0:00

[Music]

0:02

hello and welcome to technically

0:04

speaking where scientists and Engineers

0:05

come together to chat about a common

0:07

interest to share knowledge and satisfy

0:08

some curiosity I'm Laura and I'm joined

0:11

by Sarah and Antonia to talk about

0:13

artificial intelligence and how it could

0:14

be used in healthcare to help with

0:16

Diagnostics or what else it might do for

0:18

medicine so Antonia I think the

0:21

inspiration for this episode started off

0:23

with you so tell us about it so I was

0:25

reading an article and the headline said

0:28

chatgpt4 can pass the US GP medical exam

0:32

and I thought hmm interesting we've

0:36

heard all sorts of things coming out of

0:38

chat GPT

0:40

and I thought about one of my friends

0:42

has recently passed her GP exam in the

0:45

UK maybe it'd be a good time to talk

0:47

about it fair enough and that's you

0:49

Sarah so you're a medical professional

0:51

so um do you have any thoughts about

0:53

this just a caveat I actually haven't

0:54

passed my GP exam yet I've passed my GP

0:57

entrance exam so as the exam to get into

1:00

GB training so I'm both a medical

1:02

professional and a surgical educator my

1:05

initial response to AI or the concept of

1:08

AI passing these exams is it it

1:11

threatens my job

1:13

um so that's my immediate gut reaction

1:16

and I also know that it's going to be

1:19

heavily invested in because

1:22

the NHS will need to reduce Workforce

1:25

Workforce costs over time so I think

1:28

that it's going definitely going to be

1:29

an area of interest in the future all

1:32

right and I guess one big question there

1:34

is is it better to have an AI doctor or

1:38

GP or surgeon even or is it better to

1:42

have a human doing that job and what's

1:43

the difference

1:45

so I guess we should start as we always

1:47

do in this show with defining what AI is

1:50

now I know it's something that needs to

1:52

be trained by feeding it data I don't

1:55

really have much experience that I've

1:56

never used GPT I've heard a lot about it

1:58

but I don't know how any of these things

1:59

actually work so if you guys got an

2:00

experience of AI or have a better

2:02

definition I don't have a definition as

2:05

much as just a AI stands for artificial

2:08

intelligence and I think it it gets

2:11

mixed with some of the terminology like

2:13

machine learning and neural networks and

2:16

I think those are two of the key tools

2:18

that are making up the AI that we see

2:21

around at the moment such as in chat GPT

2:24

or Bing's search engine is now also AI

2:28

powered and my experience is we've got a

2:31

bunch of data and fed it into the

2:34

machine as it were and told it to either

2:37

identify patterns or replicate something

2:42

and try to do the next iteration that

2:44

follows based on that information that's

2:47

been given and so it's learning and

2:50

sometimes it will develop a new way to

2:54

pick up those patterns and so it becomes

2:57

its own teacher and that we're not

2:59

telling it exactly what pattern it

3:01

should be reading the neural network I

3:03

think helps it combine more things so

3:07

it's not just a set data set that's my

3:10

understanding as someone who isn't in

3:12

computer science

3:13

I think I mean I'm also not in computer

3:16

science but I think that the term

3:17

artificial intelligence is a bit of a

3:19

misnomer because it's only as

3:21

intelligent as the information that we

3:24

feed it so it is based on the data that

3:27

we give it it can't create its own new

3:31

data and it only learns as much as we

3:35

reinforce that learning it it can't

3:38

reinforce its own learning if that makes

3:40

sense or if it does it creates its own

3:43

feedback loop and in which case it has

3:46

its own biases that will get Amplified

3:49

as it goes through more and more

3:50

iterations in a popular culture example

3:53

I want to say it was Twitter that tried

3:56

an AI on people's tweets and then it

3:59

found it got more aggressive more races

4:02

more bigoted that they had to shut it

4:05

down within a very short time frame

4:07

because they realized that what people

4:08

put on the internet should not be

4:10

repeated and Amplified this is something

4:13

that's I think is important to consider

4:15

when we talk about bias in healthcare

4:17

and bias and AI because any data that we

4:22

put into it any bias that we put into it

4:25

will be perpetuated and there's already

4:28

something that I'm very passionate about

4:29

is equity in healthcare and any

4:31

inequities that currently exist in

4:33

healthcare is only going to be

4:35

perpetuated by this AI it can't go

4:38

against any of the societal prejudice is

4:41

that we have one of the things that's

4:43

important to me is critical thinking so

4:46

the ability to take in information and

4:48

weigh it against other information and

4:50

say well does that match what I already

4:52

know about a particular thing or am I

4:55

spotting some sort of patent or

4:56

something doesn't fit into that pattern

4:57

that doesn't make sense that I should be

4:59

questioning I I don't know if AI has

5:01

that same capability to think critically

5:03

about something and decide whether

5:04

something is true or correct or

5:08

appropriate or not I don't know is the

5:11

honest answer I think it's able to say

5:14

this doesn't fit the patterns I've been

5:15

fed but able to truly assimilate

5:20

knowledge

5:21

and create new theories I don't think so

5:26

I think it'd be interesting because um

5:28

like you know asimov's Laura robotics we

5:32

put some certain rules in to ensure you

5:35

know the Safety and Security of our

5:37

future as we developed robotics at the

5:39

time I'm not going to repeat them

5:41

because I don't remember them off the

5:42

top of my head

5:44

but I wonder if we do put those barriers

5:47

in

5:48

can we put those barriers in in a

5:51

