Circulation March 30, 2021 Issue

For this week's Feature Discussion, please join authors Michael
Ackerman, Christopher Haggerty, editorialist Michael Rosenberg,
and Associate Editor Nicholas Mills as they discuss the original
research articles “Artificial Intelligence-Enabled Assessment of
the Heart Rate Corrected QT Interval Using a Mobile
Electrocardiogram Device,” “ Deep Neural Networks Can Predict
New-Onset Atrial Fibrillation From the 12-Lead Electrocardiogram
and Help Identify Those at Risk of AF-Related Stroke,” and
“Trusting Magic: Interpretability of Predictions from Machine
Learning Algorithms.”


 


TRANSCRIPT BELOW:


Dr. Carolyn Lam:


Welcome to Circulation on the Run, your weekly podcast summary
and backstage pass to the journal and its editors. We're your
cohosts. I'm doctor Carolyn Lam, associate editor from the
National Heart Center and Duke National University of Singapore.


Dr. Greg Hundley:


And I'm Greg Hundley, associate editor, director of the Pauley
Heart Center at VCU Health in Richmond, Virginia. Well Carolyn,
this week's feature, it's kind of a new thing for us. It's more
than our double feature; it's actually a forum, where we're going
to have two papers discussed, we'll have both authors represented
from each of those two papers, we'll have an editorialist, and
we'll have one of our associate editors. And the topic, Carolyn,
just to keep you in suspense, is really on machine learning and
actually how that can be applied to 12 lead electrocardiograms.
But before we get to that, how about we grab a cup of coffee and
start off on some of the other articles in this issue? Would you
like to go first?


Dr. Carolyn Lam:


Yes, I would, but you're really keeping me in suspense. But
first, let's focus on health related quality of life. We know
that poor quality of life is common in heart failure, but there
are few data on heart health related quality of life and its
association with mortality outside of the Western countries.
Well, until today's paper. And it's from the Global Congestive
Heart Failure, or GCHF study, the largest study that has
systematically examined health-related quality of life as
measured by the Kansas City cardiomyopathy questionnaire 12, or
KCCQ, and its association with outcomes in more than 23,000
patients with heart failure across 40 countries, in eight major
geographic regions, spanning five continents.


Dr. Greg Hundley:


Wow, Carolyn. That KCCQ 12, that has been such an interesting
tool for us to use in patients with heart failure. So what did
they find in this study?


Dr. Carolyn Lam:


Really important. So the health-related quality of life differs
considerably between geographic regions with markedly lower
quality of life related to heart failure in Africa than
elsewhere. Quality of life was a strong predictor of death and
heart failure hospitalization in all regions, irrespective of
symptoms class, and in both preserved and reduced ejection
fraction. So there are some important clinical implications,
namely that health-related quality of life is an inexpensive and
simple prognostic marker that may be useful in characterizing
symptom severity and prognosis in patients with heart failure.
And there is certainly a need to address disparities that impact
quality of life in patients with heart failure in different
regions of the world.


Dr. Greg Hundley:


Very nice, Carolyn. Well, I'm going to turn to the world of basic
science and bring us a paper from David Merryman from Vanderbilt
University. So Carolyn, myocardial infarction induces an intense
injury response, which ultimately generates a collagen dominated
scar. Cardiac myofibroblasts are the cells tasked with depositing
and remodeling collagen and are a prime target to limit the
fibrotic process post myocardial infarction. Now Carolyn,
serotonin 2B receptor signaling has been shown to be harmful in a
variety of cardiopulmonary pathologies, and could play an
important role in mediating scar formation after MI. So Carolyn,
these investigators employed two pharmacologic antagonists to
explore the effect of serotonin 2B receptor inhibition on
outcomes post myocardial infarction and characterized the
histological and micro structural changes involved in tissue
remodeling.


Dr. Carolyn Lam:


Oh, that's very interesting, Greg. What did they find?


Dr. Greg Hundley:


So Carolyn, serotonin 2B receptor antagonism preserved cardiac
structure and function by facilitating a less fibrotic scar,
indicated in their results by decreased scar thickness and
decreased border zone area. Serotonin 2B receptor antagonism
resulted in collagen fiber redistribution to a thinner collagen
fiber. And they were more anisotropic. They enhanced left
ventricular contractility and the fibrotic tissue stiffness was
decreased, thereby limiting the hypertrophic response of the
uninjured cardiomyocytes.


Dr. Carolyn Lam:


Wow. That is really fascinating, Greg. Summarize it for us.


