from the
(19:09):
particular patient who is in theICU or the CCU, because there's
lots of data and you're finding so far are a greater degree of
accuracy and accuracy in what specifically that is producing
the result. We are still in the very early
(19:32):
stages and developing you know what we can do in the real life
clinical situation. So we're all of what I'm saying
is based on hopes that depend onwhat we've been able to validate
then verify in animal experiment.
So we're still at the animal experimentation phase and so we
(19:53):
are not looking at population level data in animals.
We're even beginning without that kind of referenced model.
Of parameters for animals, except what we can get from the
literature very simply and stillthis autonomous system is able
(20:14):
to perform. We demonstrated last year.
So, but your description of the process is correct.
That's what we hope to achieve and there are many challenges as
you can imagine. Again, we're at the early stages
achieving this in animals and before we get anywhere near
(20:35):
regulatory approval, we have to go through and I'm sure you're
aware of these types of hurdles.We have to use this kind of
system to do clinical decision support before we can go to
complete autonomy. And so that means the feedback
would be going through a physician, the physician would
be in this closed loop and we advised the physician and the
(20:57):
physician would make a decision whether to do what the software
says or not. So we have many things to
achieve before we before I can really answer your question
well. I guess similar to autonomous
vehicles, before you let a fleetof autonomous taxis on the road,
(21:18):
you have a driver in the car because you really don't know
what's going to happen until youdie.
It try it because it's new. So it's basically the same
concept. That's exactly right.
So some people have suggested that we not call it autonomous
intervention, but like a copilotsystem.
But we're ambitious in thinking that this kind of a system can
(21:40):
replace the need for a specialist to be at the
patient's bedside. It could be a system that's
monitored by, you know, another clinical professional, but may
not need, probably will not needa specialist, which makes it
accessible to kind of remote hospitals with fewer resources.
(22:04):
And so that's part of our visionfor this, this HealthEquity.
We have an important question from Twitter, from Arsalan Khan,
who's a regular listener. He always asks excellent
questions and he says this. It's this is a very good way
into getting to to dive into theethical and equity aspects of
(22:30):
this. He says this How do you make
sure data from these medical instruments is correct and the
doctor makes the right decision based on the data versus his or
her personal experience? My bias to be honest, and one
clinician made this comment. He said that he'd rather have an
(22:53):
autonomous system that focusing beat to beat on adjusting meds
for a patient than to have an exhausted resident standing in
the hallway checking his phone and doesn't have enough time to
look in on the patient. We had the bias that if the
system is built correctly, then it will correct even for errors
(23:17):
that the position might make. But it's a, it's necessary like
you described for levels of autonomy in vehicles to have
that supervision. FDA is not going to approve an
autonomous device no matter, no matter how well it's tested
without that step of clinical decision support.
(23:40):
And so we we have to do that. We can't take the position out
of the loop. And then we have another related
question from Elizabeth Shaw, who says how do you know you can
trust the model and then the decisions in healthcare, the
standard is let AI guide you, but the doctor should make the
decision. Yes, it's related.
(24:01):
Our first pass is to include thedoctor in loop so the doctor can
judge, you know, what the the system is telling him, him or
her. And let me just say that the
thing about our system that's different from AI is that it's
causally based. Even though cardiologists don't
(24:23):
do as much Physiology as they once did, the decisions that
would be coming out of our system will be based on clinical
practice, cardiovascular Physiology, cardiovascular
dynamics, and it won't be a black box deep learning type
(24:45):
output that cannot be reviewed and and criticized.
I'm a technologist and I absolutely support AI and
autonomous systems when they're used properly with the right
type of governance. But if we look for example, at
the Boeing 737 Max crashes that were caused by faulty sensor
(25:10):
algorithms and then incorrectly documented or incompletely
documented recovery procedures, it's an example of an autonomous
life and death system. And, and so the idea of of
having this kind of autonomous system relating to the to
(25:31):
medical medicine is kind of scary.
There's certainly lessons to be be learned, and some of the some
of the lessons are avoiding shortcuts and some of the
lessons have to do with hubris. Certainly fail safe mechanisms
have to be built in and certainly you don't want to put
critical components in the hands, at least I don't of AI.
(25:53):
The short answer is the work we do is a beautiful dance between
creativity and skepticism and we're skeptical always.
