Latest / What's Your Problem? / The Job AI Was Supposed to Kill

Transcript

The Job AI Was Supposed to Kill · What's Your Problem?

• 0:15
  Speaker 1: Pushkin.
• 0:20
  Speaker 2: Ten years ago, Jeffrey Hinton, the computer scientist who's one
• 0:23
  Speaker 2: of the founders of modern AI, came out with a
• 0:26
  Speaker 2: very strong take. He said, and this is a quote,
• 0:30
  Speaker 2: people should stop training radiologists. Now. Radiologists, as you probably know,
• 0:36
  Speaker 2: are doctors who, among other things, read medical scans, X rays,
• 0:40
  Speaker 2: CT scans, MRIs and Hinton said it was just completely
• 0:45
  Speaker 2: obvious again his words, completely obvious that within five years
• 0:49
  Speaker 2: AI would do a better job than radiologists at doing that,
• 0:52
  Speaker 2: at reading scans. And this makes sense. Radiology is largely
• 0:57
  Speaker 2: pattern matching. Does this scan match the pattern of scans
• 1:00
  Speaker 2: that show disease or the pattern of scans that show
• 1:02
  Speaker 2: a healthy patient? And pattern matching is basically what AI does.
• 1:07
  Speaker 2: And yet here we are ten years later with amazing AI,
• 1:11
  Speaker 2: and we still need radiologists. In fact, there's a radiologist shortage.
• 1:16
  Speaker 2: So the question is what happened? Why do we still
• 1:20
  Speaker 2: need so many radiologists, and more broadly, what is the
• 1:23
  Speaker 2: story of AI and radiology tell us about the future
• 1:27
  Speaker 2: of AI and work for the rest of us, those
• 1:29
  Speaker 2: of us who are not radiologists. I'm Jacob Goldstein, and
• 1:44
  Speaker 2: this is what's your problem, My guest, today is Soreb Jaw.
• 1:47
  Speaker 2: He's a radiologist at the University of Pennsylvania. He's written
• 1:50
  Speaker 2: about his own use of AI, and he's also traveled
• 1:53
  Speaker 2: to India and Nepal and has seen the very different
• 1:56
  Speaker 2: ways that AI is being used for radiology in the
• 1:59
  Speaker 2: developing world. To start, Soreb and I talked about why
• 2:02
  Speaker 2: radiologists have adopted AI more slowly than some people expected
• 2:07
  Speaker 2: they would. What was it like for you using AI
• 2:11
  Speaker 2: five years ago? Like when you tried to use AI,
• 2:13
  Speaker 2: you know, specifically sort of what happened?
• 2:16
  Speaker 1: So to give you an example, when I was using
• 2:19
  Speaker 1: AI for a chest sextuory, I found that it slowed
• 2:23
  Speaker 1: me down quite honestly, because what it would do is
• 2:28
  Speaker 1: it would pick up stuff and highlight it. And now
• 2:33
  Speaker 1: I had to do two things. Firstly make my own
• 2:38
  Speaker 1: assessment and then make the assessment with it. AI was
• 2:41
  Speaker 1: right wrong, and whereas before I didn't have to do that.
• 2:46
  Speaker 1: So AI was kind of like that irritating medical student.
• 2:53
  Speaker 1: I didn't mean to say that. They're all very nice.
• 2:56
  Speaker 2: That go on, like what was it like? We'd be like, hey,
• 2:59
  Speaker 2: what about this? Should we worry about this?
• 3:01
  Speaker 1: Yeah? And for most parts it wasn't really finding things
• 3:04
  Speaker 1: that I wasn't finding, and it was finding things that
• 3:07
  Speaker 1: weren't real. It increased my cognitive burden.
• 3:12
  Speaker 2: There is this interesting story from a kind of before
• 3:15
  Speaker 2: the modern era of AI. There was computer assisted mimmography, right,
• 3:22
  Speaker 2: there was There was an earlier version of this, and
• 3:25
  Speaker 2: as I understand it, in that earlier version, in fact,
• 3:28
  Speaker 2: it was unhelpful. Right. It led to more false positives.
