The Future of AI in Healthcare Insights from George Dorffner

Transcript

00:00:06 Carita
Welcome to the E Health 24 podcast, where we discuss the digitalization of healthcare and
the possibilities of technology.

Today, we have the pleasure of interviewing a Georg Dorffner Ms. PhD.
Associate professor at the Institute of Artificial Intelligence Center of Medical Data
Science at Medical University of Vienna, Georg is also one of the vice presidents of the
Austrian Society of Artificial Intelligence.
He specializes in medical predictions models and signal processing, particularly from
the machine learning perspective.

Welcome, Georg.

00:00:50 Georg
Thank you.

00:00:51 Carita
Could you tell us a bit about yourself and your career? How did she come to work in AI
research and medical data science?

00:01:00 Georg
Sure. First of all, thanks for the invitation. Yeah, it dates back actually to my time as
a student. I studied computer science and electrical engineering in Vienna and was very
much also interested in language. I did a little linguistics on the side and then I came
across that department for Artificial intelligence, the same institute that I work for
now, and they offered courses in natural language processing and so that's how I got
involved.

Then I went to the US, did my PhD at Indiana University and there I learned about neural
networks. This is the method or the prime method of machine learning today and I've
really got fascinated and did work on them ever since.

00:01:38 Carita
Fascinating. How did you see the role of AI in healthcare and machines today? Do you
think it offers the greatest strengths and opportunities?

00:01:49 Georg
Well, I think AI is becoming more and more an important tool in health care. I mean it
provides capabilities of interpreting data, recognizing something in recordings or in
images for instance, and also with the latest development in large language models, it
offers systems that at least implicitly have a large amount of knowledge, medical
knowledge. I mean, they can be trained on practically the entire medical literature that
exists and can use that in reformulating medical answers.

So think already today it plays an important role, like in radiology everywhere where
images or signals play a large role. That's where the real major breakthroughs happen,
and it will, I mean, it will certainly could it's here to stay. I have to say so future
health care without the use of AI is probably not possible.

00:02:43 Carita
Your research focuses on machine learning in the development of medical prediction
models. Could you explain how machine learning particularly works for example in
improving patient's prognosis.

00:03:00 Georg
Sure, in prediction it is, this sense is special type of classification. You have data of
a patient and you want to predict whether that patient will survive in a certain
situation or develop a certain disease or be treated well, healed from the disease, and
machine learning is basically it's an algorithm that provides a lot of values that can be
a lot of so-called variables. You can imagine like little screws, you can fine tune in
order to change the behavior of the system and the fine tuning happens in the so-called
learning phase and you present the data so you present all the patient data and what you
also need to know is whether that patient has that certain disease or has developed that
disease, so you need so-called labeled data, somebody needs to say whether that's a
pathological case or a healthy case and with that feedback, whether it's one or the
other, those little screws, those little values get fine-tuned that you have to repeat
that very often and then in the end you have a trained system that you can apply to new
data and make valid predictions.

00:04:08 Carita
When it comes to prediction models and their application in patient care, how confident
can we be in the results they provide? What challenges are involved in ensuring their
reliability.

00:04:23 Georg
Well, I mean, we can be become confident by validating the system. So every   such system
after it was trained on data needs to be applied to new unseen data, data that it has
never seen and then we evaluate statistically the performance how well it classifies. One

has to keep in mind it's never perfect. No, no AI software will recognize a disease in
100% of the cases, but neither do the doctors.

So, we to be confident that the AI system is doing the right thing we usually compare the
performance like say 95% of the cases are classified correctly with that of experts, and
if that's also, say 95% of the same level, we can say we have validated to be as good as
an expert.

00:05:09 Carita
There are many ethical questions related to AI, especially concerning patient privacy and
the transparency of AI-driven decision making. How do you view these ethical challenges
and what do you think is the best way to address them?

00:05:30 Georg
Well, certainly we need to address them but there's more than what you just mentioned. I
mean, privacy, of course, is one of them but that only that not doesn't directly concern
the AI. It concerns whether and how we collect patient data in any case. So if it's
usually stored in a hospital and we as machine learners would like to access it, we need
to be authorized to read that data and if that data is not protected enough theoretically
some unauthorized person could read it, but I think today computer security is good
enough to keep the data protected.

Transparency is in the sense of I mean, these neural networks, these machine learning
methods are often called the black box because we don't know how it works internally. But
then I always say well a doctor is also not very transparent. If I ask a dermatologist,
why is this a melanoma this this thing on my skin. He might also not perfectly be able to
explain it and might say well, I've seen so many I just know this is one. So and with the
latest development in language models, I think we can train those systems in future also
to explain themselves.

But what's other challenges that we need to address or, for instance, are something like
bias. So the systems can or are biased in a sense that if it's seen more data from one
gender, for instance inspired toward that, or from one ethnical group so typically like
in dermatology systems that are trained on mostly on white Europeans and then they will
not perform as well on Asians or Africans, and we just need to be aware of that and need
to make sure that the system is applied properly and or that we collect more data to
count the balance that bias. That's a challenge to face.

00:07:08 Carita
Interesting. What do you see as the next big step for AI in healthcare Where do you think
the field will be in say 5 or 10 years?

00:07:21 Georg
Well, such predictions are almost impossible because nobody would have predicted what AIs
today, three or four years ago, like these large language models that I keep referring to
surprised us all the with respect to their capabilities. But on the other hand, they are
still only trained on language, so they don't really understand what they're saying.
They've reproduced the knowledge, but they cannot associate it to appearance as we can
and now the big next step will be to enrich those models with like images and other
modalities.

So they will become more and more Intelligent and they will be interesting to see how
much of a role they will take over. I mean we should never think of AI ever replacing
doctors. But they will be able to take over more and more of the role of a physician
today in diagnosing something, but in communicating to the patients, I think the human
element is still the important one and we need to see to make sure that this will happen
that not due to reasons of costs, patients can no longer talk to human, only to the AI so
we'll have to make sure.

So that's, that's where I think in 5-10 years I think, AI will play a very prominent
role, but hopefully we will have made sure that the human in the loop is as important as
it is today.

00:08:39 Carita
I think a lot of patients will appreciate that too that they see a human, not only the AI.

00:08:42 Georg
Yes. Right, exactly.

00:08:47 Carita
What advice would you give to young researchers and students who are interested in AI and
it's applications in healthcare?

00:08:57 Georg
Well, I would say if you're interested in that area, the advice would be to always listen
to the medical professionalists. I mean, I see too often that computer scientists develop
new algorithm and they use the medical data just as their playground is, OK, I can
improve my algorithm, but they don't look enough at whether that's really clinically
relevant.

So I made very good experience with the constant exchange between technician engineers
like computer scientists, myself and the physicians, so that if you focus on that early
on, I think your research will be very fruitful.

00:09:32 Carita
Thank you so much Georg, for joining us on our podcast and sharing your expertise and
insights into the role of AI in healthcare. This discussion has given our listeners
plenty to think about. Thank you, Georg.

00:09:51 Georg
Well, you're welcome. Thank you for letting me be here.

00:09:53 Carita
Thank you. Bye.

00:09:54 Georg
Bye bye.