0:07

Welcome to the Skeptics Guide to

0:10

Emergency Medicine. Meet 'em, greet 'em, treat 'em

0:12

and street 'em. Today's date is April

0:14

15th, 2025, and I'm your skeptical host,

0:16

Ken Milne. The title

0:18

of today's podcast is Together

0:20

in Electric Dreams,

0:23

or is it reality? And

0:26

our guest skeptic is Dr.

0:28

Kirstie Chalen. She is a

0:30

consultant in emergency medicine at

0:32

Langshire, U .S. Schler Teaching

0:34

Hospital. Welcome back to the

0:36

SGM, Kirstie. Hey,

0:39

Ken. I

0:41

look forward to mispronouncing that every single

0:43

time you're on, because I just can't

0:45

get my head around it. What is it?

0:49

Divided by a common language or

0:51

much. Lancashire. Lancashire.

0:55

Not shire like a hobbit,

0:57

but Lancashire. Getting closer.

1:00

Moving on. I've

1:03

been waiting all month to be doing

1:05

this one because I'm just all so excited

1:07

about artificial intelligence. So give us a

1:09

case. It

1:11

may be April, but as

1:13

you sit in your departmental meeting

1:16

with your emergency physician colleagues, you

1:18

will note that the winter surge

1:20

of patients doesn't seem to

1:22

have stopped and the decision fatigue

1:25

at the end of shifts

1:27

is as present as ever. Surely

1:29

AI will be making some of

1:31

these decisions better than us students. That's

1:34

one of your colleagues only half

1:36

joking. Another colleague chips

1:38

in that the med students

1:40

at the nearby university have been

1:42

warned against using chat GPT

1:44

to create differential diagnoses. And

1:46

you're left wondering whether AI

1:48

might be working in

1:51

the ED in the near future. It

1:53

can seem like a conveyor belt of

1:55

human misery sometimes in the emergency

1:57

department where you're going, when

2:00

will it end? But there are ebbs

2:02

and flows and the whole idea

2:04

of don't say the Q word gets

2:06

back to regression of the mean.

2:08

But we're not here to talk about

2:10

some of those fallacies. Let's talk

2:13

about emergency departments and how they can

2:15

be real high pressure environments. Clinical

2:17

decisions, boom, we need to make

2:19

them quickly sometimes and accurately. often

2:22

with incomplete information. Clinical

2:25

decision support or CDS

2:27

tools aim to address

2:29

this challenge by offering

2:32

real -time evidence -informed recommendations

2:34

that could help clinicians

2:36

make better diagnostic, prognostic

2:38

and therapeutic decisions. And

2:41

CDS span a wide

2:43

spectrum from traditional paper -based

2:45

clinical decision rules or

2:47

tools to smartphone apps.

2:49

She's trying to trigger

2:51

me with the rules

2:53

there. Oh,

2:56

they can be on smartphone

2:58

or web -based apps like

3:00

MDCalc to more integrated

3:02

systems within electronic health records

3:04

or EHRs. These tools

3:06

function by combining patient

3:08

data with expert -driven algorithms

3:11

or guidelines to inform care

3:13

pathways. They can

3:15

help determine disease likelihood, risk

3:17

stratify patients, and even

3:19

guide resource utilization. such

3:21

as imaging or admission

3:23

decisions. And remember,

3:26

they are called

3:28

guidelines, not

3:30

guidelines. Thou shalt. Recent

3:33

years have seen a growing

3:35

interest in applying artificial intelligence, AI,

3:38

particularly machine learning, to

3:40

clinical decision support. Unlike

3:43

traditional knowledge -based CDS,

3:45

that relies on literature -based

3:47

thresholds, AI -driven tools

3:49

derive patterns from large

3:52

data set or, the

3:54

air quotes, big data. To

3:57

identify associations and make

3:59

predictions, these non -knowledge

4:01

-based systems promise to augment

4:03

human decision -making by uncovering

4:05

insights that may have

4:07

been overlooked by clinicians or

4:09

the static rules. However...

4:13

the majority of

4:15

AI based CDS

4:17

tools. Remaining

4:20

early development, few have

4:22

been rigorously tested in the ED, even

4:25

fewer have demonstrated improvements

4:27

in patient outcomes or clinician

4:29

workflow. Despite FDA

4:31

clearance for some tools,

4:34

evidence for real world impact remains

4:36

limited. And emergency

4:38

physicians are right to

4:40

approach this technology with

4:42

skeptical optimism. We'll

4:44

need to balance the

4:46

transformative potential of AI with

4:49

a critical eye toward

4:51

evidence, safety, and

4:53

usability. Clinical

4:55

workflow. It is very important,

4:57

but ultimately we want to see

4:59

the poo. We want to

5:01

see patient -oriented outcomes. So

5:03

what's the clinical question we're going to

5:05

be asking on today's SGM? It's

5:09

a two -parter. One.

