0:00

End of life care can be

0:02

really quite tricky and people may

0:04

have heard of advanced care directives

0:06

but only 14% of older Australians

0:08

actually have one in place. Whilst

0:10

they can help guide end of

0:12

life care there can be some

0:14

complexities and often all situations aren't

0:16

covered so they're impossible to predict.

0:18

Like are you going to accept

0:21

antibiotics if you've got a sore

0:23

ear? You can't necessarily think of all

0:25

of it and this is where artificial

0:27

intelligence may have a role. Who knows? Our

0:29

producer Shelby Trainer has looked into

0:31

this, the role of AI in

0:33

this space and, as you alluded

0:36

to earlier Norman, there are some

0:38

convincing arguments here. I think it's

0:40

really important, having someone that can

0:42

speak in your behalf, that knows you

0:44

well, that knows how you'd like to

0:46

live your life and what sort of

0:48

things you would or wouldn't want at

0:50

the end of your life. Nicola

0:52

Champion is more familiar than most

0:55

people with the end of life.

0:57

She's a pallied of care nurse

0:59

who also took care of her

1:01

dad in his final weeks in

1:03

their hometown of Port Piri. Dad

1:05

was an interesting character, very independent,

1:08

very strong views on things, great

1:10

sense of humour. Nicola's dad Charlie

1:12

had prostate cancer but he'd lived

1:14

a relatively healthy and independent life.

1:16

until two years after his diagnosis,

1:18

when he arrived on Nicola's doorstep

1:21

out of sorts. And I opened the

1:23

door and he said, I thought you were

1:25

going to look after me when I got

1:27

sick. And I was really taking him back

1:30

because I thought, well, yeah, I will. It

1:32

turned out he was in a lot of

1:34

pain. Charlie's cancer had spread. He would have

1:36

six weeks left if he didn't do anything,

1:38

six months if he did. And my dad

1:40

just said, I'll take the six weeks. He

1:43

never once said, why me? He's just so

1:45

pragmatic and he was just like, well that's

1:47

it, that's my lot, I've got six weeks.

1:49

Charlie had an advanced care directive

1:51

that Nicola helped him write out.

1:53

An advanced care directive lets doctors

1:55

know what you do or don't

1:57

want at the end of life.

2:00

Would you want to be given antibiotics?

2:02

Would you want a blood transfusion if

2:04

it could extend your time? I don't

2:07

remember so much the conversation, but what

2:09

I know is that he fully trusted

2:11

me. So we didn't go down to

2:13

every scenario that he might want to

2:16

consent to or not consent to. He

2:18

just appointed me as his medical decision

2:20

maker. surrogate decision-makers are people you legally

2:22

designate to make decisions for you in

2:25

the event you become incapacitated. For example,

2:27

if you're in a coma or aren't

2:29

cognitively aware of what's going on. So

2:32

for me it was about really knowing

2:34

my dad's values in life and the

2:36

way he liked to live and the

2:38

way I think he wanted to die

2:41

and so that's what was going to

2:43

guide my decisions. Not everyone will have

2:45

a surrogate decision-maker. And even if they

2:48

do, those surrogates aren't always available in

2:50

an emergency. And they don't always feel

2:52

equipped to make tough calls. It can

2:54

be incredibly stressful to be the voice

2:57

of someone who can't speak for themselves.

2:59

It happens so frequently in the ICU where patients

3:01

are critically ill and can't make decisions for themselves

3:03

and you see their family and their surrogate decision

3:05

makers struggling with the burden of that choice. This

3:07

is Dr Teva Brenda, an internal medicine resident at

3:09

the University of California, San Francisco. He saw this

3:11

situation play out almost daily. The patient has a

3:14

breathing tube, they can't speak, or they're so ill

3:16

that they're confused, and every decision becomes this huge

3:18

pivot point. Do we continue antibiotics? Do we pursue

3:20

this procedure or this surgery? Only about 14% of

3:22

older Australians have an advanced care directive. So life-saving

3:24

or life-ending decisions can come down entirely to a

3:26

family member. And there's a lot they need to

3:28

know to make an informed decision. Dr. Brenda and

3:30

his mentor wondered if there was a better way,

3:32

a less stressful and more accurate way, to come

3:35

to these decisions. And so we were just chatting

3:37

together and... AI has obviously been in the news

3:39

since 2020 with ChatGPT and we thought, well, how

3:41

could we leverage artificial intelligence to help surrogates? Because

3:43

this is such a common problem in the ICU.

