Latest / Becker’s Healthcare Digital Health + Health IT / Turgay Ayer, Professor at Georgia Tech’s School of Industrial and Systems Engineering

Transcript

Turgay Ayer, Professor at Georgia Tech’s School of Industrial and Systems Engineering · Becker’s Healthcare Digital Health + Health IT

0:00
Exciting things are happening at Becker's Healthcare. Stay
0:04
ahead of industry trends with the new Becker's
0:06
CFO plus Revenue Cycle podcast,
0:08
your go to source for insights from top
0:11
healthcare finance leaders. Tune in wherever you get
0:14
your podcasts.
0:15
And don't miss the tenth annual Health IT
0:18
plus Digital Health plus RCM conference, happening September
0:22
30 to 10/03/2025
0:24
in Chicago.
0:25
Join thousands of executives,
0:27
engage with industry leaders, and explore the future
0:30
of health care innovation.
0:32
Learn more about our upcoming events at beckershospitalreview.com.
0:36
See you there.
0:38
This is Laura Dierda with the Becker's Healthcare
0:40
podcast. I'm thrilled today to be joined by
0:42
Tur Guy Ayer, professor in the school of
0:45
industrial and systems engineering at Georgia Tech. Tur
0:48
Guy, it's a pleasure to have you on
0:49
the podcast today.
0:50
It's my pleasure. Thank you for having me.
0:53
Now I'm looking forward to our conversation and
0:55
really digging into some of the cool things
0:56
you're doing, within your research at Georgia Tech.
0:59
But before we dive in, can you tell
1:00
us a little bit more about yourself and
1:01
your background?
1:03
Absolutely.
1:04
So I was trained as a decision scientist.
1:08
I did my t PhD
1:10
thesis on breast cancer diagnosis using artificial neural
1:15
networks.
1:17
And I work extensively with radiologists and some
1:19
other physicians. Even though I was trained as
1:21
an engineer,
1:23
I was trained almost like a pseudo doctor.
1:26
And as for my background, I was raised
1:27
to be a physician by my parents. They
1:29
really wanted to want wanted me to be
1:32
a physician.
1:33
And up until
1:34
junior high, I was thinking of being a
1:37
physician.
1:38
But my sister is a veteran medicine,
1:41
doctor, and and I was helping her when
1:43
I was in high school, and I fainted
1:46
as soon as I saw some blood.
1:48
And it made it obvious that I will
1:50
make a good physician. So I decided to
1:52
study engineering
1:54
because I was good at math. I I
1:55
liked, you know, analytical thinking.
1:58
And I had a lot of, you know,
2:01
I I I had a lot of trials
2:03
and errors, in my in my studies. I
2:06
start as a double e,
2:08
major and then switched to computer science
2:11
in my junior year
2:13
in college.
2:14
And only my senior year, I,
2:17
figured out a discipline called operations research within
2:20
industrial engineering,
2:22
which is a science of
2:23
decision making through mathematical modeling and and data
2:27
driven approaches.
2:28
So I ended up majoring in operations research
2:31
and did a PhD
2:33
in operations research slash industrial engineering focusing primarily
2:37
on on health care applications.
2:41
Fascinating. That's, you know, such a a cool
2:43
journey into the career you have today and
2:45
and really meaningful research you're doing. Can you
2:47
share a bit about that research and why
2:49
it's important?
2:51
Absolutely.
2:52
So I would summarize my research as,
2:55
data driven mathematical
2:57
slash computational modeling for important medical decision making
3:00
and health policy analysis.
3:02
Oftentimes, in clinical decision making
3:05
and health policy making, decisions are too complex,
3:09
and you need a framework for the systematic
3:12
decision making
3:13
while incorporating
3:14
data to capture important analysis.
3:18
And and such decisions become too complex for
3:20
human beings. To me, it's more like making
3:22
multiplications on the order of millions.
3:25
When you ask such a question to a
3:26
computer, it's pretty straightforward.
3:28
But as humans, if we try to do
3:30
these computations in our head, it becomes pretty
3:32
challenging.
3:33
So I help that building
3:36
frameworks, mathematical
3:38
computational frameworks
3:40
using utilizing data,
3:42
to help it solving such important,
3:44
pressing clinical and policy problems.