5:12

What is the current

5:14

landscape of AICDS tools

5:16

for prognostic, diagnostic and

5:18

treatment decisions for individual

5:20

patients in the ED? And

5:23

2. What

5:26

phase of development have

5:28

these AICDS tools achieved? So

5:31

what's the reference for this episode? Karimi

5:33

et al. Artificial

5:36

intelligence -based clinical decision

5:38

support in the emergency

5:40

department. a scoping review. And

5:43

that's in Academic Emergency

5:45

Medicine, April 2025. Yes,

5:48

that's right. It's another

5:50

hot off the press episode.

5:53

So let's run through the Peacot.

5:55

What was the population? Studies

5:58

involving AI, that's

6:01

artificial intelligence or ML,

6:04

machine learning based clinical decision

6:06

support tools applied to

6:08

individual patient care in the

6:10

ED. published 2010 to

6:13

2023. And there were a number

6:15

of exclusions and we'll list those in the

6:17

show notes. How about the intervention or what were

6:19

they looking at? AI

6:21

or ML based clinical decision

6:23

support tools used during patient

6:25

care in the ED. And

6:28

did they have a comparison group? It's

6:30

not really applicable for a scoping

6:33

review. However, the

6:35

review identified whether the

6:37

studies involved any comparison

6:39

with usual care. clinician

6:41

judgment or non -AI tools.

6:44

And what were the outcomes? So

6:47

the review didn't focus

6:49

on a single outcome, but

6:52

instead categorized studies by

6:54

their targeted clinical decision task,

6:56

diagnosis, prognosis, disposition, treatment,

6:58

etc. Outcomes

7:00

were only included if they

7:02

were relevant to emergency

7:05

clinicians' decision making, such as

7:07

predicting ICU admission, mortality.

7:09

only for intervention. And

7:11

I know we've mentioned it a couple of

7:14

times already, but what type of study is

7:16

this? This is a scoping review. Yeah,

7:18

which is different than a

7:20

systematic review. Well,

7:24

it is also an SGM

7:26

hot off the press, and we

7:28

were pleased to have the

7:30

first author on the show. He

7:33

is an emergency physician researcher

7:35

at Vancouver General Hospital who is

7:37

exploring ways to improve the

7:39

development and implementation of artificial intelligent

7:41

models in emergency medical care.

7:44

Welcome to the SGM Hashem. Hi,

7:47

thanks for having me on the show. Well

7:49

as you know because we we talked

7:51

prior to recording when I set this up

7:53

in the last week or so. You

7:56

know, I'm doing this D fill

7:58

or PhD in artificial intelligence and evidence

8:00

based medicine the intersection between those

8:02

two in an attempt to overcome my

8:04

natural stupidity, of course, but you

8:06

know, you look like a smart guy.

8:08

So what got you interested in

8:10

AI? Honestly,

8:12

it was my life. She

8:17

was completing a master's

8:19

in finance and her main

8:21

project was around using

8:23

machine learning to predict stock

8:26

prices. So I

8:28

was hearing about the variety of

8:30

factors and the pure volume of

8:32

data they had to consider and

8:34

that made me think of our

8:36

work in emergency departments, you

8:38

know, our complex patients, all the variables

8:40

we're considering in our decisions. I

8:43

think also the fact that the

8:45

hospital I was training in at the

8:47

time was transitioning to an electronic

8:49

health record also just made clear to

8:51

me how much data we would

8:53

potentially be sitting on and so I

8:55

felt we should explore ways to

8:57

leverage that data to care for our

8:59

patients better and make our jobs

9:01

easier. I don't know about

9:03

you, Kirstie, but I love hearing these backstories about

9:05

how did you end up where you ended

9:07

up because it's not usually, hey, I've got this

9:10

master plan and I'm going to set up

9:12

all these steps and there I am. I

9:14

mean, the same reason I did

9:16

my MBA was because Barb suggested I

9:18

do my MBA because we really

9:20

couldn't communicate at the dinner table very

9:22

well because she's very very good

9:24

at finance and business and stuff like

9:26

that and I would not appropriately. So

9:29

Hashim were you just nodding appropriately at the

9:31

dinner table as your wife talked about things and

9:33

you said maybe I should look into this

9:35

so I can contribute to the conversation. Yeah,

9:39

more or less. I still do

9:41

that a lot with anything else

9:43

finance related, but yeah, I'm trying.