non-biased way but also isn't too

5:54

restrictive because then there's that

5:55

idea that you've put in a limitation but

5:58

then does that create other consequences

6:00

that we didn't foresee when we release

6:03

it out into world and actually it has a

6:06

complete blind spot because we said

6:07

you're not allowed to think about X so

6:10

then

6:12

never thinks about X I guess it depends

6:15

what you're using it for I guess pop

6:16

culture example is iRobot this robot was

6:19

sort of almost like a human being and

6:21

seemed to be doing certain things and

6:23

seem really independent but a more

6:25

practical example of what artificial

6:26

intelligence can be used for is I think

6:29

I saw um something a while ago now I've

6:31

mentioned it in a previous episode about

6:32

training AI to be able to tell when a

6:35

tumor is cancer or not based on an image

6:38

of it and it was actually better at

6:40

doing it than the specialist who had

6:42

been using these images to try and

6:44

diagnose cancer for decades maybe and it

6:47

seemed like the AI had been fed

6:48

sufficient data that it could spot I

6:51

guess really subtle color using the

6:52

images that sometimes a person might

6:54

miss because they're tired or they've

6:56

not eaten or they're just having a

6:57

really bad day or they're distracted by

6:59

something AI doesn't forget and it

7:01

doesn't get distracted necessarily so I

7:03

can I can see how in that instance it's

7:05

a lot more useful than just doing some

7:06

big picture out in the real world doing

7:08

everything thing yeah so for focused

7:11

tasks absolutely

7:13

and that specifically I was talking to

7:16

some histopathologists

7:18

um some colleagues of mine who very that

7:21

to explain the term

7:22

um basically they're the people when you

7:24

take a tumor out you send it off to the

7:27

lab they're the lab people that then

7:28

slice it up and look at it under a

7:30

microscope and tell you what is in that

7:32

biopsy

7:34

um so I was speaking to some of them and

7:36

they were saying that actually the way

7:37

that they study for their exams is

7:39

purely by looking at hundreds and

7:42

hundreds of slides hundreds and hundreds

7:43

of cases and that really interested me

7:45

because actually that's exactly what we

7:47

do for AI is that we just feed it all of

7:49

this data and eventually it comes to

7:51

recognize those patterns the other thing

7:53

that I was talking to them about because

7:54

I raised the idea of oh well that's

7:57

exactly what AI does so how is your job

7:59

any different

8:00

um and they were talking about actually

8:02

it's something that they are worried

8:04

about in the the world of histopathology

8:06

that they now need to look at career

8:09

tracks that are AI proof because there

8:12

is AI in some hospitals in America and

8:16

they're about to bring it into a

8:18

hospital in Nottingham as well to start

8:21

looking at the more simple cancer slides

8:23

add to free up histopathologist to do

8:25

something else

8:27

it's not a particularly attractive like

8:28

career or part of medicine

8:32

um which is possibly why they're so

8:34

heavily investing in AI but also they're

8:36

heavily investing in AI because it's so

8:38

easily replicated by AI

8:42

yeah and I imagine there are a lot more

8:43

Niche things that those professionals

8:45

could be doing not just looking at

8:47

images and saying yes cancer neural

8:49

cancer they can be talking to people but

8:51

actually

8:53

so these specific doctors don't have any

8:56

patient contact right what do they do

8:59

genuinely as they look at slides and

9:01

they tell you what is on the slides that

9:02

is it that is their entire job okay so

9:04

their job is effectively being taken by

9:06

the artificial intelligence yeah which

9:08

is why they've been talking about oh

9:09

well you know if you go into things like

9:11

prostate core biopsies which are much

9:13

more

9:14

specialist and require a human eye as

9:17

opposed to ai ai isn't as good good with

9:19

those

9:20

um that's like the route that they're

9:22

going because they're they're already

9:23

having to think about AI proof

9:25

sub-specialities all right wow you get

9:28

this idea that all medical professionals

9:30

talk to patients all the time but I can

9:33

see that's not necessarily true there

9:34

are a lot of people in the lab just

9:35

doing analysis supporting the people

9:37

that are I guess on the front line it's

9:39

one way of looking at it

9:41

yeah the the patient client facing part

9:44

of a person those doctors did they go

9:46

through the same training that you would

9:48

have gone through so the same amount of

9:50

like time and sweat to get to that point

9:53

so we all have the same degree and then

9:55

after two years of like basic training

9:59

um and rotating through hospitals you

10:01

then choose what specialty you want to

10:02

do so at that point they they will have

10:04

had a different life compared to me in

10:06

that sense that means loads more trainee

10:09

doctors could be free to go to other

10:12

disciplines it's just the challenge of

10:14

people who are already trained in that

10:17

field will have to find another field

10:19

which they can apply their skills to

10:21

yeah and

10:23

um in a Utopia uh Ai and machines taking

10:27

our work would mean that we have more

10:30

free time and the ability to enjoy our

10:32

lives but of course we understand that

10:34

that's not how Society works so it is a

10:37

threat to people's livelihoods and that

10:38

I think is going to be one of the core

10:40

reasons why if you you speak to a doctor

10:42

their initial response will be no don't

10:44

like it this is my job only I could ever

10:47

do this and AI of course is always going

10:49

to be inferior but actually there will

10:51

be some ways in which is superior

10:53

I I can see again that um being able to

10:56