And we really don't trust and certainly don't fully rely on
systems like AI, or I would say we depend on AI for particular
and certain algorithms for particular applications or
(26:15):
particular aspects. That's why our approach is
crucially causal and mechanisticand based on Physiology so that
we can see better than what is being called, I guess,
explainable AI. Explainable AI doesn't give the
kinds of explanations that cardiologists understand in
(26:37):
terms of Physiology and mechanisms and we do provide
those. There will be rationales for
what our systems would advise the physician when the
physician's in the loop. Now when the physician done in
the in the loop, then we can't say that the system will be
(26:57):
perfect all at once, but we'll suddenly be designing with with
safety in mind and risk. I should also just just mention
that some of the, the, the deeper questions that have to do
with testing implementation, they will involve large scale
(27:22):
clinical trials just like drugs.So the same questions that are
being raised about algorithms and how do you know if they're
safe or not, you know, can be raised about drug development.
And we accept the results of clinical trials even when we see
these commercials saying that there's a risk of death with,
(27:47):
you know, some of the with some of the the medications under
certain conditions and that wereproven in the trial.
And so there is no perfect solution without some level of
risk. And the goal would be to
minimize those risks. Now as entity research is Bell
Labs type moonshine ideas place,we're looking to develop proof
(28:09):
of concept to develop a regulatory strategy and to hand
off to those those companies to industry that knows how to go
into that risk safety environment like pharmaceutical
companies, like medical device companies.
And we haven't touched on medical devices, but everything
I'm describing about drug treatment also applies to the
(28:33):
management of devices controlling patients who are
more severely I'll. So we will be looking to partner
with pharmaceutical companies, medical device companies that
are able to afford that have expertise and ethical aspects,
safety aspects and conducting those, affording those large
scale clinical trials. Then we'll move on to the next
(28:55):
place so we can paint beautiful pictures and brilliant
landscapes on our small prison walls.
You mentioned several times thatyour work is causal and
mechanistic. Can you dive into that and
explain what you mean with in comparison to things that that
(29:16):
are not causal or mechanistic? If we think about a patient who
has heart failure, heart failureusually in patients end up with
some degree of renal failure. The renal failure causes an
increase in volume retention. The volume retention means that
(29:37):
you get fluid in the circulatorysystem that flows back to the
lungs and you get fluid in the lungs, fluid in the lungs.
Elevation of left atrial pressure means that you have
poor oxygen diffusion capacity, so you don't get as much oxygen
across the lungs into the circulatory system because you
don't have as much oxygen. The heart has to pump harder and
(29:59):
faster because in order to deliver enough oxygen to confuse
the tissues, in order to do that, the heart has to work
harder. This patient with heart failure,
that patient's heart has to workharder and consume more
myocardial oxygen. It needs more oxygen, but it
can't get more oxygen. So the heart is feeding itself,
(30:20):
but it can't get enough oxygen to feed itself, and you quickly
deteriorate. That is a sequence of events set
in place by various causes that originate with heart failure.
That's different from, for example, having some speckled
(30:40):
tracking images that are fed into AI and making some
hypotheses from that alone aboutwhy a patient is not ventilating
well. That for me is pattern
recognition with an attempt to extrapolate to causality.
(31:01):
But this sequence of events I just described, which also by
the way includes anemia and which exacerbates the situation
more, these are all things knownphysiologically and
mechanistically to play a significant roles in heart
(31:21):
failure. And all of the drugs that have
been developed are mostly addressing these various
aspects. Diuretics to relieve, remove
volume, contract catecholamines to improve contractile function.
But at some point, the cost in myocardial oxygen consumption is
(31:42):
just really the last draw for a patient with chronic heart
failure. It's very, very serious.
And the intervention that your system is doing, correct me if
I'm wrong, is injecting certain types of medications and
precisely the right dosages and the right times.
(32:03):
Is that correct or is there moreto it?
That's correct. For our particular use case of
acute myocardial infarctum, acute heart failure where we
have control over everything andwe're measuring everything.
It does those injections to manage the acute case.
But the best we can expect for that autonomous close up
(32:25):
intervention system is the acutecase.
And so four or five day hospitalstay it doesn't.
It gives us some knowledge that's helpful in the chronic
case when the patient's not in the hospital.
We have some starting places, but we need much more data, we
need much more laboratory chemistry.
(32:45):
We need to be monitoring patients with wearable
technologies and so forth over longer periods of time.
However, we expect that the excellent care in the acute
situation will have benefits forthe patient when they're
released from the hospital. They won't be readmitted.
(33:07):
For example, the readmission rate we expect would be less or
would be over a longer period oftime compared to the standard of
care. You described earlier the
analogy of an electrical system with diodes and resistors, and
(33:30):
as I was reading through some ofyour papers and presentations, I
was struck by these electrical diagrams.