3:45

They decided to put an idea out there, drafting

3:47

a paper that theorized how generative AI or large

3:49

language models like ChatGPT might be used. Eventually they

3:51

brought in a geriatrician and palliative care doctor to

3:53

get his perspective. It's kind of funny, his first

3:56

reaction to our proposal was, heck no, that's frightening,

3:58

that's dystopian, do we really want to be considering

4:00

this? But Dr. Brenda and his colleagues are far

4:02

from the only people thinking about this. My name

4:04

is Brian Erp and I'm an associate professor of

4:06

biomedical ethics at the National University of Singapore. At

4:08

the same time AI was having its first of

4:10

many moments in the sun, Dr. Earp was working

4:12

as an editor on the Journal of Medical Ethics.

4:14

In that journal, experts were trying to come up

4:16

with better ways to deliver end-of-life care, in line

4:19

with a person's wishes. One proposal that was gaining

4:21

attention was called a patient preference predictor. The idea

4:23

here is that you would do a big survey

4:25

of the population, you would give people various scenarios

4:27

that they might encounter, and you would ask them,

4:29

what would you like to have happen if you

4:31

found yourself in this situation, this situation, this situation,

4:33

and so on? and then you would also collect

4:35

various demographic features about people their age and their

4:37

sex and their ethnic background, their religious affiliation, maybe

4:40

their socio-economic status. So instead of everyone filling out

4:42

an advanced care directive, a sample of the population

4:44

would fill one out, an extensive one. It was

4:46

assumed that people with the same demographic features would

4:48

make similar decisions at the end of life. This

4:50

seems to solve some of the problems, but it

4:52

also was met with a lot of criticisms. The

4:54

biggest one? People don't like being reduced to their

4:56

demographic features. We like to think of ourselves as

4:58

unique, not just in age, sex, and ethnicity.

5:01

Predictor Erb was watching the

5:03

debate about this preference

5:05

predictor play out at the

5:07

same time he and

5:09

his colleagues were playing around

5:11

with AI. My then

5:13

housemate was also my friend

5:15

and collaborator, Sebastian Porstam

5:17

Mann. He realised there were

5:19

these interfaces you could

5:21

gain access to where you

5:24

can further train a

5:26

general model in a process

5:28

called fine -tuning. Basically, you

5:30

can tweak it to

5:32

fit a specific purpose. They

5:34

ended up feeding their

5:36

AI model dozens of their

5:38

own research papers. The

5:40

purpose was to teach the

5:42

model to recognise the

5:45

relationship between a paper's abstract,

5:47

which is a short

5:49

summary of the research, and

5:51

the research itself. And

5:53

once it's learned that relationship,

5:55

you can put in

5:57

a new title and a

5:59

new abstract of a

6:01

paper that you haven't yet

6:03

written, but that maybe

6:06

you plan to write, and

6:08

then you press go,

6:10

and in a matter of

6:12

seconds it will just

6:14

generate a draft of a

6:16

paper in your voice

6:18

using your style of reasoning,

6:20

drawing on the kinds

6:22

of arguments that you've used

6:24

in the past, but

6:27

applying it to this new

6:29

topic. This proved that

6:31

the AI was able to

6:33

learn Brian's views and

6:35

apply that knowledge to new

6:37

situations. So, he thought,

6:39

what if it could be

6:41

used to learn your

6:43

views on things, for example,

6:45

whether you'd want to

6:47

be put on a ventilator

6:50

if it gave you

6:52

six more months of life.

6:54

There was a problem

6:56

though. I work professionally in

6:58

philosophical ethics, so I

7:00

write dozens of papers explicitly

7:02

stating what my views

7:04

are on various topics, and

7:06

so it's not a

7:08

big surprise that this sort

7:11

of model can come

7:13

up with a reasonable guess

7:15

about what I might

7:17

say morally about some situation,

7:19

including a hypothetical situation

7:21

in which I were incapacitated.