Dr. Greg Hundley:


Yeah, sure. So this study, Carolyn, suggests that early
inhibition of serotonin 2B receptor signaling after myocardial
infarction is sufficient to optimize scar formation, resulting in
a functional scar, which is less likely to expand beyond the
initial infarct and cause long-term remodeling. The prolonged
presence of the antagonist was not required to maintain the
benefits observed in the early stages after injury, indicating
that acute treatment can alter chronic remodeling. So Carolyn,
it's really going to be interesting to see how this research
question is pursued in studies of larger animals, including us,
or human subjects.


Dr. Carolyn Lam:


Wow, that is really interesting. And so is this next paper. Well,
we know that genetic variation in coding regions of genes are
known to cause inherited cardiomyopathies and heart failure. For
example, mutations in MYH7 are a common cause of hypertrophic
cardiomyopathy, while mutations in LMNA are a common cause of
dilated cardiomyopathy with arrhythmias. Now, to define the
contribution of non-coding variations, though, today's authors,
led by Dr. Elizabeth McNelly from Northwestern University
Feinberg School of Medicine in Chicago and colleagues evaluated
the regulatory regions for these two commonly mutated
cardiomyopathy genes, namely MYH7 and LMNA.


Dr. Greg Hundley:


Wow, Carolyn. So this is really interesting. So how did they do
this and what did they find?


Dr. Carolyn Lam:


You asked the top questions, because the method is just as
interesting as the findings here. They used an integrative
analysis that relied on more than 20 heart enhancer function and
enhancer target datasets to identify MYH7 and LMNA left
ventricular enhancer regions. They confirmed the activity of
these regions using reporter assay and CRISPR mediated deletion
of human cardiomyocytes derived from induced pluripotent STEM
cells. These regulatory regions contained sequence variants
within transcription factor binding sites that altered enhancer
function. Extending the strategy genome-wide, they identified an
enhancer modifying variant upstream of MYH7. One specific genetic
variant correlated with cardiomyopathy features derived from
biobank and electronic health record information, including a
more dilated left ventricle over time. So these findings really
link non-coding enhancer variation to cardiomyopathy phenotypes,
and provide direct evidence of the importance of genetic
background. Beautiful paper.


Dr. Greg Hundley:


Very nice, Carolyn.


Dr. Carolyn Lam:


But let me quickly tell you what else is in this issue. We have
an ECG Challenge by Dr. Lutz on flash pulmonary edema in a
70-year-old; there's an On My Mind paper by Dr. Halushka,
entitled (An) Urgent Need for Studies of the Late Effects of
SARS-CoV-2 on the Cardiovascular System.


Dr. Greg Hundley:


Ah, Carolyn. Well, in the mailbox, there are two Research
Letters, one from Dr. Soman entitled (The) Prevalence of Atrial
Fibrillation and Thromboembolic Risk in Wild-Type Transthyretin
Amyloid Cardiomyopathy, and a second letter from Dr. Berger
entitled Multiple Biomarker Approaches to Risk Stratification in
COVID-19. Well Carolyn, now let's get on to that forum discussion
and hear a little bit more about using machine learning in the
interpretation of a 12 lead ECG.


Dr. Carolyn Lam:


Wow, can't wait. Thanks, Greg.


Dr. Greg Hundley:


Well listeners, we are here today for a double feature, but this
double feature is somewhat unique, in that we are going to
discuss together two papers that focus on machine learning
applications as they relate to the interpretation of the
electrocardiogram. With us today, we have Mike Ackerman from Mayo
Clinic, Chris Haggerty from Geisinger, Mike Rosenberg as an
editorialist from University of Colorado, and then our own Nick
Mills, an associate editor with Circulation. Welcome, gentlemen.
Well, Mike Ackerman, we will start with you first. Could you
describe for us the hypothesis that you wanted to test, and what
was your study population and your study design?


Dr. Michael Ackerman:


Thanks, Greg. The hypothesis was pretty simple, and that is could
an artificial intelligence based approach, machine learning, deep
neural network, could that solve the QT problem? Which is one of
the big secrets among cardiologists, which, as you know, one of
your associate editors, Sammy Biskin, published a sobering paper
over a decade ago, showing and revealing the secret that
cardiologists are not so hot at measuring the QT interval, and
heart rhythm specialists sometimes don't get it right either. And
we all know that the 12 lead ECG itself is vexed by its computer
algorithms at getting the QTC just right, compared to those of us
who would view ourselves as QT aficionados. And so we were hoping
that a machine learning approach would solve this and help us
glean, one, a very accurate QTC, as accurate as I can make it
when I measure it, or core labs that do QT measuring for living.