9:45

Yeah, it's one of our physician

9:47

superpowers. It's the rare physician that

9:49

has a really good business acumen. All

9:51

right. Well, we're not doing a business

9:53

podcast. We're doing a nerdy podcast about the

9:56

medical literature. So, okay, Hashem, can you

9:58

give the actual conclusions that your authors came

10:00

up with? Yeah, so

10:02

our conclusion was that we

10:04

found a large number of

10:06

studies involving a variety of

10:08

clinical applications, patient

10:11

populations, and artificial

10:13

intelligence models. But

10:15

despite an increased rate of

10:17

publication in recent years, few

10:19

studies have advanced from

10:21

pre -clinical development to

10:23

later phases of clinical

10:25

evaluation and implementation. Alright,

10:28

so we're going to go through a

10:30

checklist that we derive from systematic reviews,

10:32

but we're going to use it for

10:34

this scoping review. Kirstie, what was the

10:37

main question being addressed? Was it addressed

10:39

clearly? Yes, it was.

10:42

Was the search for studies detailed and

10:44

exhaustive? Yes, they

10:46

used five databases and

10:48

searched the grey literature. You

10:50

know, whenever they say the gray

10:53

literature, I'm wondering if they just called

10:55

up a bunch of gray hairs

10:57

and no hairs in that. What do

10:59

you look it into? Do you

11:01

have anything in your bottom drawer? But

11:03

really they mean things like asking

11:05

experts, which can have gray hair, but

11:07

sometimes not. And also looking into

11:09

conference publications and those types of things.

11:12

Were the criteria used to select

11:14

articles for inclusion appropriate? Yes,

11:16

they were. Do you think

11:18

the included studies sufficiently valid for the type

11:20

of question that they were asked? Yes,

11:23

I do. Were the

11:25

results similar from study to study? They

11:28

were, although this is

11:30

slightly less relevant in a

11:32

scoping review that aims

11:34

to map where the literature

11:36

is. Yeah, they're sort

11:38

of spreading the net wide and

11:40

trying to capture as much as

11:42

they can within a scoping review.

11:44

So they're not looking for things

11:47

like heterogeneity and stuff. Were there

11:49

any financial conflicts of interest like

11:51

Hashem? Was he getting a whole

11:53

bunch of money from big AI? So

11:56

the project was funded

11:58

by the Canadian Institutes of

12:01

Health Research and several

12:03

of the authors have declared

12:05

their financial interests in

12:07

AI companies or research. But

12:10

as we always say,

12:12

conflicts of interest don't mean

12:14

the research. Shouldn't be

12:16

looked at and thought about

12:18

you just need to know

12:20

about them Yeah, no transparency

12:23

is really really good and

12:25

having conflicts of interest doesn't

12:27

automatically negate any findings We

12:29

just need to put a

12:31

little bit more thought and

12:33

skepticism into evaluating those findings.

12:35

So what did they find

12:37

well? They found over

12:39

5 ,000 records were identified

12:42

electronically, and they drilled it down

12:44

to just over 600 studies that they could

12:46

include in the final analysis. So you can see

12:48

that there's a lot of studies going on

12:50

right now. Publication rates

12:52

have increased significantly from

12:54

2019. Many of

12:56

the studies, 40%, came from North

12:58

America, and of interest, less than

13:01

1 % came from Africa. So

13:04

Kirstie, what was the key result? Despite...

13:07

rapidly increasing volume of studies

13:09

across the breadth of clinical

13:11

applications. Few studies

13:13

describe advanced phases of

13:15

testing or implementation of

13:17

these clinical decision tools.

13:21

So why don't we alternate back

13:23

and forth about their four outcomes

13:25

here? I'll go first. The majority

13:27

of the data came from retrospective

13:29

studies and when I say majority,

13:31

I'm talking 79 % were looking backwards.

13:35

The most common outcome

13:37

category was for prognosis

13:39

at 44 .6%. There were

13:41

only a few high

13:44

quality trials with the

13:46

only randomized trial protocols

13:48

and one quasi -experimental

13:50

study. There were no

13:52

published RCTs done in

13:55

a live ED setting. The

13:58

majority of studies were

14:00

in the preclinical in

14:02

silico phase. Under

14:04

3 % had reached

14:06

clinical implementation or post -market

14:08

surveillance. Oh, we're going

14:10

to have to ask Hashem. What

14:12

does in silico mean? So

14:16

it's basically referring

14:18

to silicon computer

14:20

chips. So you're

14:22

just doing that

14:25

study using data. There's

14:27

no, as opposed

14:29

to in vivo, you're not.