3:50
That's fascinating to hear. And, you know, really,
3:52
such a unique angle to be part of
3:54
the health care space as you mentioned using
3:56
your mathematics background and certainly,
3:59
you know, your your talent for,
4:02
computation to get this into,
4:04
the health care field.
4:06
Now I'm curious. So what challenges in health
4:08
care are you aiming to address through this
4:10
research?
4:11
Absolutely.
4:12
So in health care, the goals health care
4:16
research, the gold standard is randomized controlled trials.
4:20
In randomized controlled trials,
4:23
researchers do AB testing, meaning that, you know,
4:26
if there is a new intervention, new drug,
4:28
new technology,
4:30
clinical trials are designed to to assess if
4:33
the new technology, new intervention, let's say, a,
4:36
is better than the status quo b.
4:39
And in such randomized controlled trials, they are
4:42
typically expensive, very expensive, millions of dollars. They
4:45
are time consuming.
4:46
And and also it takes a lot of
4:49
efforts,
4:50
to conduct these clinical trials, recruit patients into
4:53
those trials. And typically, they are short term.
4:56
Typically, you know, the trials
4:58
last for about three to five years.
5:01
And
5:03
by the time the trials are completed, we
5:06
have answers to the question of whether a
5:08
is better than b. But oftentimes, in in
5:11
in practice, we are looking at multiple options,
5:13
not just a versus b. We are looking
5:15
at a versus b versus c versus d.
5:18
And and sometimes these competitors are more than
5:21
ten, fifteen, 20.
5:23
And it's not plausible to
5:25
assess those mini interventions in a clinical trial
5:28
setting because, a, it will be extremely,
5:32
costly,
5:33
and, b, it will be very time consuming.
5:36
And, c, it perhaps wouldn't be feasible to
5:39
recruit,
5:41
thousands of patients into such a trial.
5:45
Because and and the last limitation of randomized
5:47
controlled trials is that, as I mentioned earlier,
5:50
their time span is typically short, and oftentimes,
5:53
you would like to understand longer term implications
5:55
of interventions.
5:57
And at that point,
5:58
our research chimes in and and helps with
6:01
understanding some longer impacts
6:03
of such
6:04
interventions.
6:05
We collect evidence. We basically analyze evidence coming
6:08
out of randomized controlled trials, published studies, meta
6:11
analysis, real world evidence,
6:13
and build virtual trials,
6:15
meaning that, you know, we replicate
6:18
these trials in computer settings.
6:21
We validate these computational models.
6:24
And once validated that, they they they become
6:26
a test bed for us to ask a
6:28
lot of interesting what if questions
6:31
in the short term. So we are no
6:32
longer limited to three years or five years.
6:35
We can expand the time horizon to twenty
6:38
years, thirty years lifetime.
6:39
We are no longer limited to, you know,
6:42
a versus b. We can compare multiple,
6:45
interventions.
6:46
And and that gives us a lot of
6:48
flexibility to, again, ask a lot of interesting,
6:51
pressing policy or clinical level questions.
6:57
That's amazing to hear. You know? And certainly,
6:59
really cool to see that evolution of how
7:01
you can approach clinical trials in a really
7:04
meaningful way, especially
7:05
knowing how important they are in advancing medicine
7:08
and the treatment that is provided for patients
7:10
across the board in a variety of, settings,
7:13
as well as, you know, just the the
7:16
demographics of the patients can be so different.
7:19
How do you approach conducting research to ensure
7:21
that it has the kind of real world
7:23
impact on patient care? I know a lot
7:24
of the things we've been talking about,
7:26
you know, are looking at the design of
7:28
the trials and those kinds of things. And
7:29
so how do you see that impact?
7:32
So there are a couple of ways that
7:33
we ensure that our findings are practically useful.
7:37
One, we
7:39
collaborate extensively with it, end end users, stakeholders,
7:43
typically physicians, policymakers,
7:45
depending on the nature of the project, and
7:48
that is useful for multiple purposes. One, we
7:51
ensure that we capture important clinical nuances,
7:55
medical nuances that we may not be aware
7:57
of. So we have the clinical slash policy
8:00
expertise.
8:02
And two,
8:03
because it's a co collaborative
8:06
development, collaborative research,
8:08
those decision makers, end users are engaged early
8:12
on in the process,
8:13
and we deploy our solutions to those end
8:16
users.