Dr. Michael Ackerman:


And two, could we get that QTC from just a couple of leads to be
as accurate as what the whole 12 lead ECG would be seeing so that
we can move it to a mobile smartphone enabled solution? And so
that was our hypothesis going forward, and we studied a lot of
patients. And that's something that machine learning and the
power of computation does, that in my world, I'm used to studying
a hundred or a thousand patients with congenital long QT syndrome
and thinking that I've assembled a large cohort, but for this
study, we started with over two and a half million ECGs from over
650,000 people. And then ultimately, through training, testing,
and validation of about 1.6 million ECGs from over a half a
million individuals to sort of teach the computer or have the AI
algorithm get the QT interval not too hot, not too cold, but just
right. And as we'll discuss, I think we hit the mark.


Dr. Greg Hundley:


Thanks so much, Mike, what did you find?


Dr. Michael Ackerman:


Ultimately, we were able to show that with this drill, we could
get the deep neural network derived QTC to be give or take two
plus minus 20 milliseconds from what would the standard of care,
and that being a technician over-read QTC. But then we took, I
would say, pretty unique to AI studies, as many AI studies, just
do training, testing, and validation for study number one. And
then a future paper of a prospective study. But we did that
prospective study within this single paper with a subsequent
about two year enrollment of nearly 700 patients that I evaluated
in our genetic heart rhythm clinic at Mayo Clinic. And half of
those patients have congenital long QT syndrome, half did not.
And what we showed was that the deep neural network derived QTC
from a mobile ECG approximated the subsequent or the just prior
12 lead ECG within one millisecond, +/- 20 millisecond territory.


Dr. Michael Ackerman:


And it's ability to say is the QTC above or below 500, which we
all know is sort of a warning sign, that's a very actionable ECG
finding, do something about it, that that 500 millisecond cutoff
by the deep neural network gave us an area under the curve of
0.97, which from a screening perspective, that AUC is far higher
than a lot of AUCs for a lot of screening tests done in the
cancer world and so forth. And so we think we are very close to
what I've called a pivot point, where we will soon pivot from the
way we've been doing the QTC since Eindhoven over a century ago
to a fundamentally new way of deriving a QTC that's precise and
accurate and mobile enabled.


Dr. Greg Hundley:


Very nice, Mike. So using machine learning to accurately assess
the QTC from just two leads of an electrocardiogram. Well Chris,
you also have a paper in this issue of circulation that pertains
to another application of machine learning and looking at the
electrocardiogram. Can you describe for us your study population,
study design, and then also the question you were trying to
address?


Dr. Christopher Haggerty:


Sure. Yeah, thanks Greg. Great to be here with you all today.
Very similar to Mike's study, the motivation for us was we
believe very strongly that there's opportunities with using deep
learning applied to ECG data to uncover not only new knowledge
latent in the ECG itself related to the current patient context,
but also to try to predict future outcomes, future events. And
that was really our motivation, was to take that paradigm of
looking forward, in this case to predict new onset of atrial
fibrillation within a year. We used our Geisinger patient cohort,
which is a largely rural population in central Pennsylvania. We
have very longitudinal data for a lot of our patients, which
allows us to have this kind of design going back in our
electronic health records, in this case, our ECG database to 30
plus years.


Dr. Christopher Haggerty:


Similar big numbers that Mike described, and in our case, 1.6
million ECGs over 430,000 patients used to train the model. And
we had several different study designs that we employed. One just
being a simple proof of concept, asking can we accurately predict
new onset atrial fibrillation one year? And then a second study
design that was intended to simulate a real world deployment
scenario. Obviously the main rationale for trying to predict
atrial fibrillation is to then be able to treat and try to
prevent stroke. And so we tried to, as best we can in a
retrospective fashion, simulate a scenario in which we might use
this model to identify patients who went on to have a presumably
AFib associated stroke.


Dr. Greg Hundley:


And what did you find, Chris?


Dr. Christopher Haggerty:


So I think there are three main findings that we highlighted
here. So first, obviously we were building on the great work that
Mike and some of his colleagues at the Mayo Clinic have done,
showing that looking at AFib using deep neural networks needs to
be feasible. We extended it in this case by looking out further
than just an acute sense, looking at that one-year outcome. And
we had an area under the curve for our proof of concept of 0.85.
So area under the curve of 0.85 to identify patients with new
onset of atrial fibrillation within one year in our millions of
ECGs. Looking at it another way, the second main finding was that
that one year prediction was shown to have prognostic
significance beyond that one year, which is really interesting
and warrants a lot of further study. Looking over 30 years of
follow-up, patients predicted to be at high risk at baseline had
a hazard ratio of 7.2 for developing atrial fibrillation,
compared to those deemed to be low risk.