14:32

testing it in live patients.

14:34

So it's not in vivo, it's

14:37

not really in vitro unless of course

14:39

in vitro is considered you know some

14:41

simulation like you're part of the matrix

14:43

is that what in silico means like

14:45

being part of the matrix? In

14:49

a way yeah you're

14:51

just you're zeros and ones

14:53

you know it's just

14:55

pure data it's all it's

14:57

really We can kind

14:59

of think of it as an in

15:01

vitro analog. But

15:04

yeah, it's just basically you're

15:06

running these models or developing

15:08

these models on data and

15:10

not testing it in any

15:12

kind of live clinical setting

15:14

at all. So

15:19

I'm glad we got you in early

15:21

there, Hashem, because this is the part

15:23

of the program where we really love

15:25

having authors on board because we can

15:27

talk. nerdy to you and find out

15:29

a little bit more about your research,

15:31

the depth and maybe stuff that ended

15:33

up on the editor's floor. So

15:36

let's go through five nerdy questions

15:38

and I will start with the

15:40

first one. This is

15:42

a different kind of nerdy. There's lots

15:44

of technical details in this field and

15:46

in your data in particular. Could

15:48

you start by helping us

15:50

with some definitions? Is

15:52

there a difference between

15:54

AI and machine learning?

15:56

And you also mentioned

15:58

supervised versus unsupervised machine

16:00

learning. So is somebody

16:02

watching the machines? Yeah,

16:06

it's a really good question.

16:08

There's a lot of variety in

16:10

the definitions. So I'll

16:12

just try to keep it very simple. Artificial

16:16

intelligence is an umbrella

16:18

term for the ability

16:20

of computers to replicate

16:22

how humans think or

16:24

behave. Machine

16:26

learning is one of

16:28

the processes through which

16:30

computers can achieve this

16:33

intelligence. So machine

16:35

learning is really

16:37

the algorithms that computers

16:39

can use to

16:41

identify patterns or make

16:43

predictions and some

16:45

would say without being

16:47

explicitly programmed so

16:49

they can kind of

16:51

learn on their

16:53

own. And that kind

16:55

of distinguishes it

16:57

from what I would

16:59

call traditional statistical

17:01

models, like most logistic

17:03

regression. That's kind of the

17:05

most common one we use in medicine. Although

17:07

many people would argue that

17:09

logistic regression is also a

17:12

type of machine learning. So

17:14

there's a bit of debate,

17:16

but in general terms, that's

17:18

what machine learning is. It's

17:20

kind of how artificial intelligence

17:22

can be achieved. one

17:24

of the ways. Now

17:26

supervised machine learning means

17:28

the model learns to

17:31

make predictions on data

17:33

that is labeled with

17:35

the outcome and that

17:37

labeling is done by

17:39

humans. So

17:42

a human may say

17:44

the outcome is x,

17:46

y, or z and

17:48

the model then tries

17:50

to predict or identify whether

17:52

it is truly x, y, or

17:55

z and it's told whether it's right

17:57

or wrong in that prediction so

17:59

that when it sees a new data

18:01

set it can get better at

18:03

making that prediction. So that's

18:05

supervised learning. Unsupervised

18:08

machine learning means the model

18:10

learns on unlabeled data. So

18:12

there's no x, y, or

18:14

z. It's not being

18:16

told whether it's right

18:18

or wrong in that prediction,

18:20

but it actually just

18:22

identifies its own patterns and

18:24

group's observations based on

18:26

certain characteristics that it is

18:28

finding itself in the

18:30

data. Most

18:33

machine learning models we're talking

18:35

about, especially in clinical decision

18:37

support, are supervised. But

18:39

there's some really fascinating work

18:41

being done with unsupervised models

18:43

that's coming through the pipeline.

18:46

I can think of one

18:48

unsupervised model, and that's

18:50

usually EM residents or sometimes

18:53

unsupervised models of training.

18:55

But they're not in silico.