diagnose something that frees up time to

10:59

do something else would be particularly

11:01

useful if you've got any specific

11:01

examples of how it could be superior one

11:04

of the main ways that I I think it will

11:06

be very useful is when you have for

11:10

example healthcare workers who are

11:12

working in isolation in remote remote

11:13

areas so you're the only doctor for a

11:16

community

11:18

um it's useful to be able to bounce

11:19

ideas off of something that we do quite

11:20

a lot in hospitals as we speak to

11:22

consultant colleagues about I've got

11:25

this really interesting case and I'm not

11:26

really sure what to do I think it might

11:28

be this I think it might be that where

11:30

should I go with this and you can kind

11:32

of have that conversation with an expert

11:34

opinion

11:36

um so I think it would definitely

11:39

you know benefit people who are working

11:41

in isolation

11:43

um it might also improve access to

11:45

health care so for example if you've got

11:47

an impoverished Community then AI is

11:50

going to be cheaper than hiring a doctor

11:53

so it could also improve access to

11:55

Healthcare in those sorts of ways

12:03

statistician compared to an adequately

12:07

trained doctor

12:09

but you have to train those doctors so

12:12

it does then mean that you don't have to

12:14

train people in that skill anymore

12:17

um

12:18

so you can cut out quite a lot of money

12:21

and effort there if you wanted to I

12:23

would argue that becoming overly reliant

12:26

on AI could be quite dangerous though if

12:29

for I mean stuff that happens all the

12:31

time in the NHS like our servers go down

12:33

and then we suddenly have no access

12:35

genuinely we suddenly have no access to

12:37

things like drug charts to things like

12:40

um patients observations so like their

12:42

blood pressure and their heart rate so

12:44

suddenly we have no access to any of

12:46

this sort of stuff and we have to go

12:47

back to being on paper so if we become

12:50

overly reliant on AI we have to have the

12:52

infrastructure there to support it all

12:54

right you touched on a lot of points

12:55

there one that I wanted to pick up on

12:57

was the idea that yeah it'd be great to

13:01

to have a sort of colleague inverted

13:04

commas

13:05

um as an AI to to discuss ideas but

13:09

we've also seen like people talk about

13:11

AI hallucinating and just getting basic

13:14

facts wrong that's what people have

13:16

called it in the industry of just it

13:18

just literally

13:20

says so you could look up on a spec

13:23

sheet sorry specification and it says

13:25

this camera phone has this megapixels

13:28

and instead the the AI just picks up a

13:31

random one and just says now it's got

13:34

like 50 billion

13:36

I have heard this for teaching purposes

13:38

essentially a lot of lecturers were

13:39

saying that students are using AI to do

13:43

the coursework for them and the air the

13:45

eye had been um set up in a way that it

13:48

was deliberately putting errors in there

13:49

so the lecturers could spot it so I

13:51

wonder if that's what the Hallucination

13:52

is about I I'm gonna let you into a

13:55

little trade secret here the majority of

13:58

my medical degree was learned on

14:00

Wikipedia

14:01

one of the one of the skills of being a

14:04

doctor is not necessarily being able to

14:07

have the entirety of medical knowledge

14:11

in your mind but being able to to like

14:15

to know where to go to look it up and

14:18

you have to retain that skill even when

14:20

you're using AI if it outputs something

14:22

that you go well that's a bit weird you

14:25

need to have other options to be able to

14:28

to know that it's you know giving you

14:30

the correct answer I would agree with

14:32

that actually because um I've been a

14:34

scientist working in labs and

14:35

supervising students for years now we

14:37

would all say the same thing that it's

14:39

not that you've got an entire textbook

14:40

in your head you definitely have a lot

14:42

of knowledge in there and that knowledge

14:43

helps you infer things from other bits

14:46

of knowledge that you may be not quite

14:47

as familiar with so same sort of thing

14:49

yeah absolutely there was something you

14:51

said fairly near the start about a

14:53

patient facing roles essentially so I

14:56

have a story from many many years ago I

14:58

went to my GP my general practitioner

15:00

saying I've got a sore throat and I've

15:02

had sore throat for a while but I've not

15:03

been able to do anything about it

15:05

because of my situation my situation has

15:07

now changed so I can get to you during

15:09

normal working hours take a look at it

15:11

tell me what's going wrong you can

15:13

definitely see something is wrong back

15:14

there and they literally laughed at me

15:16

an AI wouldn't have done that no it

15:18

would have asked me questions yeah and

15:21

Antonio you actually had a paper that

15:24

you sent me today saying that AI has

15:26

been proven to have better bedside

15:27

manner

15:29

um than doctors and I can entirely see

15:33

that uh as you said we all are human we

15:36

all have human error and empathy is

15:39

proven to be a skill communication has

15:41

proven to be a skill it can be learned

15:43

but people have to put the effort in to

15:45

learn it so yes on the one hand AI would

15:49

never make a mistake such as outputting

15:51

hahaha are you giving them uh symptoms

15:54

and I should joined it too unless you

15:57

joined it too

15:58

unless it's taught or that all of our

16:01

imperfect data but

16:03

I also think as I said it may be an

16:07

equal or superior diagnostician but it

16:09

might will not necessarily be an equal

16:11

or Superior clinician and the reason why

16:13

I say that is because a lot of my job is

16:15

delivering bad news especially diagnoses

16:18

of cancer and for that you need a very