Dr. Christopher Haggerty:


And then really the third, and I think perhaps the most exciting
finding that we had here, was this simulated stroke experiment
that we had, where we identified patients from an internal stroke
registry and identified patients who had new diagnosis of AFib at
the time or up to a year after the stroke. So we can assume that
they were an AFib associated stroke. And subsequently, or I
should say previously, had an ECG that we could use to run
through the algorithm to predict their atrial fibrillation risk.
And we showed that the model performed well in this setting, that
of the 375 strokes that we identified, for example, over a
five-year period in our registry, we were able to identify 62% of
them within three years based on that ECG. So a number needed to
screen for an atrial fibrillation associated with stroke about
162, which compares favorably well to other screening techniques
that are out there, obviously. So we took that as a great proof
of concept that this type of AI technique might have benefits for
screening for atrial fibrillation and preventing strokes.


Dr. Greg Hundley:


Well congratulations, Chris. Well, we're now going to turn to our
associate editor, Dr. Nick Mills. And Nick, you have a lot of
manuscripts come across your desk. What attracted you to these
two papers, and what are the significance of the results as they
apply to ECG applications as we move forward?


Nick Mills:


Thanks, Greg. Yeah, this is a rapidly growing field, where the
availability of data scale with digital archiving and lots of
really interesting new methodologies are available to our
researchers. So we are receiving a lot of content in this area.
What I loved about these two papers is not just the quality of
the work, but also the really tangible benefits, potentially, for
patients. So AI does not need to be complex, but it does need to
solve a tangible problem. I guess what we look for in the
journal, beyond the kind of innovation and methodology, is
quality, and these studies used prospective validation, really
reliable end points, ascertainments, transparency, reporting, all
the things that we know are important for high quality clinical
research. I think the idea that we can bring QT monitoring to the
drug store on a portable device for our patients is potentially
transformative. I also think that to take a technology, the
electrocardiogram that we've been using for over a century, and
provide new insights that go way beyond my ability to interpret
the ECG, that might help us recommend a different course of
action for our patients is also just really exciting.


Dr. Greg Hundley:


Very nice. Thank you, Nick. Well Mike ... we're going to turn to
Mike Rosenberg now, listeners. And Mike wrote a wonderful
editorial, and I would invite you to work through this. As you
have an opportunity to read the journal and interact with it.
Mike, there are two different types of machine learning, I think,
that you described were used by the two respective investigative
groups. Could you describe those for our cardiology listeners?
What were the differences in those two approaches?


Dr. Michael Rosenberg:


Yeah, sure. And thank you for the opportunity to write the
editorial. Two very fascinating papers. I should say that they
both use the same approach of what's called supervised learning,
where you basically have a set of data inputs, and you're trying
to predict a labeled outcome. And what I talk about in the paper
is that what we've learned is if you have enough data and enough
computing power, you can predict almost anything highly
accurately. What's interesting about the two papers, and what I
sort of tried to contrast in the editorial, is that the one from
the Mayo Group and Dr. Ackerman, was basically predicting what's
already a known biomarker for sudden death, which is the QT
interval. And essentially, almost trying to automate that process
of predicting it accurately and in a way that, in essence, could
allow a home monitoring of patients for QT prolongation, which
obviously would be a huge benefit for clinicians, all those
alerts and things, to be able to have patients taking drugs that
are known to prolong the QT interval and feeling comfortable that
if they have any prolongation, it could be detected accurately.


Dr. Michael Rosenberg:


The second one, which is sort of interesting, and in contrast is
from the Geisinger Group and Dr. Haggerty, was the approach of
... where actually the prediction itself is actually the
biomarker. And we don't actually know exactly what it's using,
which I talk about a little bit of what that means and the
implications clinically, but in essence, what they showed was
that it actually is a very good biomarker and on par with what a
lot of us would consider to be very strong predictors of agents.
So I think it was two very interesting approaches to, again,
applying the same type of machine learning, but really
approaching it one from a more discovery side and another from
sort of validated or almost automating something that we do on a
daily basis.


Dr. Greg Hundley:


Thank you, Mike. So Mike, just coming back to you again, as we
read the literature, and most of us are clinicians or researchers
practicing, what should we look for when these new machine
learning manuscripts and research studies come out as to gauge,
"Ah, this is a really good study," or maybe not so much?


