Name: Navigating the digital healthcare frontier
Uploaded: 2023-10-31T23:47:51.370Z
Duration: 388.45
Description: Join us for an enlightening interview with Professor Ian Scott, Director of Internal Medicine and Clinical Epidemiology at Princess Alexandra Hospital, as he discusses the challenges and immense potential of AI in healthcare.

P.mp4

P.vtt

Subtitles

{{00:00:06}}
Well, I think there's a couple

{{00:00:07}}
of major challenges with AI.

{{00:00:08}}
First of all, we've got

{{00:00:09}}
to make sure that the data

{{00:00:10}}
that's used to train AI

{{00:00:11}}
is, is high quality data.

{{00:00:13}}
It's accurate, it's complete,

{{00:00:15}}
and it's representative.

{{00:00:16}}
So there's no hidden

{{00:00:17}}
biases in the data.

{{00:00:38}}
Well, I think the first

{{00:00:40}}
there's a few things

{{00:00:41}}
we need to think about,

{{00:00:42}}
first of all, and that is.

{{00:00:44}}
Clinicians, will accept

{{00:00:45}}
AI if they think it's

{{00:00:47}}
serving a useful purpose.

{{00:00:49}}
In other words, it's solving

{{00:00:50}}
a problem that they grapple

{{00:00:52}}
with on a daily basis.

{{00:00:53}}
So if the AI can be shown to

{{00:00:54}}
improve quality of decision

{{00:00:56}}
making, and therefore quality

{{00:00:58}}
of care, then clinicians

{{00:01:00}}
will be happy to use it.

{{00:01:02}}
So I think  in allowing

{{00:01:04}}
clinicians to develop

{{00:01:04}}
that trust in AI, we need

{{00:01:06}}
to make sure that the AI

{{00:01:07}}
does work effectively.

{{00:01:09}}
It's been proven in good

{{00:01:10}}
clinical trials that

{{00:01:11}}
it's effective, it's

{{00:01:12}}
safe, and it's reliable.

{{00:01:14}}
And that it's

{{00:01:14}}
also easy to use.

{{00:01:16}}
So in other words, it

{{00:01:16}}
doesn't disrupt clinical

{{00:01:18}}
workflows, it doesn't

{{00:01:19}}
make for inefficiencies in

{{00:01:20}}
how people do their work.

{{00:01:22}}
So if AI can satisfy those

{{00:01:24}}
two key principles, then

{{00:01:25}}
I think clinicians will be

{{00:01:26}}
more than happy to accept it.

{{00:01:28}}
I think it really starts

{{00:01:29}}
from early training.

{{00:01:30}}
So I think it's important

{{00:01:31}}
that, for example, in

{{00:01:32}}
the  undergraduate degrees

{{00:01:34}}
when medical students, for

{{00:01:35}}
example, have been trained,

{{00:01:36}}
that  they're exposed to AI

{{00:01:38}}
and they become familiar with

{{00:01:39}}
AI concepts, understand data

{{00:01:42}}
and data science, and actually

{{00:01:44}}
have an understanding of how

{{00:01:45}}
these things integrate with

{{00:01:46}}
electronic medical records and

{{00:01:47}}
other digital technologies.

{{00:01:49}}
And that then needs to be

{{00:01:50}}
carried through right through

{{00:01:51}}
to postgraduate training,

{{00:01:53}}
indeed, in even into advanced

{{00:01:55}}
training for specialties.

{{00:01:56}}
So I think that

{{00:01:57}}
this is happening.

{{00:01:58}}
I think colleges now

{{00:01:59}}
are grappling with this.

{{00:02:00}}
I think universities

{{00:02:00}}
are grappling with this.

{{00:02:02}}
So I think over the coming

{{00:02:03}}
years, there'll be a more

{{00:02:04}}
literate workforce in relation

{{00:02:06}}
to AI usage and how and

{{00:02:08}}
how to use it effectively.

{{00:02:10}}
AI can't compromise

{{00:02:11}}
the clinician

{{00:02:11}}
patient relationship.

{{00:02:12}}
So at the end of

{{00:02:13}}
the day, I see AI as

{{00:02:14}}
being like a copilot.