16:21

specific amount of tact and the ability

16:24

to respond to the patient in front of

16:25

you to the person in front of you and

16:27

yes probably eventually you could get to

16:29

the point where if they look down

16:30

they're sad or if they make that very

16:32

specific twitch of their facial muscles

16:34

they want to hug eventually the AI will

16:36

get to that point but will

16:39

will it ever be able to replace

16:42

the warmth and the emotion that comes

16:46

from a real person connecting with you

16:48

and holding your hand in that moment I

16:50

don't think so no I think that human

16:53

connection is what would make a big

16:54

difference in a situation like that and

16:57

I will say I've been treated by some

16:58

healthcare professionals who have been

17:00

absolutely amazing you don't have to

17:02

defend them it's absolutely

17:04

I'm just thinking of some nurse

17:06

practitioners that are just not knowing

17:07

exactly what to do and sorted things out

17:09

for me and just understood what I'm

17:11

going through without me having to say

17:13

anything to them other than my arm is

17:15

swollen and numb help yeah yeah and I I

17:19

would argue that it's their life

17:20

experience that means that they're able

17:22

to provide you with the exact support

17:24

that you need and it's their experience

17:25

of dealing with people who are similar

17:27

to you and there's no reason why we

17:28

can't input that data into an AI but do

17:31

we want to yeah

17:34

laughs is that what we expect you know

17:37

there's also that kind of managing

17:38

expectations that you know someone's

17:41

always wanting to go I want a second

17:43

opinion I want the person with the most

17:46

credentials or do we then start putting

17:48

AI with the best success rate in front

17:51

of them and say well they've got a

17:53

99.9999 correct rate true which you will

17:58

never be able to say of a human

17:59

absolutely I don't know thinking about

18:01

the film AI

18:03

the one with Jude Law and the other

18:06

child actor that I can't remember the

18:08

name of um

18:11

Haley Haley Osmond Haley Joel Osmond

18:13

that was it anyway

18:15

AI that film because they have robotic

18:18

sex workers they had lots of companion

18:20

robots

18:21

who were able to emulate and then they

18:24

had AI robot children that were able to

18:27

emulate those sorts of human connections

18:30

and human emotions

18:32

um

18:33

so I think the fact that we are creating

18:38

films and literature about it

18:41

probably means that actually ultimately

18:44

we will end up wanting it

18:46

um if you could and it's it's something

18:48

I don't know that I've done with my

18:50

friends of like if you could build your

18:52

perfect man what would they be like

18:55

I mean if you could build your perfect

18:57

doctor what would they be like and if

18:59

you could why wouldn't you because they

19:01

would be perfect

19:03

yeah what I don't really know what I'd

19:05

want a doctor to look like I kind of

19:06

only go to them when I'm told to really

19:09

like um

19:10

oh well I've got a an implant that needs

19:13

to be replaced every three years and I

19:14

get letters saying you have to go for

19:15

this test

19:16

it's not often I get ill or think I've

19:20

got something that says I've got

19:22

symptoms I should go to visit my GP or

19:24

sit in an emergency room so I don't know

19:26

I've not had a lot of experience on that

19:28

front to be able to say what I'd want

19:31

them to look like well like most of what

19:32

I see is what I see on TV that's

19:33

probably not very representative Gray's

19:35

Anatomy yeah Antonio yeah I've I have

19:39

been Cinder docs for about all sorts and

19:42

also just when I had a niggle you know

19:44

when I wasn't really sure what what it

19:47

is but it had some sort of measure

19:50

measurable impact on my life like you

19:53

know can't sleep as well because of said

19:56

sad thing

19:58

um and they do take a while to to get to

20:01

the bottom of it you know there's a bit

20:03

of trial and error that because

20:06

you know it it's not always obvious what

20:10

it is and I think I think part of it was

20:12

drawing covert as well where we didn't

20:15

have face-to-face interaction and so

20:18

there was a little bit of like well we

20:20

had um asked my GP so you could send

20:22

photos which felt really weird like here

20:25

here's this weird patch of my elbow can

20:27

you please see what it is I know black I

20:30

can't really see anything it doesn't

20:32

look that bad and I'm trying to describe

20:34

it with my vague non-common medical

20:37

terms because I feel like you know

20:39

doctors have that like measure of is it

20:42

a stabbing pain is it a throbbing pain

20:45

is it a sharp pain and I'm like I don't

20:48

know I've never been stabbed yeah

20:55

this is my own experience of pain

20:58

so yeah it's almost like you need a

21:00

doctor to be able to interpret your

21:04

human foibles because like if an AI is

21:07

only very good at understanding very

21:09

precise instruction

21:11

and we are imprecise with our uh

21:15

descriptions then they would have to

21:18

figure it out you know but I guess the

21:21

argument is that over time with enough

21:23

data put in they would learn that you

21:26

know the majority of people when they

21:27

say this they mean this and if it's

21:29

unclear then they ask clarifying

21:30

questions I can see that getting really

21:32

frustrating though it's but like on that

21:34

the point Antonio made about pain if

21:36

you're sickness saying well it's it's

21:37

kind of starving but I don't know I also

21:39

throbbed as well and it's like which is

21:40

it is it stabbing or throbbing pick one

21:42

but yeah and I keep I feel like I'm

21:44

defending AI a lot but then you would

21:46

train it to say okay that's it let's

21:49

move on from that question because