8:17
And once they believe in the value, once
8:19
they see what these solutions tools could do
8:21
for them, they are typically the champions to
8:24
take take these solutions to the next step
8:27
and, you know, bring it to clinical implementation
8:30
or use it as a policy policy tool,
8:32
etcetera.
8:33
And the last thing that we do is
8:35
we ensure that our findings
8:37
are valid.
8:38
Before we deploy any solutions,
8:41
we do a lot of stress testing
8:43
compared against any available
8:45
real world data.
8:48
So it's
8:49
just to ensure that, again, the results are
8:51
clinically meaningful, and we also have the binds
8:53
from these end users.
8:58
That's helpful to hear. Thank you so much
9:00
for digging a little bit deeper there. Now
9:02
I'm curious. What are some of the big
9:04
obstacles that you face while you're,
9:06
doing this type of research and trying to
9:08
make sure that it's deployed in a meaningful
9:10
way?
9:12
Data challenges
9:13
is is a big issue. Oftentimes, health care
9:16
data is fragmented,
9:18
and they are kept in silos.
9:21
You don't oftentimes, you don't have national level
9:25
databases,
9:25
access to that data. You know, even even
9:28
if it is local level data may be
9:30
a challenge. We have to go through a
9:31
little bit pool processes, and
9:34
and in the end, it may not work
9:35
out. So as an example, you know, we
9:37
have recently approved
9:39
to have access to a data
9:41
all by,
9:43
kept by CDC.
9:44
And because of these communication freezes CDC is
9:47
going through, we basically even though our our
9:51
data access request has been approved after several
9:53
months of back and forth communications
9:56
and documentation, you know, report preparation, etcetera,
10:00
basically, this is this is on freeze now.
10:02
We won't be able to access this data
10:03
until this uncertainty is resolved.
10:07
So accessing the data,
10:09
compiling, augmenting datasets, and and also
10:13
validating, you know, these
10:15
including missing data, validating these datasets is is
10:19
one of the challenges. And and,
10:21
sometimes communication
10:23
gaps with the end users is a challenge.
10:27
Yeah. I would I would list these two
10:29
as the biggest challenges.
10:33
That's helpful to know. And, you know, certainly,
10:35
it seems like,
10:36
having that reliable data is a challenge for
10:38
a lot of health care organizations, as you
10:40
said, and and really truly something that, can
10:43
make a big difference,
10:45
within
10:45
the the research and development in,
10:48
many fronts, but but certainly with the work
10:50
that you're doing as well. I I think,
10:53
you know, when you look at that communication
10:56
aspect of it,
10:57
too, you know, is there anything you found
10:59
that has worked particularly well in,
11:02
communicating
11:03
some of the research and the data, to
11:05
those who, you know,
11:07
need to know about it or or would
11:09
most benefit from it,
11:11
or is that still very much a work
11:13
in progress?
11:15
Absolutely. So in one of our research projects,
11:18
we work with radiologists,
11:20
breast imaging radiologists,
11:24
from from one of the hospitals in Wisconsin.
11:27
And we were basically the key question research
11:30
question we were asking was whether
11:33
machine learning models, artificial intelligence models was useful
11:37
in diagnosing breast cancer based on mammographic images.
11:41
So in other words, radiologists look at suspicious
11:44
mammograms,
11:45
or or they look at a set of
11:47
mammograms if if and if they notice any
11:49
suspicion such as, you know, some some mass
11:53
some messes, calcifications,
11:55
distortions,
11:56
asymmetry, etcetera,
11:58
then they may refer them to a biopsy,
12:02
to
12:03
see if if patients have a tumor, a
12:05
malignant tumor, and and they need breast cancer
12:08
treatment.
12:09
And it turns out that this decision making
12:11
process is is very difficult and in more
12:14
process. There are, like, you know, over 50
12:16
risk factors that physicians need to consider.
12:19
And
12:20
and you need to again, to me, this
12:22
is like making multiplications on the order of
12:24
millions. And it turns out that about ninety
12:27
eight percent of the biopsy cases
12:29
are, false positives, meaning that out of ninety
12:32
eight percent of all mammograms referred to biopsies,
12:36
they are in reality, you know, just okay
12:38
mammograms. Right? Nothing is wrong with them. But
12:40
think about the psychology of a patient.