I mean literally like schematic diagrams that you'd see for
electrical devices. Can you tell us more about that
that aspect? When we naturally, when we look
(33:51):
at that diagram, we see what looks like the electrical
circuit and the circulatory system is a circuit.
And so that part, it kind of works for us naturally.
And actually, elements in that electrical analogue diagram
represent left ventricular contraction, right atrial, right
ventricular contraction, atrial contractions.
(34:13):
The person who is familiar with those symbols can see the entire
circulatory system represented there, but they're represented
very simply. For example, just to give you a
notion, the left ventricle ejects into the systemic
arterial system across the aortic valve.
(34:33):
So for us the left ventricle is a capacitor where for a certain
volume, you know it generates a certain force and we call it a
time varying capacitor because we let the capacitor level
change or electric ventricle become stiffer to generate
pressure. That pressure in the electrical
allogue sense is a voltage, and that voltage in the heart in
(34:56):
that left ventricle gets higher than what's downstream across
the valve. Then that valve will open or
that diode will open and flow will go out of the diode or
across the aortic valve into theanother set of resistors and
capacitors which we say represent the systemic arterial
system. So resistance in the electrical
(35:20):
analogue network is the relationship between flow and
pressure in the real system capacitor.
A capacitor in that electrical analogue network is
representation between volume that goes into an elastic
artery, it stretches the wall ofthe artery and that generates a
(35:43):
certain pressure. So by a series or a system of
resistors and capacitors and diodes representing different
compartments around the arterialsystemic arterial system, the
venous systemic venous system, the pulmonary arterial system,
the pulmonary venous system, we can represent the entire, simply
(36:07):
represent and I know simple, simple is a relative term, but
simply represent the entire circulatory loop in the
essential measurements or parameters or variables that are
important for our application toacute myocardial infarction and
acute heart failure. And how does this type of
representation interplay with the creation of the digital
(36:33):
twin? This representation is our
digital twin. It's a mathematical
representation of the patient inthese terms.
Now what I described to you was the simple circulatory loop.
We have other subsystems that are more specialized that are
not represented in probably the diagrams you looked at that
(36:56):
represent the the coronary circulation.
Now, if you think about the heart, the heart generates a
pressure and it generates a flow.
That flow goes into the systemicarteries across the aorta.
But that same aorta is has a coronary circulation that
supplies blood to the heart itself.
(37:18):
That's why I say the heart feedsitself in supplying its own
blood supply. That coronary, that coronary
circulation is has a little moresophistication that we need to
account for. If there's auto regulation of
coronary flow, we need to account for that.
When the heart beats, it squeezes the vessel surrounding
(37:38):
it, so it squeezes its own coronary vessels.
And so we need to account for that in describing what flow
gets to an ischemic heart. So we have to do a more
sophisticated model drilling in on the heart and this
metabolism. Now it turns out that myocardial
oxygen consumption, cardiac metabolism has been thoroughly
(38:02):
described in simple terms that relate to pressure and volume
mechanics. Again, those signals that we're
measuring all the time, we're interested in pressures and
volumes you can measure. You can use a pressure volume
loop to represent the mechanics of a Chamber of the heart, where
the slope of the pressure volumerelationship is an index of the
(38:24):
ability of the heart to contract.
So a simple index of the abilityof the heart to contract can be
derived by pressure volume measurements as well.
Those same simple signals and the areas in various aspects for
the pressure volume can show potential energy of the heart,
the work that the heart does, and they correlate with
(38:45):
myocardial oxygen consumption. So there's a tremendous amount
of data about the heart that's implicit in the way that we
study it using pressures, flows and volumes.
And Arsalan Khan on Twitter comes back again and he asks,
with your research, what did youfind or did you find that the
(39:06):
systems and processes, traditional ways of describing
the body were either not useful or overly complicated?
How did you how did you end up with this electrical analog?
In the world of cardiology, although the electrical analog
(39:28):
model is not so thoroughly represented as what we're using,
it has become a traditional approach to characterizing
hemodynamics. And so pressure volume area in
systolic, pressure volume relationship, the representation
(39:51):
of the dynamics of the circulatory system and neural
control of the circulation usingbaroreceptors and so forth.
This way electrical analogue wayis traditional, It's just that
many practitioners don't understand it on this level.
(40:13):
They're using volume conductancecatheters in the Cath lab to
measure volume and pressure and generate pressure volume loops
and generating in indices of or indexes of contractility.