7:23

So, we did some

7:25

informal experiments with our academic

7:27

paper model, and we

7:29

just asked it, you know,

7:32

suppose I was in

7:34

X, Y, and Z conditions,

7:36

what do you think

7:38

I would want to have

7:40

happen, and why, and

7:42

the model would typically give

7:44

an answer that's at

7:46

least plausible. So, if you

7:48

haven't written dozens of

7:50

papers on medical ethics, who

7:52

has, how would an

7:55

AI model learn about your

7:57

preferences? Well, one way

7:59

could be through social media.

8:01

There already are existing

8:03

digital duplicates. of us out there in the

8:05

world that are owned by technology companies. So there's

8:07

a digital twin of me that Amazon owns. And

8:09

there's a little twin of me that Facebook owns

8:11

and that Twitter owns and so forth. And they

8:13

use these to predict my preferences in particular domains,

8:16

namely so that they can sell me stuff. Clearly

8:18

it's possible to predict some things about people based

8:20

on information you might think is not obviously or

8:22

directly relevant. Another approach would be to train

8:24

AI on your medical records. He is Dr

8:26

Bender again. So this morning I was

8:28

in our lung transplant clinic. We have

8:30

this technology. It listens to the conversation

8:33

you're having with the patient. And at

8:35

the end of the visit, it summarizes that

8:37

in an after visit summary that the patient

8:39

can take home. And it's really truly impressive.

8:41

So what if we could record these visits

8:44

and capture all of the nuance that happens

8:46

in that 20 or 30 minute visit, all

8:48

of the chit-chat in the door, that really

8:50

speaks a lot about what's important to that

8:53

person, what do they do over the weekend,

8:55

how is their family, etc. Some of these

8:57

things really do seem like they might be

8:59

far-fetched and you might think that you'd have

9:02

a much better go if you focused on

9:04

specific prior treatment decisions people have made.

9:06

So it's one thing to just be

9:08

chit-chatting with my doctor, but if I

9:10

train a model on my electronic health

9:13

records, for example, which show all the

9:15

various other decisions I've made, there's at

9:17

least a question about the extent to

9:19

which I can extrapolate from those kinds

9:21

of decisions to novel cases that

9:23

I won't have yet encountered. But

9:25

imagine this AI could find out

9:27

your vaccination status, or whether you

9:29

went for a more aggressive approach

9:31

to cancer treatment. Dr. Urb has even

9:34

suggested training the AI not passively,

9:36

using data collected from the internet

9:38

or from your medical records, but

9:40

directly, you could take charge of

9:43

building your own digital duplicate. Whether

9:45

an AI should therefore make a

9:47

decision is a different question, because

9:49

it's an ethical question. It's not

9:51

a technical question. And this raises

9:54

another problem, one that AI may

9:56

or may not solve. Nicola was

9:58

familiar with the health... care system.

10:00

She trusted she could make the right

10:02

choices for her dad. But even with

10:04

her expertise and an advanced care directive

10:07

in hand, she struggled to get doctors

10:09

to listen. The way we'd worded his

10:11

advanced care directive was in the terminal

10:13

phase of a terminal illness. or if

10:15

he was in a persistent vegetative state

10:17

that I would then speak on his

10:20

behalf. But when she tried to discharge

10:22

her dad from hospital following a procedure

10:24

so he could die at home, she

10:26

was told he wasn't necessarily terminal because

10:28

they hadn't explored all options. options that

10:30

Charlie had stated he did not want to

10:32

pursue. I just couldn't get anyone to listen

10:34

to me that dad was told he would

10:37

have six weeks, he accepted that he was

10:39

terminally ill, was going to have six weeks,

10:41

that he wanted to be back in the

10:43

country town where we were from, and I

10:45

really had to fight hard to try and

10:47

get him home. It's not good enough to

10:49

have an advanced care plan. People have to

10:51

respect it. And they have to believe that

10:53

if someone's been named as a decision maker,

10:55

that that that person didn't do that lightly.

10:57

Nicholas thinks that if there had been a

11:00

digital duplicate of her dad in the room

11:02

that day, it might have swayed the doctors.