18:57

They're in carbon. Nerdie

19:01

point number two is

19:03

on article screening. Your

19:06

initial search found over

19:08

5 ,000 records and you

19:10

screened the full text of

19:12

721. That

19:14

is impressive. Weren't

19:17

you tempted to use AI at any

19:19

point to help with that? I

19:23

was and I'm sure

19:25

my co -reviewers were even

19:27

more so. but I'm very

19:29

grateful for their hard

19:31

work on screening all of

19:33

those articles. Frankly,

19:35

I have limited experience

19:37

with AI research tools

19:39

and I haven't established

19:41

trust with them just

19:43

yet. I'm trying

19:45

them out with varying

19:48

degrees of success. And

19:50

in fact, actually one

19:52

of my co -authors on

19:55

this paper recently developed an

19:57

AI model to streamline

19:59

screening for systematic reviews. And

20:02

he's recently published this

20:04

in, I believe, Annals of

20:06

Medicine. It's fantastic work

20:08

and these models and these

20:10

tools are coming through the

20:12

pipeline. But this

20:14

scoping review was from my

20:16

master's thesis and I really

20:18

just wanted to keep it

20:20

as simple as possible and

20:22

Sometimes simple is more work,

20:24

but at least it's safe.

20:26

So that's why we chose

20:29

to do this all by

20:31

hand The third nerdy point

20:33

is about anglo -centricity You

20:35

excluded papers which didn't have

20:37

either the full text or an

20:39

abstract in English. We get

20:41

this from Figure 1. And

20:43

this was six papers in the end, so

20:45

under 1 % of the papers. And

20:48

despite this, most of the literature came from

20:50

North America, Asia, and Europe. Arguably,

20:52

AI could be more

20:54

potentially beneficial in parts of

20:56

the world like Africa,

20:58

where healthcare resources are more

21:00

stretched. But these aren't

21:03

represented in your data set. Did you

21:05

get a feel from your review of

21:07

whether this might happen or why it's

21:09

happening or isn't happening? Yeah,

21:12

it's a valid critique. I

21:15

can't say the review

21:18

itself will speak directly

21:20

to this, but personally,

21:22

I think there is a

21:24

real risk that AI could

21:26

widen the socioeconomic gaps

21:28

that are prevalent not just

21:30

on a global level but you

21:32

know right here at home. If

21:35

your hospital is not

21:37

using an electronic health record,

21:40

the data and therefore

21:42

the effectiveness of any potential

21:44

AI solution is extremely

21:46

limited. So you know

21:48

obviously there are many

21:50

other priorities for underserved populations

21:53

but as a profession

21:55

and especially as researchers I

21:57

think we need to take

21:59

steps to mitigate worsening the disparities

22:01

that already exist. And

22:04

perhaps AI can be a

22:06

tool used for good, but

22:08

those wielding it need to

22:10

acknowledge that responsibility. So

22:13

while our review can't

22:15

really didn't really dig

22:17

into those disparities and,

22:19

you know, regional ones

22:21

in Africa, for example,

22:24

I think certainly that

22:26

that gap does exist

22:28

and hopefully there's a

22:30

way to address that

22:32

in the near future. Yeah,

22:35

one of the things we

22:37

recognize in this space is

22:39

that biases can be amplified

22:41

using artificial intelligence and we'd

22:44

like to use artificial intelligence.

22:46

as a tool ethically and

22:48

responsibly and not contribute or

22:50

magnify or make biases worse.

22:53

I would rather it be used

22:55

for great good and hopefully break

22:57

down some of those barriers and

22:59

see that everybody can contribute to

23:01

this area of research. Yeah,

23:04

absolutely. And the counterpoint

23:06

that I've heard from

23:08

many AI researchers is

23:10

that, well, these biases

23:12

exist already. And AI

23:14

can actually shine light

23:16

on these biases and

23:18

help us mitigate them.

23:21

There's tools, for example, that

23:24

can help with

23:26

translation for patients coming

23:28

to triage. There

23:31

are tools that show

23:33

that patients who don't speak

23:35

as the native language

23:37

often English. are not getting

23:39

as good prediction accuracy with

23:41

some of the tools that we're

23:43

using. So it helps us

23:45

kind of shine a light into

23:47

these issues and these gaps. And

23:51

then we have to kind of reckon

23:53

with how do we solve that gap?

23:55

The AI tool may not be the

23:57

solution to it, but it may expose

23:59

the gap. And then it's on us

24:01

to figure out ways to stop that

24:03

gap. Like

24:05

most things, it's a tool and

24:07

it's how we use the tool that's

24:09

more important than the tool itself

24:12

necessarily. Agreed.