Dr. Michael Rosenberg:


Yeah. And it's a good question. I think one of the biggest
challenges, as I talked about, is interpretability. I think in
the clinical world, we're used to understanding the code for the
variables that go into our risk prediction model. And so I think
first and foremost is can I even understand what this is
predicting or am I sort of expected to take the predictions as
sort of a black box, it is what it is type of approach? I think
that there's other things that I just look at when I'm reviewing
these manuscripts. I mean, as I sort of mentioned, what these
models are really doing, it's not anything magical. What they're
doing is identifying patterns in the data and then using those to
make predictions, again, toward whatever label that you've
assigned them to.


Dr. Michael Rosenberg:


It's important that your data sets are split and that you're
training at one data set and then testing it in one that's
separate. And again, you can't ignore epidemiology. Is the data
set that you're training it reflective of the population that
you're going to be using those models in? And we know from
outside of healthcare, there's issues with models that have been
trained in one population where it's potentially biased or it's
potentially offering predictions that are using information we
may not necessarily want to use. Recidivism is a big example of
that. So I think that that's, first and foremost, it's sort of
taking a step back as a clinician and saying, "If this was a
biomarker that someone was proposing to use to predict some new
disease, what would I expect to use to evaluate that?" And that's
probably what I would start with.


Dr. Greg Hundley:


Excellent. Well, I'm going to turn back and go back to our
panelists here, listeners. And we're going to ask each of our
panelists in about 20 seconds to describe for us what they think
is the next most important aspect of research in their respective
areas. So first I'll start with Mike Ackerman. Mike, can you tell
us what's coming next in this area of assessment of QT
prolongation or other aspects of the electrocardiogram?


Dr. Michael Ackerman:


I think next is implementing this in the real world. We are
having our suite of the AI ECG as a  hypertrophic
cardiomyopathy detector. We've shown that as an ejection fraction
detector, and now as a QT detector in AFib, from our work and
Chris's work. And for the QT itself, I think where we are is
we're really, really close to now having a mobile enabled digital
QT meter. And a digital QT meter, once FDA cleared, then allows
the QTC to truly emerge as the next vital sign. And it really
deserves to be a vital sign. We use it as a vital sign. We know I
want to know my patient's QTC every bit as I want to know his or
her weight, blood pressure, saturation. It's an actionable
finding, and we're now getting really close. We're just on the
cusp of having a true digital QT meter.


Dr. Greg Hundley:


Excellent. Chris?


Dr. Christopher Haggerty:


I think for us to, in part address some of the comments that Mike
brought up about the reproducibility of these types of models,
we're very keen to demonstrate the prospective capabilities of
our models to enroll patients in a prospective fashion, run their
ECG through our predictor, and then screen them for AFib to
determine how well we actually do moving forward, instead of just
relying solely on our retrospective data. So we're very excited
to do that. We're ramping up for that trial now and hope to be
able to demonstrate similarly positive findings from our
technique.


Dr. Greg Hundley:


Great. How about you, Nick?


Nick Mills:


I'd like to see the same quality and rigor applied to the
implementation of these technologies as we have to other
important areas in cardiovascular medicine. I think that's a
really important step, not just to develop the tools, but to
demonstrate their value. But I also think what we've done so far
is relatively simplistic. We've taken an ECG and we've ignored
almost all the other information that we have in front of us. And
as these algorithms are trained and evolved, these and other
vital clinical biomarkers and information, and integrating them
into these neural networks will really enhance their performance
for predicting things that are less tangible, like sudden death
in the future or stroke.


Dr. Greg Hundley:


And then finally, Mike Rosenberg.


Dr. Michael Rosenberg:


Yeah, I actually see two challenging areas in this field. One is
the access to data. And I think one of the things that a lot of
companies are realizing is that even if they make hardware, that
the data may be more valuable than the technology that they're
getting the data from. So I think one is figuring out ways to get
access to data so that people can reproduce findings from these
studies. And the second is deliverable. A bottle like this is not
like the CHADS-VASc score that I can calculate in my head in the
clinic. I mean I need a way to actually run these models within
an EHR, within a computer system like that. And I think it's
going to be a big challenge to take a model like this and to
deploy it at scale the way we would with the drug or any other
innovation.


Dr. Greg Hundley:


Fantastic. Well listeners, we want to thank Mike Ackerman from
Mayo Clinic, Chris Haggerty from Geisinger, Mike Rosenberg from
University of Colorado, and Nick Mills from University of
Edinburgh for really providing us with a wonderful discussion
regarding the use of machine learning applications in one study
to predict the QTC interval from two leads that may be applicable
to wearable devices. And in the second study, predicting the
future occurrence of atrial fibrillation and even stroke as an
adverse event in people at risk.


Dr. Greg Hundley:


On behalf of both Carolyn and myself, I want to wish you a great
week and we will catch you next week on the run. This program is
copyright of the American Heart Association, 2021.