11:04

If they'd asked the AI, okay, what's important

11:06

to you, he would have been able to

11:08

say, I don't want a fast, I want

11:10

my daughter to look after me. But as

11:12

Dr. Earp points out, that all rests

11:14

on AI being proven accurate and being

11:17

trusted, and for that to happen, there

11:19

needs to be a level of transparency

11:21

around how the AI reaches the conclusions

11:23

that it does. Suppose the model says

11:26

it's likely maybe with 80% confidence

11:28

based on all the various factors

11:30

that I've considered that John would

11:32

want to have treatment withheld in

11:34

this situation and some of the

11:36

major evidence I'm using for this

11:38

claim is that John wrote an

11:40

email on January 22nd of 2015

11:42

explicitly saying to his friend Bob

11:44

that if he ever was unable

11:46

to feed himself, he would definitely

11:48

not want to live under those

11:50

conditions. Well, that would be like

11:52

explicit evidence that people could then

11:54

evaluate. They could say, oh, well, let's go check

11:56

with Bob and confirm that that's true. And oh, okay,

11:58

so that seems like a... pretty strong expression of

12:00

his values. I haven't ever heard him

12:03

say anything otherwise, so maybe we should go

12:05

with that. Whatever is the type of

12:07

evidence that it raises is the sort of

12:09

thing that Canon should be then evaluated

12:11

by people who know the person. It might

12:13

be that the person's spouse is there

12:15

and asked to make the decision, and the

12:18

model comes up with some reason, and

12:20

the spouse says, look, that's not anything that

12:22

actually reflects the John I know, and

12:24

then, okay, now you have a difficult discussion

12:26

to make, and you have to do

12:28

some further investigation. I think most people recognize

12:30

this isn't ready for primetime quite yet.

12:32

So we're not saying that this could or

12:35

should be done, but it's interesting to

12:37

think about because that would be ideal, right?

12:39

If there was a future where as

12:41

a physician, I could sit down with the

12:43

family and say, hey, this is what

12:45

this algorithm suggests. Hopefully there is some transparency

12:47

there. How does that sit with you

12:49

as somebody who knew and loved this person?

12:52

And that can sort of be a

12:54

jumping off point. There's a long way to

12:56

go in proving AI decision makers are

12:58

accurate and finding out whether people would ever

13:00

actually trust them to help make life

13:02

or death decisions. Which is why I asked

13:04

Brian what he would do. I have

13:06

thought about this. I mean, a lot depends

13:09

on how predictively accurate these models turn

13:11

out to be once we've trained them and

13:13

tested them. And so right now I

13:15

sort of shrug my shoulders and say, well,

13:17

it kind of depends. If it turns

13:19

out that they're far more accurate on average

13:21

than human surrogates, then I'd be more

13:24

likely to want to have it used in

13:26

cases where I lost capacity. Whether I

13:28

would want to use it if a loved

13:30

one lost capacity and I was the

13:32

proxy decision maker, I suppose that my sense

13:34

of curiosity would be great enough that

13:36

I would be very interested in what the

13:38

prediction was. But then I would also

13:41

want to know that I could trust myself

13:43

enough to not be unduly swayed by

13:45

the prediction if I know that I have

13:47

good reason to trust my own judgments

13:49

about the person. Nicola looks back at her

13:51

time as her dad's carer and is

13:53

glad she could be his voice in his

13:55

final weeks. If it weren't for her,

13:58

it's possible Charlie would have died in hospital

14:00

rather than back at home. I felt

14:02

I knew what was important to my dad

14:04

and I feel like top of his

14:06

list was me to care for him and

14:08

my understanding was for that to be

14:10

in our home. He never specifically said that, but I still think about

14:12

him turning up to my front door. You know, I thought you'd look after me

14:15

when I was sick. And she only felt reassured about her decision when they arrived

14:17

back in Port Piri. He had the biggest smile on his face as we wheeled

14:19

him in that back door, and it's something that I will always treasure, because I

14:21

just thought I know I made the right decision. That

14:24

was Nicola Champion finishing off that story

14:26

by Shelby Trainer and early you heard

14:28

from Dr. Brian Erp and Dr. Tever

14:31

Brenda.