• 3:32
  Speaker 2: Women were getting more biopsies, more follow up tests because
• 3:36
  Speaker 2: they maybe had cancer and they didn't, and in the
• 3:38
  Speaker 2: end it was not improving outcomes. Right. So I understand
• 3:42
  Speaker 2: that that's an old story, but like that points to
• 3:44
  Speaker 2: the risk of AI. It's not always good to have
• 3:47
  Speaker 2: more flags raised.
• 3:50
  Speaker 1: Yeah, I'm glad you raised that because that does address
• 3:53
  Speaker 1: two very important things. You know, we think about physicians
• 3:57
  Speaker 1: as sort of people that spot diseases and as somebody
• 4:03
  Speaker 1: that I trained a long time ago, but not that
• 4:07
  Speaker 1: long time ago. The easy bit was finding disease. The
• 4:14
  Speaker 1: easy bit was actually finding people that were very ill.
• 4:17
  Speaker 1: The difficult bit was telling somebody that you're not ill,
• 4:22
  Speaker 1: that you're actually normal. So appreciating what normal is appreciating
• 4:27
  Speaker 1: that you don't need to come into the hospital. You
• 4:30
  Speaker 1: don't need to be investigated left, right, and center. You
• 4:35
  Speaker 1: don't need to be on a bunch of medication. Is
• 4:39
  Speaker 1: probably the most challenging task.
• 4:41
  Speaker 2: I say more about that. Why do you say that.
• 4:44
  Speaker 1: Because when disease looks at you, it stares at you
• 4:47
  Speaker 1: straight in the face. Yea, But for many reasons, when
• 4:52
  Speaker 1: you're at that border zone between disease and not disease,
• 4:57
  Speaker 1: it's very fuzzy. It gets very foggy, and at that
• 5:01
  Speaker 1: point that you have to really instill all your expertise
• 5:05
  Speaker 1: to make sure that it's not that you're not overcalling.
• 5:09
  Speaker 1: So the false positive part is a very big part.
• 5:12
  Speaker 1: And what happened in the past was that the mammography,
• 5:19
  Speaker 1: the computer aided detection of findings was fine tune. You noticed,
• 5:25
  Speaker 1: like a thermostat, you can find tunit to some accent
• 5:29
  Speaker 1: the way you want. And they said, we can't miss
• 5:31
  Speaker 1: a single case of cancer. Can't make that mistake. It's
• 5:35
  Speaker 1: okay to overcall cancer, but we can't have a false
• 5:37
  Speaker 1: negative because that would be disastrous. And so that the
• 5:42
  Speaker 1: radiologist who had the computer aided detection, now it became
• 5:47
  Speaker 1: like a monkey on the shoulder. It kept saying positive
• 5:50
  Speaker 1: positive positive you know, you know the story of the
• 5:53
  Speaker 1: boy who cried Wolf. Right, So when there truly was
• 5:56
  Speaker 1: a cancer, people were like, yeah, are you are you
• 6:00
  Speaker 1: just being AI? Or is this truly cancer? So it's
• 6:04
  Speaker 1: not very helpful if it has a lot of false positives.
• 6:07
  Speaker 2: Right, And in fact, there was a study published in
• 6:10
  Speaker 2: the New England Turnal of Medicine. Right, this is not
• 6:12
  Speaker 2: just anecdotal, like that early system did not help, Like
• 6:15
  Speaker 2: women had more biopsies, more fear of having cancer, but
• 6:20
  Speaker 2: they didn't. They were not healthier, Like the computer was
• 6:22
  Speaker 2: doing more harm than good at that point, whatever it was.
• 6:25
  Speaker 2: And fifteen years ago.
• 6:27
  Speaker 1: I would say that things have improved, okay, And trust
• 6:31
  Speaker 1: is a very important thing. Trust is a major thing.
• 6:34
  Speaker 1: So as AI improves, my trust of using eye increases,
• 6:40
  Speaker 1: and then at some point I just kind of relax
• 6:44
  Speaker 1: and I'm not so uptight with AI and more like yeah,
• 6:47
  Speaker 1: relinquishing some stuff. So we're getting to that point where
• 6:52
  Speaker 1: we can basically start trusting, not fully trust, but start
• 6:56
  Speaker 1: trusting AI and start asking ourselves, starting to yourselves, maybe
• 7:02
  Speaker 1: AI has a.
