0:06

Behind the Knife, the surgery

0:08

podcast, relevant and engaging content

0:10

designed to help you dominate

0:12

the day. Welcome

0:23

back to Behind the Knife. This is Kevin and I'm

0:25

here with Jason and we have some exciting news to let you know

0:27

about. Yeah, thanks Kevin.

0:29

So we are now pleased to offer

0:31

bundle and institutional sales for our Behind

0:33

the Knife premium products. The

0:35

bundle includes our general surgery

0:37

oral board audio review, our

0:39

trauma surgery video atlas and

0:42

our oral board audio reviews

0:44

for vascular, colorectal, surgical oncology

0:46

and cardiothoracic surgery. We

0:49

offer institutional discounts to our bundle

0:51

and individual premium products. Yeah,

0:54

so all of this is available via our new

0:56

website and app where you can take

0:58

notes, pen chapters and download chapters for offline

1:01

viewing both on the website and the app.

1:03

So please email hello at behindtheknife.org and

1:05

check out the show notes to learn

1:08

more. And remember,

1:10

this isn't only for board review. You

1:12

can use our content to prepare throughout

1:14

the year and provide didactic learning across

1:16

all levels for all residents. So

1:19

level up the education at your institution and let

1:22

your program directors know that Behind

1:24

the Knife can be bought in institutional packages

1:26

and in bundle pricing. All

1:28

right, let's get back to the episode. Hello

1:30

and welcome to this Behind the Knife

1:32

episode in surgical education. We're the general

1:34

surgery education team for gliimophthale. I'm

1:37

Nicole Brooks, our current surgical education research

1:39

fellow and general surgery resident. And

1:41

I'm Judith French, I'm the PhD education

1:43

scientist for the Department of General Surgery. And

1:46

I'm Jeremy Libman, I'm the DIO and

1:48

director of graduate education at Vauclat. On

1:51

today's episode, we'll discuss the use

1:53

of artificial intelligence in surgical education.

1:56

As more and more advances in surgery

1:58

involve applications of AI. Many surgeons

2:00

and trainees, including us, are

2:03

left without much understanding of the technology behind it.

2:05

It can be overwhelming to imagine where the

2:07

field might be going without really comprehending these

2:10

events. Luckily today we're joined

2:12

by an expert in the field who can help touch

2:14

us up to speed. Dr. Dan

2:16

Hashimoto is an assistant professor and boarder

2:18

and endoscopic surgeon at the hospital or

2:20

the University of Pennsylvania. He's

2:23

also an affiliated faculty in the Penn School

2:25

of Engineering and Applied Science. He

2:27

completed medical school and a master's offense

2:29

in translational research at the University of

2:32

Pennsylvania, provided general surgery training at Massachusetts

2:34

General Hospital, followed by a

2:36

fellowship at University Hospital's Bleakville Medical Center.

2:39

Dr. Hashimoto is also a leading expert in

2:41

the use of AI in surgery. He's

2:44

the director of the Penn and

2:46

Computer Assisted Surgery and Outcomes Laboratory,

2:48

which focuses on using technology to

2:50

improve surgery performance and decision making, with

2:52

a special interest in the translation

2:54

of AI and computer vision for

2:56

surgical video analysis. Dr.

2:58

Hashimoto is a co-founder of the Global

3:01

Surgical AI Collaborative and has held leadership

3:03

positions in many surgical organizations, including the

3:05

Sages AI Excess. He

3:08

has over 75 publications and is the

3:10

editor of the textbook Artificial Intelligence and

3:12

Surgery, Understanding the Role of AI in

3:15

Surgical Practice. We're thrilled to welcome

3:17

you to the show. Thank

3:19

you so much for having me. I love

3:21

behind the knife, so I really appreciate the

3:23

opportunity. All right. So AI

3:25

is currently a hot topic, not

3:27

just in surgery, but it seems

3:29

that across the spectrum. Can

3:32

you briefly describe some of the

3:34

advances that fall under this broad

3:36

umbrella of AI and how they're

3:38

being applied specifically in surgical education?

3:42

Absolutely. No, thank you so much. I know

3:44

I knew that it was kind of hot

3:46

where my grandmother started texting me about it, and

3:48

she doesn't text me about much other than to

3:50

ask me about my kid. And so I knew

3:52

it must have captured some attention

3:54

outside of the direct research

3:57

field. But no, absolutely. I think... that

4:00

when we think about artificial intelligence, obviously

4:02

it's a broad field of study that

4:05

really thinks about how machines can, quote

4:07

unquote, reason or work

4:09

through tasks in a manner that is analogous

4:11

to how a human being might do it.