24:15

Which brings us neatly on

24:17

to nerdy point number

24:19

four, which is about outcomes.

24:22

You found that the

24:25

largest group of tools,

24:27

270 out of 606,

24:29

used the AICDS to

24:32

inform prognosis. And

24:35

as a clinician, this makes

24:37

me wonder, well, so what?

24:40

Okay, knowing the expected clinical course

24:43

of a patient can be

24:45

useful, but I'm not sure I

24:47

need AI to tell me

24:49

that the frail 102 year old

24:51

with kidney failure is unlikely

24:53

to do very well. And

24:56

I have limited scope to

24:58

change that. Did

25:01

you find there was

25:03

any exploration of what the

25:05

AICDS tools added from

25:08

the patient's point of view.

25:11

So given the breadth and

25:13

volume of studies we found,

25:16

we weren't really able to

25:18

delve into whether and to

25:20

what extent patient perspective was

25:22

a part of the tool

25:24

development. But it certainly

25:26

needs to be, as

25:29

with any intervention or therapy

25:31

in. medicine or emergency

25:33

medicine specifically. I

25:35

think the point you raise about this

25:37

100 year old with kidney failure

25:39

is actually even more relevant to have

25:41

the clinicians point of view. How

25:44

is this tool going to help

25:46

me in my work? Where does it

25:48

fit into my clinical workflow? Clinicians

25:50

need to be a part

25:52

of defining the problem. And unfortunately

25:55

this just isn't always the

25:57

case with these AI tools. And

26:00

so we end up with

26:02

a lot of irrelevant studies and

26:04

research waste. And I

26:06

wonder if this is behind

26:08

the main finding of our study

26:10

that so many of these

26:12

models are being developed because data

26:15

is available and it's relatively

26:17

easy to create a model. Get

26:19

a data scientist or a

26:21

computer scientist and build a build

26:23

a nice big model that's

26:25

really accurate. What's not so easy

26:28

is identifying a really important

26:30

clinical problem for which that data

26:32

is available. and from

26:34

which you can create a

26:36

timely effective solution. So

26:38

if a model can tell me

26:40

that, hey, you know, of the

26:42

30 CTAS -3 patients that are

26:45

sitting in your waiting room right

26:47

now, they've been sitting there for

26:49

10, 12 hours, this one

26:51

is going to end up in the

26:53

ICU within 12 hours. Well,

26:55

I think that's actually a

26:57

very useful model, but you need

26:59

those key stakeholders. You need

27:01

physicians, nurses, patients to help identify

27:04

what the problem is and

27:06

what the potential solution can be.

27:09

I'm with you on that one. Personally,

27:13

I develop it even further and

27:15

I'd want my model to tell

27:17

me of those 30 CTAS patients

27:19

in the wedding room, which one

27:21

can I stop going to the

27:23

ICU if I do something about

27:25

it now instead of in two

27:27

hours? And I

27:29

don't think we're quite there yet unless you

27:32

know something about the literature that I don't.