So it's not non traditional, it's just that we're taking it
on a whole nother level and extending it and using some non
(40:33):
traditional or, or some elementsthat we discover and that have
been published but have not beentraditionally used because of,
I'll just say, complexity of biomedical engineering compared
to clinical medicine. So a lot of clinical
(40:54):
cardiologists are using the principles but don't understand
it so thoroughly. Now the I mentioned, I may have
mentioned a consultant that we have on the project, his name is
Kenji Sonogawa. He discovered and prove many of
the principles that are being used in clinical cardiology
(41:15):
today that are based on pressurevolume characteristics and
electrical analogue models. There's an electrical analogue
model of the systemic arterial system called impedance that not
only represent is not only represented representing the
resistances, but the frequency dependent nature of pressure and
flow in the dynamics of the arterial system.
(41:36):
People are using results from impedance characteristics
without being able to measure impedance, which is another
electrical term because it's just a little bit more
complicated than what people areare ready for.
And so all of these insights, for example, another example is
ventricular vascular coupling. Cardiologists talk about
(41:58):
ventricular vascular coupling and one concept is to think of
the heart that's a as a, if you will, a battery with a charge
and the arterial system represented by pressure
installed. Pressure and stroke volume
relationship where you have two capacitors and opening the
(42:20):
aortic valve is the closing of aswitch between the capacitors
and volume will move from one capacitor to the other.
That notion of ventricular vascular coupling is something
also derived by Doctor Kenji Senegal, who's a consultant.
And I'll just add that a lot of the work I did together with
(42:40):
Kenji Sunugawa when I was a graduate student at Johns
Hopkins. And so we have a long history of
work using electrical analog models dating back to the 70s.
How much of the math did you need to develop versus relying
(43:02):
on traditional or existing understanding of these various
relationships? As a biomedical engineer, we've
used, you know, whatever popularmathematics was available at the
time that might render some new piece of information about
biology to be understandable or to make sense.
(43:23):
So I, I started way back in the 70s doing neural networks, you
know, when it was a fashionable thing.
We did wavelet analysis of waveforms and reflection
coefficients reflected waves. So we as biomedical engineers,
we love complexity. And you know, we'll use
(43:45):
homeomorphic deconvolution mathematics of, you know,
measuring earthquakes, anything that's available, we'll learn
and use how much, as much as necessary and practical.
So we built back in the day we built, we had RadioShack
computers controlling analog computers that we patched in in
(44:08):
order to have real time control of an isolated perfused dog
heart preparation. That was just the way we do
things. So I don't know how much
mathematics, but as much as necessary to be practical.
There's a image near Ian Hunter at MIT who once said to me, he
(44:31):
said scientists like taking things apart and you engineers
like putting things together. So I in retrospect, everything
I've done, I'm more an engineer than a scientist, has been
integrative. How to explain the system when
it's already put together. And sometimes that takes a lot
(44:54):
of mathematics, and we'll do what's necessary about them.
Sounds like you're building the bionic system.
Remember there used to be ATV show called the $6 million man
that was sounds like you're you're building the bionic
approach. What about the the team required
(45:16):
to build this? There are so many different
skills that are involved. One comment to your comment
about the bionic man by like woman, woman is that Kenji and I
once wrote a paper around the bionic baroreceptor reflex with
exactly that notion in mind, where you have a device that
(45:38):
automatically can regulate arterial pressure reproducing a
failed autonomic system. And there's a syndrome called
Scheidrager syndrome where the receptor doesn't doesn't work
and that kind of system has beenproven to work.
So we back in the 70s or 80s, wewere developing A bionic
(46:00):
baroreceptor reflex. Yes, on our team in the May lab,
immediately we have a number of PhDs in biomedical engineers, in
biomedical engineering, we have MD's, we have masters people and
everybody can do programming. Everybody knows Physiology
(46:24):
either from the beginning or they've learned from our
consultants, and we have routineconferences with cardiologists
and anesthesiologists doing experiments together with them.
So we have that level of expertise.
And at the National Cardiovascular Center, I
mentioned earlier, those guys are seeing patients on the one
(46:47):
hand, then doing experiments in animals, then writing programs
and then building electrical analog circuits.
The same person is doing all of those things.
And so we have access to all of their training and their
expertise. And this is an occasion where I
should also mention that I've talked about our main
collaborator and our main projects on the larger scale.
(47:12):
We also have collaborations looking at other aspects of
digital twins with Harvard Disease Biophysics Group, where
we're working on building developing a bio hybrid digital
twin strategy using stem cells to manufacture tissue engineer
(47:34):
hearts from patients own stem cells and then study those
chambers beating and isolation. And we have a goal or they have
a goal and we're trying to facilitate them to build a heart
based on on patients own stem cells and tissues generated in a
bioreactor from those stem cells.