4:14

And what we've seen, particularly I think

4:16

in the last 10 or

4:19

11 months or so, is sort

4:21

of a second explosion in interest. I would

4:23

sort of think that in the last 10

4:25

years or so, the first explosion came around

4:27

deep learning in 2012, 2013,

4:31

and then most recently, obviously, with large

4:33

language models and really sort of capturing

4:35

the imagination of many fields. And

4:38

what's really interesting though is that

4:40

these types of advances have, I

4:42

think, really made it possible for

4:44

us as education researchers to really think

4:46

about what are the types

4:48

of data that we can now look

4:51

at more

4:53

quantitatively than we

4:55

could before the growth of these

4:57

types of methods. So that is

4:59

to say, taking into account things like

5:02

video, taking into account text entries, or

5:04

perhaps even audio recordings, can

5:06

we potentially analyze those in a

5:08

way that's scalable so that you

5:11

can do it across multiple trainees

5:13

instead of doing things one at

5:15

a time? And can you do it in

5:17

a bit of a quantitative fashion? And I

5:19

think that's where there's a lot of sort of

5:21

interest in the education space about what new world

5:24

has this opened up for us. So,

5:27

I mean, your grandma's talking about

5:29

it, but your grandma's not doing

5:31

AI coding and that kind

5:33

of work, and those surgeons aren't either. So

5:35

what should the surgeon's role

5:37

be as this stuff gets developed?

5:40

Yeah, absolutely. I think it's a question

5:42

I get fairly often. It's, hey,

5:45

I just signed up for this Python course.

5:48

What's my next step to becoming an AI

5:50

researcher? And while I do

5:52

think it's important for surgeons to have sort of

5:54

a base understanding of sort of what's going on,

5:57

I can tell you we're never going

5:59

to be as good as the PhD

6:01

engineers who are doing this for

6:03

the 15 hours a day, they were otherwise spending

6:05

in the operating room. So

6:07

in my opinion, surgeons really need

6:09

to leverage their expertise. And what's

6:12

their expertise? It's around clinical care.

6:14

It's around thinking about how this is going to impact

6:17

our patients. It's about

6:19

thinking on how this is going to

6:21

impact our trainees and the next generation.

6:23

And how do you use these types of technologies

6:26

in a safe way and in a

6:28

meaningful way? I

6:30

think there's a lot of times where an

6:33

idea can come up that comes

6:35

from non-clinicians that says, oh, we

6:37

built this technology and we created

6:40

this application of it. Can

6:42

you please use it and tell us how it works? And

6:45

then we sort of look at it and say,

6:47

well, actually, this doesn't fit our workflow at all

6:50

and this doesn't give us any meaningful information from

6:52

which we might be able to do assessment. And

6:55

rather than saying, go back to the drawing board and

6:57

come back to me again with a new idea, I

6:59

think surgeons would say, hey, let me

7:01

talk to you about what my experience is

7:04

and how I see this technology

7:06

potentially impacting education,

7:09

training, outcomes, et cetera.

7:12

You tell me, is that feasible? Is that

7:14

doable? It really needs to be about having

7:16

a conversation and building an interdisciplinary team that

7:18

can tackle these topics together. So

7:21

you've been involved in a lot

7:23

of projects that use AI intraoperatively,

7:25

which are very interesting, including the

7:27

Go No Go Net project, which

7:30

uses AI to help with intraoperative

7:32

decisions for it during a lab

7:34

fully dissection. Can you

7:36

discuss what your experience was like in

7:38

the creation and implementation of this project?

7:41

Yeah, happy to. And I need to give a

7:43

shout out to my friend, Amin Madani, who's an

7:46

endocrine surgeon at the University of Toronto. He

7:48

really sort of led the charge on this. And

7:51

the way that came around was it was

7:53

never initially intended to be a, can we

7:56

use AI in the OR type of project.