{{00:02:16}}
But it's not going to do

{{00:02:17}}
away with the human empathy

{{00:02:19}}
and the human communication

{{00:02:21}}
and the personalized shared

{{00:02:22}}
decision making that between

{{00:02:24}}
clinician and patient decide

{{00:02:26}}
what's best care for that

{{00:02:26}}
particular individual.

{{00:02:27}}
So AI assists, but it

{{00:02:29}}
doesn't actually replace

{{00:02:31}}
or substitute for human,

{{00:02:32}}
for clinician judgment or

{{00:02:34}}
that human interaction.

{{00:02:35}}
I think there's a couple of

{{00:02:36}}
major challenges with AI.

{{00:02:38}}
First of all, we've got

{{00:02:38}}
to make sure that the

{{00:02:39}}
data that's used to train

{{00:02:40}}
AI is high quality data.

{{00:02:43}}
It's accurate, it's complete.

{{00:02:44}}
And it's representative,

{{00:02:45}}
so there's no hidden

{{00:02:46}}
biases in the data.

{{00:02:49}}
Then we need to make sure

{{00:02:50}}
that the tool, that the

{{00:02:51}}
model that's been developed

{{00:02:52}}
or trained on that data,

{{00:02:54}}
is then well validated.

{{00:02:55}}
So in other words, we then

{{00:02:57}}
test it in a number of

{{00:02:58}}
different environments,

{{00:03:00}}
separate from the training

{{00:03:01}}
environment in which

{{00:03:02}}
the model was developed.

{{00:03:03}}
So in other words, does

{{00:03:04}}
it work as good in other

{{00:03:05}}
hospitals and other sites

{{00:03:06}}
with other clinicians as

{{00:03:08}}
it did in the original

{{00:03:09}}
training environment?

{{00:03:10}}
So if that rigorous external

{{00:03:12}}
validation is then carried

{{00:03:13}}
out, Then the next step is how

{{00:03:15}}
you then develop a prototype

{{00:03:16}}
and integrate it into

{{00:03:17}}
electronic medical records.

{{00:03:19}}
Because that's where most of

{{00:03:20}}
these algorithms will sit.

{{00:03:22}}
And I think that's a

{{00:03:22}}
challenge because these,

{{00:03:25}}
electronic medical

{{00:03:25}}
records can be different.

{{00:03:27}}
there's different,

{{00:03:28}}
products,  different

{{00:03:29}}
platforms that can be used.

{{00:03:30}}
So we need to make sure

{{00:03:31}}
that the model works well

{{00:03:33}}
in that particular platform.

{{00:03:34}}
And then finally, we have

{{00:03:36}}
to make sure that clinicians

{{00:03:37}}
in a routine workspace

{{00:03:39}}
then actually use the tool.

{{00:03:41}}
Again, that they're

{{00:03:42}}
not overwhelmed

{{00:03:42}}
with too much data.

{{00:03:44}}
That they, again, they trust

{{00:03:45}}
the tool and are prepared to

{{00:03:46}}
use it when it's required.

{{00:03:49}}
So I think in overcoming

{{00:03:50}}
those challenges, we need

{{00:03:51}}
to make sure then that the

{{00:03:52}}
data, the original data,

{{00:03:53}}
the raw data in which it's

{{00:03:54}}
trained is comprehensive and

{{00:03:57}}
as accurate as we can get it.

{{00:03:59}}
And we therefore need to make

{{00:04:00}}
sure that when people are

{{00:04:01}}
actually entering clinical

{{00:04:02}}
notes or writing things

{{00:04:03}}
into electronic medical

{{00:04:05}}
records, that they understand

{{00:04:06}}
that this has been used

{{00:04:06}}
to train AI and therefore

{{00:04:07}}
needs to be accurate.

{{00:04:09}}
I think in terms of the

{{00:04:10}}
rigorous validation, I think

{{00:04:12}}
that requires a network

{{00:04:14}}
of sites to be involved

{{00:04:16}}
in an integrated research

{{00:04:17}}
network to then do those

{{00:04:18}}
validation studies and

{{00:04:19}}
make sure that the AI works

{{00:04:20}}
in different settings.