(47:55):
So there were tapping into tissue engineering and lots of
machine learning. We published a paper with them
in Science Robotics on machine learning inspired design of a
ray, where they posited that it may be possible with AI and
(48:17):
machine learning to have biological design evolve at a
more rapid rate than evolution on the basis of what would be
predicted from some first principles that we know about
how tissues function. Not necessarily a trial and
error approach to revolution, but a specific design inspired
(48:37):
by machine learning. Then we have a Group, A
collaboration with Technical University of Munich, where
they're developing different types of biosensors and
different types of technologies for in vivo measurements of
electrical signals. And so we have a lot of
(48:59):
expertise available in our collaboration done, a lot of
expertise local to our small team, which is only about 9
people at the moment. So you really are fusing the
Physiology with medical practice, research with robotics
(49:24):
and the math abstracted to create your models?
I mean, it's so interdisciplinary.
It is. And while we have a strategy,
we're also again we're that BellLabs type group where you get a
bunch of creative people who have lots of new ideas.
So some of this, the work that we're doing is born out of many
(49:47):
great ideas that need to be pursued at the same time that we
try to keep a larger strategy inmind so that everyone is working
around a kind of organizing principle for what we want to
achieve. And one of the main goals is
improved patient outcomes. Some of those goals we expect to
(50:10):
achieve in the longer term, but some we want to demonstrate in
the shorter term. And this autonomous closed loop
intervention system is turning out to be at least in terms of
first in human studies. We expect to see some first in
human studies within five years.You brought up earlier the
(50:35):
HealthEquity vision that you have and how does this research
relate to that and very quickly,because again, we're going to
run out of time. If you think of the quintuple
aims of healthcare and that's improved experience of the
patient, improved experience forthe clinician, improved
population health, decreased cost, improved HealthEquity,
(50:59):
then I think for example, the autonomous flows of an invention
system does all of that. So improves HealthEquity by not
requiring so many specialists inorder to treat a very deadly
cardiovascular disease. The cost, we hope, should be
(51:19):
less. Although the history of medicine
would suggest that no matter howdramatic an improvement we make
in scientific advancement, the entrepreneurs find a way to make
it more costly. But we think that this
technology will be less, less costly and should improve the
(51:41):
physician experience. I won't be so exhausted having
to give so much care to managingparticular patients and the
benefits just strike directly, Ithink with the quintuple aims of
improved healthcare. Where is all this going?
Where is your research headed right now?
In the immediate term, we're headed toward those first in
(52:04):
human studies that have to do with clinical decision support
and we're looking to develop regulatory strategy with a
partner and hand off the expertise, hand off to them and
their expertise and investments in large scale clinical trials
(52:27):
and developing technology devices or pharmaceuticals.
And then we would move to focusing on chronic heart
failure and that's where we would expect to make use more of
the collaborations we have with Harvard Disease Biophysics Group
(52:48):
and with our folks at TUM wearable devices and such.
And again, this is all starting from the notion of a human bio
digital twin or human bio digital twins, not just
cardiovascular. It was a starting place and I'm
personally interested in neurodegenerative diseases and
(53:09):
so if I'm around long enough, I hope we can get to to that
branch of work as well. All right.
And with that, I'm afraid we're out of time.
A huge thank you to Doctor Joe Alexander.
He is director of the medical and health Informatics lab at
(53:31):
NTT Research. Joe, thank you so much for
taking time to be with us. It's been a very fascinating
glimpse into your work. Thank you so much Michael for
having me. Everybody, thank you for
watching. Before you go, subscribe to the
CXO Talk newsletter so we can notify you about upcoming shows.
(53:53):
We have great shows that are coming up.
Thanks so much everybody for watching and we'll see you again
next time. Have a great day.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
My Favorite Murder with Karen Kilgariff and Georgia Hardstark
My Favorite Murder is a true crime comedy podcast hosted by Karen Kilgariff and Georgia Hardstark. Each week, Karen and Georgia share compelling true crimes and hometown stories from friends and listeners. Since MFM launched in January of 2016, Karen and Georgia have shared their lifelong interest in true crime and have covered stories of infamous serial killers like the Night Stalker, mysterious cold cases, captivating cults, incredible survivor stories and important events from history like the Tulsa race massacre of 1921. My Favorite Murder is part of the Exactly Right podcast network that provides a platform for bold, creative voices to bring to life provocative, entertaining and relatable stories for audiences everywhere. The Exactly Right roster of podcasts covers a variety of topics including historic true crime, comedic interviews and news, science, pop culture and more. Podcasts on the network include Buried Bones with Kate Winkler Dawson and Paul Holes, That's Messed Up: An SVU Podcast, This Podcast Will Kill You, Bananas and more.