7:59

In his... PhD work that he

8:01

was doing during residency, Meen

8:04

was studying sort of decision-making and

8:06

surgeons and trying to really understand

8:08

the mental models that

8:10

surgeons were developing around understanding safe

8:12

and unsafe claims of dissection. And

8:16

to do that, he and his friend

8:18

Robert Messina had developed a web platform

8:20

that allowed surgeons to view

8:23

a video of a laparoscopic

8:26

cholecystectomy and then they'd be

8:28

asked to mark up, where do you

8:30

think is a safe place to do a dissection? Or

8:33

where is an area on this particular image

8:35

that you would not want to do a

8:37

dissection because you're worried about an injury to

8:39

a critical structure? And

8:42

that allowed him to gather data

8:44

from experienced surgeons as well as

8:46

trainees and compare what the

8:48

differences were in terms of where is

8:50

that safe and unsafe area. And

8:53

when we sat down and looked at the

8:55

data together, it dawned on both of us

8:57

that, these were actually just labels or annotations

9:00

that we could feed to computer

9:02

vision algorithms to see if they

9:04

could also learn where are

9:06

safe and unsafe areas of dissection. And

9:09

that project really became sort of the first

9:11

project that led to the founding of the

9:14

Global Surgical AI Collaborative because we

9:16

wanted to train a model that

9:18

was robust to different types of

9:20

data. So coming from different types

9:22

of institutions, different kinds of practice

9:24

patterns, and we were able

9:26

to sort of scale across several

9:29

different institutions, both academic

9:31

centers, community hospitals, rural hospitals, et cetera,

9:33

to really see if this type of

9:36

algorithm could detect these safe and unsafe

9:38

areas of dissection in all manners of

9:40

different cold bladders. And

9:42

that has since sort of grown. In

9:44

fact, he released a mobile game based

9:46

off of this that you

9:48

can download from the App Store on

9:50

your iPhone or Android that actually takes

9:53

some of these frames of cholecystectomy, looks

9:55

at the safe and unsafe zones that

9:58

were generated by the... NoGoNet

10:00

algorithm and gamifies it.

10:02

So create the scenario where you can look

10:05

at the video, you can mark out where

10:07

you'd want to do your next step in

10:09

your diverse section and it gives you a

10:11

score compared to the algorithm and compared to

10:13

expert annotators who participated in the original project.

10:16

So it's really been very cool to see

10:18

how that has grown. How

10:20

good is that thing at predicting what

10:22

would be the experts' opinion? Yeah,

10:25

so it's pretty good, especially the first

10:27

iteration of it. We're hitting

10:29

somewhere around the 70 to 80 percent

10:32

mark in terms of matching up with

10:34

where an experienced surgeon might want to

10:37

do their dissection, but subsequently

10:39

we've taken advantage of newer types

10:41

of algorithm architectures and have improved

10:43

that to above 90 percent. So

10:45

it actually fits very nicely with the mental

10:47

model of some very experienced HPV surgeons and

10:50

people who, for example, sit on the SAGE's

10:52

SafeCoal task force who are kind enough to

10:54

help label some data and review some of

10:57

the data for us. So

10:59

as you know, we're supposed to

11:01

be assessing our trainees using a

11:04

competency-based framework. In thinking

11:06

about this particular project, how do you

11:08

think that fits in with this idea

11:10

of competency-based assessment? I

11:13

think that's a great question, and not surprisingly

11:15

coming from you because you have such expertise

11:17

on this and in fact I might at

11:19

some point skim that back around to see

11:22

where your recommendations are for us. Here

11:25

I think it's helpful because

11:27

in some ways, it's one

11:29

thing to take a

11:31

trainee through a case and try to

11:33

better understand, okay, is this trainee competent

11:35

to perform this, let's

11:38

say, cholecystectomy independently

11:40

with or without different

11:42

levels of supervision as they're doing it? And

11:44

it's another to try to break

11:47

down what parts of that operation

11:49

are potentially prohibiting them from reaching

11:51

competency. Obviously, an operation

11:53

is sort of the merger of

11:56

different elements of education. So One

11:58

is obviously the decision-making. Component. So.

12:01

understanding. Where.

12:03

Is it it that I want

12:05

to place this particular Maryland grasp

12:07

or tip? So. I

12:09

know that I'm not going to injure

12:11

critical structure and that comes from understanding

12:13

what your anatomical landmarks are, where the

12:16

boundaries, understanding principles of attraction, intention and

12:18

play, and exposure, and working with your

12:20

assistant to get the optimal views. These

12:22

are all these things that come together

12:24

beyond just the can. I move my

12:26

hands in a certain way to put

12:29

my instrument. Kippur. I wanted to. And

12:31

I think this element of can you take

12:33

the video and sort of bringing out of

12:35

the operating room so we can try to

12:37

break down. What? Is it

12:40

about your mental model of

12:42

perceiving? The appropriate next step

12:45

or the appropriate plane. And

12:47

that we can try to give your feedback

12:49

on and can we say that in old

12:51

from a decision making standpoint. You. Are

12:53

appropriately visual perceiving the landmarks such that

12:55

you can make a safe decision around

12:58

what is a quote unquote goes on

13:00

or a no go zone. And

13:02

and that allows I think a an

13:05

educator. Who. With greater confidence, say okay,

13:07

I understand your decision making process. or at

13:09

least I agree with your decision making process.

13:12

Where. Does that translate into? now that we're

13:14

in the operating room and I am sort

13:16

of observing, you are assisting you in

13:18

accomplishing the goals of your mental representation

13:20

of how the servers ago. You've.