{{00:04:22}}
And then we need the technical

{{00:04:23}}
people and the IT people

{{00:04:25}}
and the informatics folk

{{00:04:27}}
to be able to integrate

{{00:04:27}}
the prototypes into the

{{00:04:29}}
electronic medical records.

{{00:04:31}}
And that requires, I

{{00:04:32}}
think  the acceptance of

{{00:04:34}}
the electronic medical

{{00:04:36}}
record vendors to be able

{{00:04:37}}
to integrate those tools

{{00:04:39}}
into their platforms.

{{00:04:40}}
And then finally, it's

{{00:04:41}}
a matter of training

{{00:04:43}}
clinicians then to use the

{{00:04:44}}
tool on a regular basis

{{00:04:45}}
to actually install, make

{{00:04:47}}
sure that that clinical

{{00:04:48}}
workflow is not disrupted

{{00:04:50}}
and that clinicians then

{{00:04:50}}
become adept at using that

{{00:04:52}}
tool in everyday practice.

{{00:04:54}}
Well, I think it's got,

{{00:04:55}}
it's really, it's a, it's

{{00:04:56}}
a very large environment

{{00:04:58}}
in which AI could be used.

{{00:05:00}}
So, I think initially

{{00:05:01}}
where AI probably will have

{{00:05:02}}
immediate effect is to assist,

{{00:05:05}}
particularly with large

{{00:05:06}}
language models and chat GPT,

{{00:05:08}}
is to be able to synthesize

{{00:05:09}}
information, to be able to

{{00:05:11}}
summarize documents, to be

{{00:05:12}}
able to create discharge

{{00:05:13}}
summaries in clinic letters.

{{00:05:15}}
In other words, do a lot

{{00:05:16}}
of the documentation work

{{00:05:17}}
that currently is done

{{00:05:18}}
very manually and takes

{{00:05:19}}
an awful lot of time.

{{00:05:20}}
So I think AI could really

{{00:05:21}}
speed that up and therefore

{{00:05:23}}
release clinicians to

{{00:05:24}}
be able to do more hands

{{00:05:25}}
on direct patient care.

{{00:05:27}}
I think the second thing

{{00:05:28}}
is then when we come into

{{00:05:29}}
diagnostics is to how we

{{00:05:31}}
can use these models and

{{00:05:33}}
large language models in

{{00:05:33}}
particular to actually then

{{00:05:35}}
be helpful in diagnosis.

{{00:05:37}}
In other words, to be able

{{00:05:38}}
to give a differential

{{00:05:39}}
diagnosis for clinical

{{00:05:41}}
presentations where

{{00:05:42}}
clinicians are struggling

{{00:05:43}}
in trying to determine

{{00:05:44}}
what's best, what is the

{{00:05:45}}
diagnosis for this patient.

{{00:05:47}}
I think there are other

{{00:05:48}}
applications in relation

{{00:05:49}}
to radiology and any other

{{00:05:51}}
form of imaging data.

{{00:05:52}}
AI can be used very

{{00:05:54}}
efficiently to localize

{{00:05:55}}
lesions, to be able to

{{00:05:57}}
pinpoint clinicians to

{{00:05:58}}
areas of abnormality on

{{00:06:00}}
an image and therefore not

{{00:06:01}}
waste their time having

{{00:06:02}}
to spend many minutes in

{{00:06:03}}
looking at the entire image.

{{00:06:05}}
And then I think finally

{{00:06:06}}
in relation to treatment

{{00:06:07}}
selection, I think again AI

{{00:06:09}}
can assist there in being

{{00:06:10}}
able to determine on the basis

{{00:06:12}}
of patient phenotypes, okay,

{{00:06:14}}
which is the best treatment

{{00:06:15}}
for this particular patient.

{{00:06:16}}
So I think the scope of

{{00:06:18}}
what AI can do is quite

{{00:06:19}}
unlimited, but there's

{{00:06:21}}
still a lot to be done.

Expert Insight

Join us for an enlightening interview with Professor Ian Scott, Director of Internal Medicine and Clinical Epidemiology at Princess Alexandra Hospital, as he discusses the challenges and immense potential of AI in healthcare.