12:44
Right, when you tell them that they have
12:45
a suspicious mammogram and they need to schedule
12:47
a biopsy,
12:48
those biopsies are typically not scheduled for tomorrow.
12:51
Right? You need to wait for two, three,
12:52
four weeks, sometimes months,
12:55
to get these biopsies done
12:58
and on top of it further wait to
13:00
hear about the results. So you go through
13:02
a lot of stress. Right? You know, you
13:04
you go through a lot of stress. Your
13:05
your family go goes through a lot of
13:07
stress.
13:09
And what we did is
13:11
we basically,
13:12
built a an artificial neural network model,
13:16
and we asked the key for the the
13:19
following key question. Right? Could an artificial
13:21
neural network model, a form of an, artificial
13:24
intelligence, machine learning models,
13:27
help with biopsy decision making
13:30
using findings or extracting data from mammographic images.
13:35
And it turned out that and and and
13:37
our results we published our results in in
13:39
a clinical journal, radiologist journal. We thought, you
13:42
know, physicians, radiologists
13:44
would be reacting to those findings because our
13:46
findings indicated that
13:48
our machine learning based solutions
13:51
could perform
13:53
better than majority of the radiologists
13:55
in our setting. Specifically,
13:58
seven out
13:59
of eight radiologists,
14:00
seven of them performed worse than the artificial
14:03
neural network model. So meaning that a AI
14:06
model
14:07
was more successful
14:08
in detecting abnormalities on mammographic images as compared
14:12
with those seven out of eight radiologists.
14:15
Whereas one radiologist performed significantly
14:17
better than the AI model.
14:20
That's
14:21
that radiologist who stood out had thirty two
14:23
years of domain experience, whereas the other seven
14:26
radiologists
14:27
who are outperformed by the AI model had
14:30
an experience ranging from two to ten years.
14:32
And we positioned our paper in a way
14:34
that we we we we basically,
14:36
you know, highlighted the key message that
14:39
AI based solutions may complement physicians, especially junior
14:43
physicians, to help it, you know, to to
14:45
act as a second reviewer or a second
14:48
reader,
14:49
especially in the resource limited settings.
14:52
The doubles double screening in radiology is very
14:54
common, meaning that, you know, a radiologist first
14:57
reads an image. And if the radiologist is
14:59
is unsure,
15:00
he or she consults with a second radiologist,
15:03
and then,
15:04
they discuss and make a decision together. But
15:07
in the resource limited settings, that's not feasible.
15:09
And we positioned our paper study in a
15:12
way that we we emphasized that in resource
15:15
limited settings,
15:16
AI based solutions could act as a second
15:18
screener, second reviewer.
15:21
And
15:23
to our surprise, pleasantly, our findings are very
15:26
well received by by the radiology community. Right?
15:29
Radiology community is not trained
15:31
as as AI experts.
15:33
But even though, you know, the findings
15:36
could have sort of, like, to to to
15:38
some extent threaten their their their their discipline,
15:41
they were very positive, very enthusiastic in embracing
15:44
our findings. And we they they ended up
15:47
accepting our paper,
15:49
to get published in in one of their
15:51
flagship journals in their communities.
15:53
This is just one example.
15:55
I can think of several such examples where
15:57
our findings were very well received by the,
16:00
by by the corresponding communities.
16:04
That's a really, powerful example, and thank you
16:07
so much for, explaining,
16:09
deeper there and,
16:10
telling us a story because I think, you
16:12
know, it's it's really cool to have that,
16:14
frame of reference.
16:16
Now have there been any surprising or unexpected
16:18
findings that you've had within your research?
16:21
Yes. Many of them. I'll just mention maybe
16:24
one or two.
16:26
In one project, we were looking at, hepatitis
16:29
c elimination.
16:30
Hepatitis c is an infectious disease, and it's
16:34
a prevalent disease. As as a matter of
16:36
fact, as of 02/2006,
16:38
the number of people dying from hepatitis c
16:41
in The US has surpassed those dying from
16:44
HIV.
16:45
Right? So when I when we think about
16:47
HIV, I don't think we need to make
16:49
a case that HIV,
16:50
has a has a big burden on the
16:52
society.
16:54
And it turns out that as of as
16:56
of, you know, as of 02/2006,
16:58
hepatitis c has even a bigger burden on
17:01
the society.