13:22

Got the experience binoculars here new

13:25

can log down further than lot

13:27

of other what do you see

13:29

coming down the road. In.

13:31

How we teach her surgical trainees,

13:34

At a eyes really going to

13:36

be impactful for either positively or

13:38

potentially negatively. Yeah. I

13:40

really thank you. Know that minute tea

13:42

party on his computer vision? Peace Because

13:44

I am very biased. It It's so

13:46

that's the majority of what our lab

13:49

dies. But. It I think this

13:51

the growth in interested in using

13:53

video. To. Sort of replay

13:55

performance in get feedback on performance I

13:57

think is going to play a very.

14:00

The drawer where this is gonna

14:02

go forward. I think that as

14:04

technologies are getting better and applications

14:06

are coming to market that allow

14:09

trainees and faculty members. To.

14:11

Just take clips of their

14:13

videos. right? And then use

14:15

that to guide us. feedback

14:17

session. At between a faculty member in

14:19

a trainee or even for pure coaching I had

14:21

so a senior as in a junior resident or

14:23

maybe two reasons for the same level. Added

14:26

you're going to start see greater engagement. With

14:28

visual media. I. And I think

14:30

that. Artificial. Intelligence tools can help

14:33

with. I have seen they can do the

14:35

automated segmentation the steps of the procedure. Was.

14:37

Shown with gonna go nods and and

14:39

said calling up from the Strasburg group

14:41

that you can automatically segment out different

14:44

anatomy and the key structures and things

14:46

like that. It's. So it takes out

14:48

a lot of the the manual labor of

14:50

can I just prepare this video to the

14:52

point where it's gonna be useful. says he

14:54

back in coaching. I. Do think

14:56

that we're gonna see an increase

14:58

in through these a quantitative metrics

15:00

of performance. So Andrew hung from

15:02

the or see I think is

15:04

now Cedar Sinai Urologist. His.

15:06

Group has been very advanced. I'm

15:08

thinking about this automated performance metrics

15:10

that they're gathering from the robotic

15:12

platforms and they've been able to

15:14

show. That. Quantification.

15:18

Of surgical gestures. In.

15:20

A robotic procedures such as possible to me.

15:23

Can. Correlate with outcome so they can

15:25

look at it but we called a killer

15:27

matic profile or how are these robotic arms

15:29

moving during a case. That

15:31

actually predict whether that patient is

15:33

going to have a better functional

15:35

outcome. And ten eyes and categorize

15:38

a surgeon. Based. Off of

15:40

those kinda mad x into and experienced

15:42

are super expert surgeon. And.

15:44

An inexperienced surgeon. And then

15:46

can I try to get that inexperience surgeon?

15:49

To. Match.kinda matic profile.

15:52

Of the super experts such had, they

15:54

also have better outcomes for their patients.