21:50

that's exactly what a doctor would do is

21:52

they would just say you're getting

21:53

frustrated by this let's move on let's

21:55

just let's talk about something else

21:56

it's an interesting point though can you

21:58

train AI to be I was going to say more

22:01

human than a doctor with no empathy

22:05

actually that's what I want to say

22:06

though I mean I I remember reading

22:09

somewhere but it might be apocryphal so

22:11

please forgive me

22:13

um but surgeons have a higher rate of

22:16

psychopaths than any other profession

22:20

um because they like the stabbing

22:23

get theme now

22:27

um potentially I mean we teach set

22:30

phrases to medical students on how to

22:33

respond to uncomfortable

22:36

outpourings from patients uncomfortable

22:39

for the patients I'm uncomfortable for

22:41

them so you say things like that must be

22:43

really hard or I can see why that would

22:46

be difficult for you or yes I I get why

22:49

that's frustrating like you you teach

22:51

them very set phrases but if you don't

22:53

apply them correctly with the correct

22:56

like vocal tone and with the correct

22:58

facial expression it does come out like

23:01

AI is just giving you yes I can see why

23:03

that's hard oh that did sound a bit

23:04

patronizing

23:08

talking about training them and pouring

23:12

information into them you mean the AI

23:14

not the DNA doctors yeah no not the

23:16

junior doctors no

23:18

um trading Ai and trading and pouring

23:20

data into AI

23:22

how much data is enough how much data is

23:26

too much

23:27

how much of our private lives should we

23:32

be putting into these machines how

23:36

secure is it

23:38

um who has access to it who has the

23:41

right to read it

23:43

um I mean and at what point

23:46

are we allowed to just not know stuff

23:49

about ourselves like if you for example

23:52

if we put people's DNA and family

23:54

history and medical history and

23:57

everything else in that

23:58

at what point it might come up saying oh

24:01

you're at risk of x y and z

24:03

do they have to know that do I have to

24:05

know that oh what if it knows it and

24:07

tries to stay towards it because you

24:09

don't want to know yeah yeah

24:12

that's a very cyclical argument yeah I

24:15

guess a good example of what you're

24:16

saying is there's a TV program it's on

24:18

Netflix I think called The Bold type

24:20

it's about three youngish women making

24:22

their way in publishing in New York

24:24

um and one of them says um I think my

24:27

mother might have had breast cancer and

24:29

she ended up getting herself checked to

24:30

see if she had the gene and she did and

24:33

she got a bit paranoid about it it was

24:34

doing all these checks and whatever else

24:35

and eventually decided to get a

24:37

mastectomy so she didn't have to worry

24:38

about it and she thought that would

24:40

solve the problem but she obviously got

24:42

implants to replace what had been

24:43

removed and then she started to feel

24:45

like it wasn't her own body and she was

24:46

uncomfortable and there was nothing she

24:48

could do about it and I think if I were

24:51

in that situation I wouldn't want to

24:52

know I'd rather take my chances because

24:55

just because she have a gene doesn't

24:57

mean you will get cancer right there are

24:59

other factors at play but why if it

25:01

becomes so good at predicting that it

25:03

just knew because it's had so much

25:06

information that the chance of it being

25:09

wrong was so little if it could say with

25:12

like 90 certainty that given your

25:15

environment your habits your genetics

25:18

and all this other information that's

25:20

not directly relevant to me well I have

25:22

to sit there and think well is it if I

25:24

am going to get cancer and I can avoid

25:26

it maybe I should maybe I wouldn't want

25:27

to I don't know maybe I'd be happy with

25:29

waiting for that point and then having

25:32

removed when the time came I don't know

25:33

I suppose it's on the one hand you have

25:37

a right to know and make an informed

25:40

Choice with your own body on the other

25:43

hand it's the right to be ignorant

25:45

because ignorance is bliss

25:47

and if you know that you're likely to

25:51

die of a heart attack at 75

25:54

that's so far away but it's always going

25:56

to be hanging over your head if you know

25:58

now know that

26:01

um compared to if you were just

26:02

blissfully up until the age of 75 living

26:04

your life and then you know one day you

26:06

just don't wake up on a lighter note I'm

26:09

sorry I was going to continue down that

26:11

model in veins

26:13

this kind of thought process was applied

26:16

in the Black Mirror episode hang the DJ

26:19

which was where a couple

26:22

had an app and they got a number and it

26:26

turned out that number represented how

26:28

long the app expected them to be

26:30

together

26:31

so if you knew the number right from

26:34

when you met how would you play out that

26:37

relationship would it change the way you

26:39

treated that relationship absolutely

26:41

absolutely if if an app told me or

26:44

you'll be together five years and then

26:46

break up I'd be like no I'm just not

26:48

gonna bother

26:50

it's been five years in a relationship

26:52

that's gonna end see I'd go the other

26:54

way and I'd want to know why it was

26:55

saying that and try and beat that number

26:56

I said you're not telling me it's gonna

26:58

last like this it's going to be

26:58

different it has to be different I

27:00

refuse to believe that's true

27:04

but and maybe maybe on on the side of

27:08

things is that a sign that there are

27:10

some people like me who will put their

27:13

entire faith in what the computer tells

27:16

them when it could be wrong

27:18

and are we you know are there people out

27:22

there who will say oh I'm gonna die age

27:24

50 of Ms no I'm gonna kill myself at 40.