17:03
And
17:04
because it has a such a so it
17:06
has such a big disease burden, I was
17:08
I was interested in, you know,
17:11
doing some modeling analysis in that space. I
17:14
elaborated with epidemiologists and physicians and decision scientists
17:18
in this work,
17:19
from Emery Rowland School of Public Health, Harvard
17:22
Medical School, among others.
17:24
And the key research question in one of
17:26
the projects, the key research question we were
17:28
asking was the following.
17:30
Does it make sense to treat hepatitis c
17:33
in correctional facilities? Does it make sense to
17:36
provide access to vast access to hepatitis c
17:39
treatment in correctional facilities,
17:41
or should it be limited in some ways?
17:44
And the reason why this question was relevant
17:46
is that by the time
17:48
the initial hepatitis C treating drugs, direct acting
17:51
agents came to market around early 2000s,
17:54
many of the correctional health facilities were limiting
17:57
access to hepatitis C treatment,
18:00
because hepatitis c treatment
18:01
by then and and even still,
18:04
was was quite expensive.
18:06
By the time they came to market,
18:08
the single
18:10
pill cost per day was about $1,000
18:13
per pill, and patients were expected to be
18:15
on treatment for twelve weeks. So meaning that,
18:17
you know, by the time the treatment is
18:18
completed,
18:19
the per patient cost would be about $84,000.
18:24
And even though correctional facilities receive some discounts,
18:28
even after discounts,
18:29
a study has shown that
18:32
if correctional facilities provide a treatment to all
18:35
hepatitis c infected patients,
18:37
the total cost would be higher than the
18:40
entire health care spending budget allocated for correctional
18:43
facilities.
18:45
And and and inmates are not suffering only
18:47
from hepatitis c. Right? They could be infected
18:50
with with HIV. They may have chronic,
18:53
diseases such as cardiovascular
18:54
disease, diabetes, etcetera.
18:56
From a resource allocation perspective,
18:58
it it it was very clear that
19:01
correctional health systems wouldn't be able to provide
19:03
universal access to all inmates. And it was
19:06
a very controversial topic. As a matter of
19:08
fact, many prison many, many inmates are suing
19:10
state prison systems,
19:12
for for for not providing them access to
19:15
hepatitis c treatment.
19:17
And our key research question was the following.
19:19
As I mentioned, I'll be repeating myself. Does
19:21
it make sense to expand health care spending
19:24
for correctional facilities
19:27
such that
19:28
access to hepatitis c treatment could be expanded,
19:32
for inmates behind the bars.
19:34
And, this was a politically controversial question because,
19:38
you know, because you know, I mean, I
19:39
don't wanna get into too much into into
19:41
political discussion. But as you may imagine, some
19:43
some politicians are arguing that, you know, there
19:46
is no point of further expanding health care
19:48
spending,
19:49
for correctional facilities.
19:52
And what we did in our in in
19:54
our analysis is we have taken a radical
19:55
approach. Right? We have asked the following question.
19:57
Right?
19:58
We have taken a radical approach and asked,
20:01
what happens if we expand
20:03
treatment to hepatitis c
20:06
in correctional facilities from a societal perspective
20:09
for the for the for the community, right,
20:12
outside the bars.
20:14
And we have shown that
20:16
it makes
20:18
a lot of
20:19
economic sense to treat hepatitis c
20:22
while,
20:24
in the correctional facilities because the idea is
20:26
the average time spent in behind the virus
20:29
is less than five years.
20:30
Correctional facilities provide an ideal setting for identifying,
20:36
infected inmates,
20:38
treating them and linking them to care. Outside,
20:40
you know,
20:42
in the outside community, it's very harder to,
20:45
diagnose patients. The prevalence is very lower. And
20:48
also linking patients to care is more difficult.
20:51
In correctional settings, it's almost like a laboratory
20:54
environment, so linkage to care is way higher.
20:57
And through, again, validated
21:00
mathematical and computational models, we have shown that
21:03
from a resource allocation standpoint, if you are
21:06
going to spend as a society, if you'll
21:08
spend any resources on hepatitis c treatment and
21:11
elimination,
21:12
it makes a lot of sense,
21:14
rational economic sense to increase spending
21:18
for for
21:19
for reducing hepatitis c infections and treating hepatitis
21:23
c,
21:24
in the prison settings, community,
21:26
in the in the correctional health settings, again,
21:29
because of this intermingling happening upon release of
21:32
these inmates.