15:57

So. i think what you're going to see is sort of

15:59

this novel use of data in terms

16:01

of providing more specific and

16:03

quantitative feedback to trainees. It's

16:05

almost like what we do in sports now, right? In

16:08

fact, the other day I was at a kid's soccer

16:10

game and the coaches had this

16:12

camera on the side of the field that

16:14

was recording the entire field and later on

16:16

they were using it to provide feedback and

16:18

they were taking very advanced measurements for a

16:20

bunch of seven-year-olds playing soccer to

16:23

try to do coaching strategy and give feedback to

16:25

these kids. And I was kind of

16:27

amazed. I'm like, well, we can do this for seven-year-olds

16:29

in soccer. I don't understand why we're not doing something

16:31

similar here for surgery to make our trainees better at

16:33

taking care of people. How

16:35

do you envision this technology will be

16:37

used in real time in the OR

16:40

about term? Like, do you think it's

16:42

ever going to stop surgeons from dissecting

16:44

in no-go zones or doing other unsafe

16:47

movements like what you're talking about? I

16:49

think that would potentially ultimately be the goal. I

16:51

don't know about stopping in the sense of stopping

16:53

the hand or stopping the robot or whatever it

16:56

might be, but I think it needs

16:58

to be a collaborative decision-making process. Obviously,

17:00

we have a long way to go. There are a

17:02

lot of hurdles to get through for FDA approval and

17:04

things like that for an algorithm

17:06

that functions intraoperatively in real time

17:09

to impact performance. That's a huge

17:11

hurdle to climb from a regulatory

17:13

perspective. I know the FDA is

17:15

thinking about it, but there's no clear guidance yet on

17:17

what those types of algorithms need to look like. But

17:20

my hope, or at least my early vision of

17:22

what I think that's going to look like, is

17:25

basically additional data that's provided to the

17:27

surgeon such that the surgeon can augment

17:30

their decision-making. It may not even be

17:32

that it's running all the time. It's

17:35

probably initially going to be a system where a

17:38

surgeon can say, I think this is what's going

17:41

on. I would like some additional

17:43

data on it. Let's turn

17:45

on the data visualization platform. If

17:48

you're in the Iron Man movies, let's

17:50

turn on Jarvis. Let's ask Jarvis what

17:52

these coordinates are, what these calculations are. You

17:56

can get a better sense of that from a

17:58

data perspective. surgeon can

18:00

take that into consideration with your

18:03

personal experience as a clinician and

18:06

then integrate those two together to make a decision on

18:08

what to do next or how to proceed in a

18:10

given operation for that patient that's in front of you.

18:13

Every fan knows the right player in

18:15

the right position can be a game

18:17

changer. Put LifeLock between your identity

18:20

and identity thieves to monitor and alert you

18:22

to threats you could miss. Plus,

18:24

with a U.S.-based restoration specialist on your

18:26

team, you won't have to face drained

18:29

accounts, fraudulent loans, or other losses from

18:31

identity theft alone. All

18:33

backed by the LifeLock Million Dollar Protection

18:36

Package. Change the game on identity

18:38

theft. Save up to 25%

18:40

your first year at lifelock.com. If

18:43

you only have a 401k, you're not getting the

18:45

most for retirement. Wait, what? Add a Robinhood IRA

18:47

on top, then they'll boost it by 3%. You

18:50

can do that? And if you transfer in any

18:52

retirement account, you get 3% on top

18:54

of that. Is there a limit to the match?

18:56

No limit. Robinhood Gold gets you the biggest contribution

18:59

match of any IRA on the market. Sign up

19:01

for Robinhood Gold at robinhood.com/boost by April 30th. Subscription

19:03

fees apply. Investing involves risk. 3% match requires gold

19:05

for one year from first match must be by

19:08

rate for five years. Match on transfer subject for

19:10

additional terms and conditions. Robinhood Financial Level Team member,

19:12

SIPC. So, you mentioned

19:15

Jarvis. I'm going to bring up

19:17

Hal from an earlier movie. Okay.

19:20

So, where Hal in the

19:22

Stanley Kubrick movie, Space

19:24

Odyssey. Yeah, 2001 Space Odyssey.

19:28

Yeah. Anyway, Hal takes

19:30

over the space station and destroys

19:33

everyone, decides Hal knows

19:35

what's best. So, how do

19:37

we prevent that sort of doomsday

19:39

model as we continue to develop these things? And

19:42

maybe it's not going to be that Hal's taking over

19:44

the robot and dissecting the

19:46

surgeon, but perhaps

19:49

making unwise decisions or

19:51

providing the wrong guidance or, you know,

19:54

as we're using this for higher and higher stakes

19:56

decisions, not giving us the

19:58

best information. I

20:00

think that's where the regulatory component becomes

20:02

key because we know that models drift.

20:04

So what do I mean by that? Once

20:07

you train a model, the current regulatory framework

20:09

is that you have to kind of lock it in place.

20:11

If you're going to do a next

20:14

iteration or an update, do you have to sort of

20:16

resubmit that data so that they can make sure that

20:18

it's safe to release that

20:20

next iteration. But as you

20:22

collect data, practice patterns

20:25

change. And as you

20:27

use technology, your practice pattern changes.

20:30

And that can cause you to

20:32

sort of drift out of the

20:34

original distribution of data on which

20:36

those algorithms were trained, such

20:38

that even just a couple

20:40

of months after an algorithm gets

20:42

released, it could potentially already become

20:45

outdated and give wrong recommendations. That

20:47

is a very real question and a

20:49

big fear that a lot of us have

20:52

as we're developing these technologies is how

20:54

do we account for that? And how

20:56

do we control for that? And how do we ensure that

20:58

that's safe? And we're exploring

21:01

different types of techniques in terms

21:03

of looking at explainability, for example,

21:05

in terms of trying to better

21:07

understand why a given algorithm might

21:10

be recommending XYZ type of step

21:12

or thinks that this plane is better than

21:15

the other plane. But ultimately,

21:17

that's why I think a lot of

21:19

us envision this as being an augmenting

21:21

technology instead of a replacement technology, because

21:23

at the end of the day, it

21:26

does require a human being

21:28

with surgical experience to

21:30

look at that and say this

21:33

is or is not appropriate for

21:35

the given clinical scenario. And

21:37

so in reality, what we really

21:40

fully expect is that the clinician is going

21:42

to have to pay just as much, if

21:44

not more attention than they do

21:46

without an AI algorithm to ensure that

21:48

this is implemented safely. I know

21:51

there's a lot of concern about de-skilled link. So

21:53

what happens if you give people an AI

21:55

algorithm that helps them do a dissection to turn

21:58

off their brain? When

22:00

you use GPS, sometimes you sort of forget

22:02

to pay attention to where you're making your left-right turn. The next thing

22:04

you know, you're at your destination, you don't know how you got there.