27:27

so I don't have to live that last 10

27:30

years and what if it's wrong you know

27:32

how how much Faith are we going to end

27:34

up putting in Ai and that is that is a

27:36

danger that we shouldn't believe that

27:41

computers are perfect AI is perfect

27:43

because ultimately it is based on human

27:47

fallibility so I guess you'd want some

27:49

way of also training it to see when it's

27:52

been fed something based on something

27:54

that's a fallacy I really don't know how

27:56

you do that though I I mean I guess you

27:58

could always just you'd do it with the

27:59

caveat of like the same way that people

28:01

oh that's just my opinion though of you

28:04

know you have a caveat of this is AI and

28:06

we cannot promise that it's perfect or

28:08

whatever and in the same way that you

28:10

know whenever I'm counseling patients I

28:12

always say based on the information that

28:14

I have this is what I think is going on

28:15

therefore this is what I would recommend

28:17

and there will be risks with everything

28:19

and it might not turn out like this you

28:21

might end up having this problem you

28:22

know

28:23

doing thorough counseling on the risks

28:26

and benefits of AI opinions and are they

28:28

better than human opinions or are they

28:31

still just opinions I guess from my side

28:35

I always see my opinion as just that

28:37

just my opinion this is my professional

28:39

opinion but do patients ever see the

28:42

advice the doctors give them as an

28:43

opinion do they see it as gospel that is

28:46

true I guess it depends on your

28:48

experience with them I think some some

28:50

people just straight up don't listen if

28:52

it's not what they wanted to hear right

28:54

like you know you've been told you

28:58

should cut this out of your diet because

29:00

it's increasing your risk and people

29:02

don't like people know doctors are right

29:05

but ultimately they don't follow it

29:07

ultimately they want to eat that

29:08

lambdish even though it's going to make

29:10

their gout worse yes yes

29:13

this friend who spent a long time trying

29:17

to get this swollen ankle diagnosed and

29:21

it makes me wonder if they had AI well

29:24

the doctor had AI would it have spotted

29:26

it faster because it was like sooner

29:29

than the typical time that you would get

29:32

this condition and also he didn't like

29:35

going to the doctor because he couldn't

29:37

walk so you just go ah it's a bit

29:41

awkward now and then you know you know

29:43

when you you go to a doctor and you say

29:45

oh I need to see a doctor about this ah

29:48

well the next appointment's like two

29:49

weeks away and by the time that two-week

29:52

appointment comes up you're like it's

29:53

gone but it comes back and so you'd book

29:55

another appointment and then it's gone

29:57

again and so they never actually get to

29:59

observe it yeah

30:01

um that's a really interesting point in

30:03

the this friend of ours who's been

30:05

diagnosed with gout

30:08

he is very young to have gotten gout he

30:12

is not the typical sort of person to get

30:15

gout either so usually you associate a

30:17

gout with older white men generally who

30:21

drink a lot eat a lot of meat like think

30:24

like old-timey ships captains that's

30:27

generally Henry VII or Henry VIII yeah

30:30

exactly and that's generally who gets

30:32

gout and he doesn't fit any of those

30:34

stereotypes so would a

30:38

computer that's purely basing it on

30:41

pattern recognition ever come to that

30:43

diagnosis

30:45

um would it come to it quicker based on

30:48

some unknown variables that we that we

30:51

haven't inputted yet so the

30:52

histopathologist I was talking to they

30:55

were saying that actually the AI now

30:57

uses criteria that they that they don't

31:00

know it's using for it it's hard to

31:03

describe but basically they inputted the

31:05

criteria that they've teached that

31:06

they've taught trainees and it's also

31:09

using some extra criteria that it can't

31:12

explain that it's using so would

31:15

eventually AI get to the point where it

31:18

would come to these conclusions quicker

31:20

because it has picked up extra patterns

31:24

that we don't see as clinicians but then

31:26

could that ever be the absolute truth

31:28

because you couldn't explain it well I'm

31:30

just wondering if this goes back to what

31:31

Sarah was saying about you wouldn't just

31:33

rely on the AI alone for various reasons

31:36

so if you can get it to explore blame to

31:39

you so it can teach you what it's

31:41

looking out for yeah because I guess

31:43

that's something with people like well

31:44

that looks like cancer to me but I can't

31:46