21:34
This was somewhat a contrary to finding in
21:36
the sense that
21:38
on one hand, you would think that, you
21:40
know, it may not make sense to further
21:42
increase health care spending for correctional stand you
21:45
know, settings. Again, this is not my standing,
21:47
but there are people who believe that way.
21:50
On the other hand, we have taken this
21:52
radical extreme approach and have shown that even
21:55
if you
21:56
don't care about the lives of these inmates,
21:58
which is obviously
21:59
unethical and doesn't make sense, but even if
22:01
you have taken such an extreme approach,
22:04
it still makes rational
22:05
economic sense to prioritize,
22:08
treatment access
22:09
in the correctional settings from a societal perspective,
22:12
because return on investment
22:14
in such cases is higher because, you know,
22:16
detecting those patients and treating these patients and
22:19
linking them to care is easier in correctional
22:22
settings.
22:24
So this is one example. I can I
22:26
can keep going? I can share more examples
22:28
if you'd like.
22:29
If I can share a second one and
22:31
stop that.
22:32
A second example is is in another project,
22:35
we looked at,
22:38
Medicare Advantage Plans. Medicare Advantage Plans are basically,
22:42
you know, Medicare plans
22:44
offered by private insurances. What what what happens
22:47
is Medicare sort of outsources
22:49
the provision health care of health care services
22:52
to to private entities, such as UnitedHealthcare may
22:55
be providing coverage for Medicare patients, and it's
22:58
part it becomes part of the Medicare Advantage
23:00
Plans.
23:02
And it is well established that it is
23:04
somewhat well understood in the published literature that,
23:07
Medicare Advantage Plans
23:10
tend to do risk selection, meaning that they
23:12
select profitable patients,
23:15
through some selection mechanisms.
23:19
And the well under there is this well
23:21
understood notion that, you know, they may be
23:23
spending
23:24
money
23:25
allocated
23:26
for high risk patients towards treatment of sorry.
23:29
Money allocated
23:31
for low risk patients towards treatments of high
23:33
risk patients, and it's referred to as cross
23:36
subsidization.
23:38
In in one of our research papers, we
23:40
have shown that
23:42
the other way around is also correct, meaning
23:44
that medic Medicare Advantage Plans tend to utilize
23:47
resources allocated for high risk patients
23:52
to affect low risk patients into their plans,
23:55
you know, and through mechanisms such as such
23:57
as,
23:58
reduced gym memberships. Right? You know, some Medicare
24:01
Advantage Plans offer reduced gym memberships.
24:03
And if you think about reduced gym memberships,
24:05
it sounds great. Right? You know, you would
24:07
think that, you know, this is a great
24:10
add on perk that they're offering, but when
24:12
you carefully think about it,
24:15
carefully think about who would be interested in,
24:18
gym membership
24:20
among elderly population. Right? We are talking about
24:22
Medicare patients, patients over age 65.
24:26
An individual suffering from multiple chronic conditions
24:29
and not more you know,
24:31
not easy to you know, not having a
24:33
hard time with movements, etcetera,
24:36
wouldn't be so much interested
24:38
in gym membership.
24:39
But healthy physically healthy,
24:42
physically active,
24:43
yeah, older individuals,
24:45
would be more attracted to gym membership, reduced
24:48
gym memberships.
24:49
Even though at first, it sounds like a
24:51
great perk,
24:52
in reality,
24:53
it is a selection mechanism. Right? It is
24:55
a it is a mechanism primarily designed for
24:58
attracting
25:00
relatively healthy healthier older adults into their population.
25:04
And we have shown that they are using
25:06
basically these resources allocated for these relatively healthier,
25:11
individuals,
25:12
towards treatment of sicker patients.
25:18
Wow. That's fascinating to hear. You know? And
25:20
really, really,
25:22
interesting to understand the applications of your research
25:25
and dig a little bit deeper here. Jorge,
25:27
thank you so much for joining us on
25:28
the podcast today. I I really appreciate your
25:30
time, and I look forward to, connecting with
25:33
you again soon. Thank you so much for
25:35
the opportunity.