22:07

I really don't think that that's going

22:10

to be wise for assertion

22:12

to toothpads when we have these intraoperative AI

22:14

systems running because it's going to be very

22:16

important to sort of keep that algorithm honest.

22:20

So keeping in line with some of the

22:22

barriers to the use and advancement

22:24

of AI, there are several

22:27

ethical and legal implications that

22:29

are related to this in

22:31

surgery, data protection, error accountability,

22:33

limiting bias with equitable data

22:35

sources. So how do you

22:37

balance those challenges in your

22:39

work? Yeah, it's

22:41

very, very difficult, particularly around

22:43

considering the appropriate data sources.

22:46

As we sort of know, just in general from anything

22:48

that we do in medicine, most,

22:51

if not all, of our data sets are biased

22:53

in one way or another. And that's just as

22:55

true, if not more true, for

22:57

the types of data that we're collecting for

22:59

our AI studies. Obviously,

23:01

there's a minority of surgeons

23:03

and institutions that elect to

23:06

record their videos

23:08

in their cases or to provide

23:11

their data for training AI algorithms.

23:14

And that can include text data, not just

23:16

video. So that can include, for example,

23:18

the EMR notes and things like that, or potentially

23:20

even your assessments as a resident, your milestones, your

23:22

EPA's, et cetera. So the

23:24

data that we do get is very much

23:26

biased to toy institutions that are already sort

23:28

of thinking about using this data in this

23:31

way, but it may not

23:33

be reflective of actual practice. And

23:35

what we have to do is really understand the

23:37

distribution from which that data comes from. And

23:40

to the question that came up earlier, what's the surgeon's

23:42

role in this? The surgeon's role

23:44

is not just in developing and using the

23:46

AI technology, but evaluating it. So

23:49

if you are told, I

23:51

have an AI algorithm that can do this for

23:53

you as a surgeon, I do think

23:55

it's the surgeon's responsibility to look at that

23:57

data critically, whether it's a paper or a

24:00

paper. pamphlet or some sort of product brochure

24:02

and ask those very serious questions

24:05

of, was this algorithm trained on

24:08

a population that is reflective of the population

24:10

in which I plan to use it? Is

24:13

this going to be really the best thing for my

24:15

patients? There have been a lot of studies coming out

24:17

in the radiology literature, for example, that

24:19

real world performance of algorithms

24:22

that were otherwise incredibly

24:24

impressive in the clinical trials that

24:26

were submitted to the FDA for

24:29

approval are absolutely abysmal when they got

24:31

in the real world and the incidence of disease is

24:33

like 2% versus 50% in the initial evaluation

24:37

data. And so

24:39

that has led to some very strong questions

24:41

around what it's going to mean to implement

24:44

these in a safe way. But

24:47

100%, we really have to think about biases

24:49

and things that come into that. And

24:51

then also we think about, you know, who are

24:53

the types of patients who are going to be

24:56

willing to donate their data? Or if they are

24:58

donating their data, is it being collected in a

25:00

way that's ethical, that's equitable

25:02

and it's fair? How

25:05

do we in our trainees is the

25:07

reading literature that's coming out and

25:10

stuff that's coming out in surgical literature?

25:13

We used an AI algorithm to do this.