explain why that's cancer I just know

31:47

that it is trust me this is my opinion

31:50

yeah

31:51

trust me I'm a doctor

31:54

I'm imagining this 200 300 years in the

31:58

future is what I'm imagining where you

32:00

have ai practitioners who are operating

32:02

independently and at that point I do

32:06

wonder whether it would come to these

32:08

sorts of realizations sooner or whether

32:10

I think definitely in the short term it

32:12

probably would never have considered it

32:14

as a differential because it doesn't fit

32:17

the standard patterns I guess there are

32:20

always going to be outliers in the data

32:22

because our understanding of the human

32:23

body and well the entire world isn't

32:25

perfect yet anyway there are still lots

32:27

of things that we don't know and there

32:29

are still lots of extra patterns to spot

32:31

and I guess that's also where some of

32:33

the biases come in healthcare already

32:34

like for example um women's problems are

32:37

often underlooked people of ethnic

32:39

minorities in a white country have

32:41

particular likely outcomes that white

32:44

people don't get um and that again it

32:46

goes back to what you're saying about

32:47

training it to already have bias because

32:50

the data set is limited yeah

32:52

absolutely to end on a slightly lighter

32:56

and more futuristic you know I've heard

32:58

little bits about advances in surgery

33:00

and like doing things that you would

33:02

never have thought possibly before like

33:03

using little robotic arms to do things

33:05

and

33:06

microscopic images so you can see things

33:09

with Incredible detail is that something

33:11

that you combine with AI so you've

33:13

basically got a robot doing surgery on

33:14

people with no human intervention so I

33:17

spoke to a couple of my colleagues so I

33:19

spoke to one person who is the head of

33:21

Robotics at Leicester

33:23

um University Hospitals of Leicester

33:25

um and he was saying that the main issue

33:28

for humans is there's no haptic feedback

33:31

during robotic surgery because you're

33:34

essentially it's it's like piloting

33:39

um a remote control car

33:41

there's no like you can't feel when the

33:44

car goes over a bump or you know Turns

33:46

Upside Down you just have to see it and

33:49

then base it on what you're seeing this

33:51

is why I don't like computer games no

33:53

feedback at that time I obviously need

33:55

the haptic feedback and not just the

33:56

visual yeah whereas with laparoscopic at

33:59

least you get some haptic feedback

34:01

because you're directly touching stuff

34:02

so sorry for laparoscopic it's Keyhole

34:04

surgery so using a camera and looking

34:07

inside using very small holes but you're

34:09

still touching things directly whereas

34:11

robotic it's a computer attached to that

34:13

equipment and then you're you know I

34:16

don't know a couple of meters away

34:18

controlling that robot with like a

34:20

little like joystick and two two

34:23

joysticks genuinely two joist joysticks

34:25

that move in three directions that you

34:27

can control all of the different

34:29

um instruments with so you've got no

34:32

idea if you're touching something that's

34:33

quite squishy yeah and it's all based on

34:36

your prior knowledge of what

34:39

uh what anatomical structures are where

34:42

which you could easily program into AI

34:45

or you could map it you know in in this

34:48

futuristic world you don't even need to

34:50

give it an estimation you could just

34:51

scan them true very true yeah you could

34:54

do an MRI and then you could program

34:55

that into a robot and away it goes

34:58

that's very true actually in which case

35:01

if you were to do that that would

35:03

actually overcome quite a lot of the

35:04

stuff quite a lot of the arguments

35:06

against

35:07

um Ai and robotic surgery because the

35:10

main problem is having the confidence to

35:14

cut something when you're unsure what it

35:15

is so when you lift something up for

35:18

example and you go well I know it can't

35:20

be this I know it can't be that and this

35:22

other really important structure I know

35:24

it can't be any of that so it can't it's

35:26

nothing important so I can just cut it

35:27

and that is

35:29

a bit of confidence

35:32

the idea of non-important things in your

35:34

body it's a bit

35:36

but yeah like the appendix yeah but the

35:40

appendix wouldn't be there but yeah

35:41

let's just cut it out fans don't need it

35:44

sorry anyway

35:47

so it's having that confidence to know

35:49

that there aren't any vital structures

35:51

that you're going to go through basically

35:53

and would and AI ever be able to

35:57

independently make those sorts of

35:59