27:35

Yeah, that's a difficult question

27:37

because I think there's

27:39

so much variety in what

27:41

the interventions may be

27:43

for that potentially critically ill

27:45

patient. But

27:47

it's a really good, I think at

27:49

least if we can get the model

27:51

to identify those high risk patients and

27:54

then we can kind of focus

27:56

our attention on them a little bit

27:58

more. Say, hey, this patient's really high

28:00

risk. I'm sure once we identify that

28:02

patient, something will come up. It's like,

28:04

oh, their heart rate's 125. Oh,

28:07

their blood pressure is actually

28:09

like 100 systolic and they're on

28:11

antihypertensives. And so we'll probably

28:14

clinically identify something that the model

28:16

cannot specifically tell us, but

28:18

at least it's shown some light

28:20

on that patient. I

28:22

also think that the

28:24

reverse is very true that

28:26

not just identifying super

28:28

high -risk patients, but a

28:31

model that can valid and

28:33

accurately identify patients that

28:35

are very low risk that

28:37

we don't need to

28:39

put extra resources towards in

28:41

the the next few

28:43

hours would be really helpful,

28:45

especially as emergency departments

28:48

are overcrowded and we have

28:50

staff shortages and resource

28:52

shortages. So knowing how to

28:54

better utilize what we

28:56

have is, I think, a

28:58

major potential boon for

29:00

this kind of technology. I

29:03

am an AI optimist, skeptic

29:06

but optimist and I'm glad you

29:08

brought up the idea of

29:10

hospital crowding because you know you're

29:12

putting all of these resources

29:14

into trying to find the signal

29:16

in the noise who's really

29:18

sick and who's really not sick

29:20

and separate them all out

29:22

and Boy, wouldn't it just be

29:24

better if we had enough

29:26

staffed inpatient beds and were adequately

29:28

staffed in the emergency department

29:30

with natural things rather than artificial

29:32

things? And yet, how

29:34

much is it going to cost to

29:36

do all of this artificial intelligence

29:38

stuff? Which, you know, I'm obviously a

29:40

proponent of. But what about

29:43

the human factor? Why can't we just have

29:45

enough staff? Can't we just hire a couple of

29:47

more nurses? And then

29:49

we don't have to sit around

29:51

going, hmm, I wonder if that

29:53

person's sick or that person's not

29:55

sick. We'd actually have one of

29:57

the best diagnostic tests known to

29:59

medicine. And that is called your

30:01

retina. Look at the patient. But

30:04

anyways, I'll get off that one because it's

30:06

a segue into number five. And

30:08

that's where most of this stuff is

30:10

in the developmental phase. And you

30:12

highlight that the literature is coming from

30:14

these early phases of artificial intelligence

30:17

and clinical decision support tools. You

30:19

know, it's looking at preclinical stuff. It's

30:21

looking at offline validation, which is I

30:23

assume the silico thing, which I'm loving

30:25

that new term. Thank you very much.

30:28

But few tools have undergone

30:30

these large -scale safety and

30:32

effectiveness trials or even

30:34

post -marketing surveillance. So

30:37

we like shiny new things.

30:40

So how do you think regulators, administrators,

30:43

and the EM community in

30:45

general should use this information

30:47

from your scoping review because

30:49

it's like, oh, that's so nice

30:51

and shiny. Me want. Yeah.

30:56

I think they should

30:58

be excited by the potential

31:00

but disappointed by the

31:03

lack of realization. Some

31:05

of these tools have been around

31:07

for over a decade and yet we

31:09

haven't found a way to test

31:11

them or implement them effectively. I

31:14

think that points to a

31:16

need for more research, more

31:18

knowledge translation and more innovation.

31:21

I think that's going to all

31:23

require more funding and better

31:26

use of that funding. It's

31:28

not enough to create that big,

31:30

shiny, highly accurate model if it

31:32

doesn't stand a chance of being

31:34

useful to clinicians or patients. And

31:37

as a researcher, I think we

31:39

need to be more creative. I

31:42

think we need to interface with

31:44

fields like implementation

31:46

science, human factors,

31:48

engineers, healthcare and

31:50

industrial design. We

31:52

need more qualitative research and we

31:54

need more patient and community engagement. Clearly

31:56

the same old approach is not

31:59

working, so I think it's time to

32:01

think outside the box a little

32:03

bit. And most of

32:05

all, I hope the emergency

32:07

clinical community stays cautiously optimistic.

32:10

about AI and how it can

32:12

enhance our work and help us

32:14

take better care of our patients

32:16

and ourselves. But yeah, I think

32:18

there is still a lot of work to be done in

32:20

this field. Hashem, that

32:22

was a great answer because I love

32:24

that you brought it back to patient

32:26

care and that they should be involved

32:28

in the development, participate in the research,

32:31

talk about their preferences and their values

32:33

and really focus on what's important to

32:35

them. And I really like the fact

32:37

that you brought it back. to

32:39

patients because it starts with patient care and

32:41

it ends with patient care and that's why we're

32:43

there. 100%.

32:46

I agree.

32:49

So those were our five nerdy questions,

32:51

but what did we forget to

32:53

ask? Is there anything else you want

32:55

to highlight from your study or

32:57

this area of research? No,

33:00

I think that's great. I'm just

33:02

appreciate the chance to talk about

33:04

it. Am I allowed to do

33:06

a little shout out here? Oh,

33:09

absolutely. Shout out. Okay,

33:12

awesome. So at

33:14

the Canadian Association of

33:16

Emergency Physicians, we have

33:18

a AI special interest

33:20

group. And we're

33:22

actually going to be hosting a

33:24

couple of events at ISEM 2025

33:27

in Montreal this year. So

33:29

it's a pretty exciting

33:31

stuff. So we have a

33:33

panel discussion with several

33:35

international AI experts. And

33:38

we also are doing an

33:40

AI networking event over some cocktails.