25:15

We used an algorithm to

25:17

show that. How do

25:20

we determine that that's really okay and

25:22

that we really understand where that's coming from and

25:24

that it's going to be applicable to

25:26

our situations? Yeah,

25:29

thank you. In that case, right, and I'll

25:31

tell you, it's a flood of literature coming

25:33

out now. Now that a lot of these

25:35

tools are much more accessible, it's

25:38

really becoming much more widespread. And

25:40

I think it's really relying on

25:42

first principles of research. So it's

25:44

not even just thinking about, you

25:46

know, oh, is this, other questions

25:48

are specific to AI research. These

25:50

are just questions about research methodology

25:52

in general, right? So you

25:54

want to look at what's the size, obviously, of

25:56

that population. Where is that population

25:58

being drawn from? And then when

26:00

you get down to the modeling questions, again,

26:03

it's really thinking about the phenomenon of

26:05

interest. So focusing less so on what

26:07

was the name of the algorithm that

26:09

they used, but really drilling down

26:12

to what was the question being asked? We're

26:14

kind of talking about education right now. I've

26:16

seen some things come up to say, oh,

26:19

we built this algorithm to automatically

26:21

assess the competency of a trainee

26:23

in doing XYZ task. And

26:25

it's always interesting. Oh, I didn't know we had

26:27

sort of a validated way of assessing competency for

26:30

that task. Let me read this. And

26:32

then you read the paper, and they

26:34

had sort of defined competency in a

26:37

very narrow way that was very specific

26:39

to their use case, but hadn't otherwise

26:41

been investigated for any real

26:43

sort of applicability outside that particular

26:45

research study. So then you got

26:48

to wonder, okay, did AI actually determine

26:50

the competency of doing XYZ task in this

26:52

paper, or was it just that AI

26:54

was demonstrated to be able to do this

26:56

task that was specifically defined for this

26:58

paper itself? Right? So those are

27:00

the types of questions to look at from the lens

27:02

of just general research

27:05

methodology. Right

27:07

now, all of our research fellows take

27:09

a statistics course because it's important for

27:11

them to understand that element of how the

27:13

research is done, where it's coming from, helps

27:16

them to better understand what they're reading.

27:19

Should we start having them taking some type

27:22

of foundational AI course or coding? I

27:25

think a conceptual AI course would probably

27:27

be more helpful. But

27:29

I will say that having a

27:31

strong basis in statistics is as

27:33

important as sort of building

27:35

in some component of AI education. Although

27:38

I will say the Royal College of Physicians and

27:40

Surgeons of Canada a few years ago released a

27:42

report where they actually suggested

27:44

that digital health literacy become

27:47

a new fundamental competency in

27:49

the Canadian system because of

27:51

the anticipated growth and

27:53

the expected importance of

27:56

digital health in delivering care. So

27:59

that means try to understand. understand, okay, what do these

28:01

technologies mean? How do I interact with a

28:03

computer scientist or a data scientist who

28:06

may become a part of a healthcare team

28:08

who helps to interpret these types of

28:10

tools such that I can be

28:12

a competent and safe physician in the new era?

28:16

Well, thank you so much for all

28:18

of your insights. It's been very helpful

28:20

to better understand this topic that I

28:22

personally don't have much understanding of. So

28:25

can you go ahead and give our

28:27

listeners an educational timeout, some key takeaways

28:30

on AI and surgical education that they

28:32

should meet this podcast with? Yeah,

28:35

absolutely. I think the number one

28:37

thing is that as magical as AI

28:39

can seem, when you

28:42

are thinking about artificial intelligence

28:44

applications and thinking about, for

28:46

example, papers and things like

28:48

that, it's really less

28:50

about the AI and more about

28:52

first principles in research. So

28:55

any other approach that you would take

28:57

in evaluating a research study, the

28:59

same approach is going to apply to

29:02

evaluating an AI study. Just

29:04

like you may not understand or have

29:06

heard about every single

29:08

statistical analysis in a clinical

29:11

trial, you may not have

29:13

heard about every single type of algorithm that's going

29:15

to be presented in a study that uses AI. But

29:18

that doesn't mean you don't already have expertise

29:20

around understanding good fundamental

29:22

research methodology and

29:24

more importantly, understanding what

29:26

is the implication of that for

29:29

clinical care, for education and training,

29:31

for learning, for teaching, et cetera.

29:34

And so it's not being

29:36

intimidated by the subject

29:38

matter and really relying on the

29:41

excellent training and education that you

29:43

have already had around science and

29:46

understanding science to get you through that.

29:48

And I think that's probably the most

29:50

important lesson to take away from this

29:52

because it's very easy to get intimidated

29:54

by it, thinking, oh, we

29:56

weren't exposed to coding and things like

29:58

that when we were... pre-meds or in

30:00

medical school, but we were

30:03

very much trained to think scientifically

30:05

and to evaluate literature in

30:07

a rigorous fashion. I'm gonna

30:09

ask one more. I'm gonna give you an

30:11

opportunity to become very very famous here. You

30:15

know we brought up the Stanley Kubrick movie

30:17

2001 Space Odyssey that

30:19

was released in 1968. So if

30:21

you're now looking down the road another

30:24

40 years, where do

30:26

you realistically see AI

30:28

taking us in surgery and surgical education?