assumptions and I suppose eventually you

36:02

would get the technology to the point

36:04

where you could do an MRI scan

36:06

but you would have to have a surgeon

36:09

label every single structure as cut hair

36:13

don't cut hair important not important

36:16

because I mean it's even been it's been

36:18

I mean anecdotally found that nurse

36:21

practitioners nurse surgeons nose

36:24

operators don't have the the confidence

36:27

to cut without having a surgeon there

36:29

saying yes to take it

36:31

and having

36:33

a person there to take that risk for

36:35

them I feel like AI probably wouldn't do

36:38

risk very well yeah I can see why you

36:41

know people have that kind of measure of

36:44

how much risk they're willing to take

36:45

whereas an AI we kind of say like you

36:48

know if it was a programmer they might

36:49

say okay let's say

36:51

99.99999

36:53

absolute but then do we get in a in a

36:57

world where these robotic surgeons go

36:59

into the operating theater and then go

37:01

yeah that's a little that's a little bit

37:03

too borderline for me yeah exactly

37:06

everyone get back out we're coming out

37:09

this yeah and I mean risk in surgery is

37:12

is I mean basically the whole premise of

37:14

surgery is risk management so what what

37:17

is the risk if we don't do the operation

37:18

what is the risk if we do what is the

37:20

risk if we do this specific operation

37:22

versus a different type versus if we

37:24

just do this or we just do that

37:26

um

37:27

so I I think risk management is going to

37:31

be a big area that AI will struggle with

37:34

um

37:35

but then there's other stuff that's

37:37

taking out uh surgery as a required

37:41

specialty which obviously we won't go

37:43

into here uh future episodes maybe

37:45

especially the the idea about risk which

37:47

is the very first episode talking about

37:49

what risk means to us do it over two

37:51

years ago now yeah and we didn't go into

37:52

that much depth though wouldn't AI be

37:54

able to judge risk

37:55

um

37:57

to sum up my my opinion on this whole AI

38:00

thing

38:01

um AI in healthcare AI is built and

38:05

trained on imperfect data from imperfect

38:09

humans

38:10

therefore expecting Perfection from AI

38:13

is impossible we shouldn't expect

38:16

Perfection from Ai and there is no such

38:19

thing as

38:21

perfection in the human world I also

38:24

don't expect Perfection from your

38:25

doctors

38:26

well that just reminds me of the Chaos

38:28

Theory episode where we said you know

38:30

with enough variables and enough

38:32

computing power we could almost put a

38:36

theory to predict something like this

38:39

but will we ever get to that point where

38:41

we have enough computing power to do

38:42

that good question what do you recommend

38:46

for people trying to train AI to be a

38:50

better doctor because you know there are

38:52

pros and cons to it we've talked about a

38:54

lot of con but they're they seem to be

38:56

if we could work it out you know we're

38:58

very early stage in it so what do you

39:00

think should be taken into consideration

39:01

today for the future what would be the

39:07

ideal clinician and then from that build

39:12

the AI into that so you don't just think

39:15

about a diagnostician because we're

39:17

basically there already with AI you also

39:19

have to think about personalized

39:21

management plans you also have to think

39:23

about communication with patients you

39:25

also have to think about

39:26

the ai's own resilience and ability to

39:29

deal with the workload you have to think

39:32

about the way that it interacts with

39:34

other healthcare workers and the system

39:36

in general that is probably what I'd

39:38

recommend is think about what your ideal

39:39

is and build the AI towards that in a

39:42

holistic way so like any good project

39:45

management yeah exactly great Okay so

39:49

we've covered a lot of different things

39:51

about Ai and how it could impact

39:54

Healthcare today or how it already

39:56

impacts Healthcare thank you Sarah for

39:59

sharing your experiences you're welcome

40:00

I hope we continue to have great

40:03

conversations like this in other

40:04

episodes

40:06

the views expressed in this podcast

40:07

belong entirely to the person that said

40:09

them they do not represent any industry

40:11

or organization if you enjoyed listening

40:13

to these views it would really help us

40:14

out if you could rate US leave a review

40:15

and tell a friend this podcast was

40:17

sponsored by no one but if you're

40:18

interested in funding us to continue to

40:20

have Frank discussions about science and

40:21

engineering please get in touch

40:23

[Music]