33:43

So if any of the listeners

33:45

are interested and want to

33:47

join, please feel free to reach

33:50

out to myself or the

33:52

event organizers. I think we're planning

33:54

to send out a communication as well

33:56

in the coming weeks about when those

33:58

events are going to be. And

34:00

there's tons of our special interest

34:02

group members are going to be

34:04

doing their own presentations on the

34:07

topics that they've been researching. So

34:09

yeah, there's lots of good AI stuff

34:11

at ICEM this year. So please come

34:13

on out. Hashem, is

34:16

this only open to the

34:18

SGM listeners or could maybe one

34:20

of the SGM hosts put

34:22

their name forward to be part

34:24

of the interest group? Yes,

34:27

as long as you're

34:29

a member of CAPE right

34:31

now, you are welcome

34:33

to be a member of

34:35

our AI Special Interest

34:38

Group and we're under the

34:40

umbrella of the Digital

34:42

Emergency Medicine Committee. So

34:44

you can join that and then

34:46

you can join AI Special Interest Group

34:48

from there. Alright,

34:52

well, those are our nerdy questions and

34:54

thanks for giving us a little extra there.

34:57

It's time to comment on the author's

34:59

conclusions and compare them to the SGM's

35:01

conclusions. Straightforward,

35:03

we agree with the author's

35:05

conclusions. Well, how about

35:07

the bottom line? Is it straightforward as well? Yep,

35:11

artificial intelligence -based clinical decision

35:13

support tools in the ED

35:15

show promise, but...

35:17

we need rigorous

35:20

evaluation before they're routinely

35:22

implemented. And how

35:24

about resolving that case that you started?

35:27

You agree with your colleagues that

35:29

this is a rapidly expanding

35:31

field, but your jobs are

35:33

probably safe for another few

35:35

years. And how are

35:37

you going to apply this information clinically? We

35:41

are going to ask what in

35:43

the UK we call the Chief

35:45

Clinical Information Officer. to

35:48

attend the next staff

35:50

meeting to discuss the potential

35:52

benefits and the potential

35:54

harms of AICDS in the

35:57

ED. And on this

35:59

episode, we don't have what would you tell the

36:01

patient because we don't have a patient involved

36:03

in this case. So we'll skip ahead to the

36:05

Keener contest. And

36:09

it's been a few

36:11

weeks. So the last episode's

36:13

winner was our good

36:15

friend from all... way

36:17

over in New Zealand, Stephen

36:19

Stelz. He knew sudden

36:21

onset shortness of breath or

36:24

dyspnea is the most common

36:26

presenting symptom for pulmonary embolism.

36:29

Kirstie, I know you got a hard one this

36:31

week. What's the question? Where

36:35

was the first

36:37

computer that could store

36:39

a program in

36:41

its electronic memory? Well,

36:44

if you know the answer to this

36:46

week's question, then send an email to

36:48

theSGM at gmail .com with Keener in

36:50

the subject line. The first correct answer

36:52

will receive a shout out on the

36:54

next episode. And

36:56

now it is your

36:58

turn, SGMers. What

37:01

do you think of this

37:03

episode on artificial intelligence? X

37:05

previously tweet your

37:08

comments or on Blue

37:10

Sky using the

37:12

hashtag SGM Hop. What

37:15

questions do you

37:17

have for Hashem and

37:20

his team? Ask

37:22

them on the SGM blog and

37:24

the best social media feedback will

37:26

be published in Academic Emergency Medicine.

37:29

Or you could ask ChatGBT for

37:31

what you think about this

37:33

episode or what questions you want

37:35

or any other AI model

37:37

that you want to use and

37:39

then you can send those

37:41

questions to Hashem. Even those might

37:43

get published in Academic Emergency

37:46

Medicine. Well, thank you,

37:48

Kirstie, for doing another SGM hop with me.

37:50

It's been a blast as ever. And

37:53

Hashem, I've learned some stuff today. I

37:55

like this in silico stuff. Yeah,

37:57

you've taught me a few things about artificial

37:59

intelligence. That the fact that CAPE has

38:01

this, you know, a special interest group.

38:03

I'm, as soon as we hang up,

38:05

I'm putting in my application. But I really

38:08

appreciate you coming and sharing your master's

38:10

thesis and the publication in AEM. Thank you

38:12

so much for having me, it's been

38:14

a pleasure. And if you

38:16

could give the SGM tagline in

38:18

your best robotic voice. Remember

38:21

to be skeptical of

38:23

anything you learn even

38:25

if you heard it

38:27

on the skeptics guide

38:29

to emergency medicine. That

38:32

was really nice. Talk to

38:34

everyone next week.