30:31

Yeah, I think that realistically 40

30:34

years out, I mean at that

30:36

point I would fully expect that

30:39

we have hopefully been able to put

30:41

together a database

30:44

of outcomes that are

30:46

linked to trainees such that

30:48

as you are moving through your training

30:51

process, right, evaluation becomes

30:53

much more quantitative and outcomes

30:55

based. So it becomes

30:57

less about, oh let me

31:00

check off these Likert scale things on your

31:02

evaluation to tell you that you're competent and

31:04

ready to graduate. It's, let

31:06

me look at the data about you

31:08

and your performance, what you've

31:10

done in the operating room, what's your kinematic

31:12

profile, what's your outcome profile, what is your

31:14

decision-making profile based off of the orders that

31:16

you have entered relative to

31:18

when you access certain results that

31:21

can allow me to infer what your

31:23

decision-making process is like for

31:25

x, y, z disease process and

31:28

really create a comprehensive picture of

31:31

who you are as a clinician and

31:34

what it is that you're ready

31:36

to graduate from a competency-based perspective.

31:38

That I think is sort of what I see about

31:41

that 40-year time frame because I do

31:43

fully expect that we're going to have

31:45

better pipelines for the data and

31:47

things like that. I assume we'll

31:49

call the computer Dan. Although

31:53

I will say one thing that I always

31:55

bring up to people is, you know, I

31:57

always ask where was the first self-driving car

31:59

demonstrated? And I'll ask you, do you

32:02

remember the first self-driving car? No,

32:05

I have no idea. I

32:07

used to think it was like, it must have

32:09

been like mid 2000s or something like that, but

32:11

it's actually 1987, if I remember correctly, on the

32:14

Autobot. They demonstrated a self-driving car. At that

32:16

point in time, you can actually

32:18

look up some of the newspaper clippings and everything. Everybody

32:21

was convinced that we were going to have self-driving cars

32:23

by the mid 90s. And

32:25

here we are in 2023, and I don't think we're that

32:27

much closer with how things are going. So,

32:30

and that was about like 30, almost 40 years

32:32

ago at that point. So I could be way

32:35

off. Well, Dan,

32:37

thank you so much for your time.

32:39

This has been incredible, really insightful, and

32:41

a lot of your optimistic view of

32:43

the future of AI. Hopefully,

32:46

we'll all be around to see it. If

32:48

it doesn't take us out first, right? Be

32:52

sure to check out our website at www.behindtheknife.org

32:54

for more great content. You can also follow

32:56

us on Twitter at Behind the Knife and

32:58

Instagram at Behind the Knife podcast. If you

33:00

like what you hear, please take a minute

33:02

to leave us a review. Content

33:04

produced by Behind the Knife is intended for

33:06

health professionals and is for educational purposes only.

33:09

We do not diagnose, treat, or offer patient-specific

33:11

advice. Thank you for listening. Until

33:14

next time, dominate today. Every

33:26

fan knows the right player in the

33:28

right position can be a game changer.

33:31

Put LifeLock between your identity and identity thieves

33:33

to monitor and alert you to threats you

33:35

could miss. Plus, with a

33:37

U.S.-based restoration specialist on your team, you

33:39

won't have to face drained accounts, fraudulent

33:42

loans, or other losses from identity theft

33:44

alone. All backed by

33:46

the LifeLock Million Dollar Protection Package. Change

33:49

the game on identity theft. Save

33:51

up to 25% your first year at

33:53

lifelock.com. On

34:00

top and they'll boosted by three percent. You can do

34:02

that and if you transfer, any were. Present.

34:05

On top of that, a no limit to the

34:07

matte, no than having to go guessing the biggest

34:09

contribution matter of any I agree on the market.

34:12

Sign up for Robin Hood Gold at running with

34:14

on com flash lose by April thirtieth subscription fees

34:16

apply. investing about refugees them as required. Go for

34:18

one year from first Match must be very for

34:20

five years match on transfer such as additional freshly

34:22

condition for a minute financial of remember as I

34:24

can. Just. A

34:26

smidge know. But if you sign up

34:28

for Robin Hood goals ira with a

34:30

three percent match, you can get up

34:33

to one hundred ninety five dollars for

34:35

the twenty twenty three tax. Or oh

34:37

yes, sign up a decimal. Don't lose

34:39

contacts day to get the biggest contribution

34:41

match on the market. Subscription fees, Opinion.

34:46

Investing involved with three percent match requires gold

34:48

for one year from first met Monkey Virus

34:50

for five years lot of his financial hello

34:52

sees member as I just. Waiting

34:56

on a tax return and hopefully it ends

34:58

up in your hands. Fraudulent tax returns due

35:00

to identity has increased by thirty percent and

35:02

twenty twenty three if you're in a bind

35:04

this tag please and Life Lock can help

35:06

or us based restoration specialists are experts dedicated

35:08

to helping solve your identity theft issues and

35:10

lowlife like plants are backed by the million

35:12

dollar protection targets so will reimburse you have

35:15

to the limits of your plan if you

35:17

lose money to the identity theft. help protect

35:19

your information that taxis and whitlow flocking set

35:21

to twenty five percent your first year with

35:23

thing or code news it later.