• 7:02
  Speaker 2: Point, And so specifically, how does that work in your
• 7:07
  Speaker 2: in your practice of medicine? Like what specific instances where
• 7:10
  Speaker 2: AI is helpful.
• 7:12
  Speaker 1: Now, so let me give you one very big example.
• 7:16
  Speaker 1: So a lot of people have lung nodules, which is
• 7:20
  Speaker 1: little dots in their lungs or varying sizes, most of
• 7:25
  Speaker 1: which the vast majority of which are not a problem.
• 7:29
  Speaker 1: So if somebody says to you you have a lung nodule,
• 7:31
  Speaker 1: I wouldn't lose sleep at all. But some of which
• 7:34
  Speaker 1: can become conser rather, some of which are precursor to consent.
• 7:39
  Speaker 1: So a big example is AI flagging them lung nodules. Right,
• 7:45
  Speaker 1: So what that does is it reduces the work on
• 7:48
  Speaker 1: my visual field, like I don't need to It's like
• 7:52
  Speaker 1: trying to find a needle in a haystack. If it's
• 7:55
  Speaker 1: finding those needles in a haystack, I don't need to
• 7:58
  Speaker 1: spend my time looking for those needles.
• 8:01
  Speaker 2: And it needs to find substantially all of them, and
• 8:04
  Speaker 2: not find substantially more than there really are.
• 8:07
  Speaker 1: Yeah, and also also needs to tell me based on
• 8:12
  Speaker 1: the shape, the size and other aspects which can become
• 8:17
  Speaker 1: cancer which can't.
• 8:19
  Speaker 2: This presumably is the harder diagnostic problem to solve, so.
• 8:25
  Speaker 1: Willing to accept a fewers, it's not that difficult. A
• 8:27
  Speaker 1: problem to solve because we can we can ratchet the
• 8:32
  Speaker 1: thermostat to tell us that you know, the vast majority
• 8:36
  Speaker 1: nodules which are not malignant. You say, they're not malignant.
• 8:42
  Speaker 1: We're in good shape.
• 8:44
  Speaker 2: Yes, Well, I mean there's always going to be some
• 8:46
  Speaker 2: number of errors, and the nature of eliminating errors is
• 8:50
  Speaker 2: the more you eliminate, the harder it is to eliminate
• 8:52
  Speaker 2: the next one the more. And so presumably the trade
• 8:57
  Speaker 2: off with AI is it needs to on net make
• 9:03
  Speaker 2: you catch more without taking so much extra time and
• 9:06
  Speaker 2: have so much extra intervention so that the benefit exceeds
• 9:11
  Speaker 2: the cost.
• 9:12
  Speaker 1: Right, Yeah, that is that is pretty much it.
• 9:15
  Speaker 2: And are we there now with with with lung nodules.
• 9:20
  Speaker 1: I think we candy. I think that we out the
• 9:28
  Speaker 1: algorithms that could tell us when there's a lum module
• 9:33
  Speaker 1: and when it's a problem without really compromising on the
• 9:38
  Speaker 1: false positives too much.
• 9:41
  Speaker 2: When you say we can be does that mean it's
• 9:42
  Speaker 2: not quite there yet?
• 9:43
  Speaker 1: Like?
• 9:44
  Speaker 2: Where is this now in your practice? Where is it?
• 9:46
  Speaker 2: More generally? And then presumably there's this sort of sociology
• 9:51
  Speaker 2: piece of it, for lack of a better word, Right,
• 9:53
  Speaker 2: there's does the machine actually work? And then there's do
• 9:56
  Speaker 2: doctors actually use it? And the first is a precondition.
• 9:59
  Speaker 2: The first is necessary but plainly not sufficient for the second.
• 10:02
  Speaker 1: Right, right, And so let's answer your first question about
• 10:06
  Speaker 1: the where are we now? We are still in the
• 10:13
  Speaker 1: relics of the past, which is that we are still
• 10:16
  Speaker 1: obsessed with making sure AI detects correctly and not missus.
• 10:21
  Speaker 1: So we're still in where AI is pointing stuff out
• 10:25
  Speaker 1: by circles and things of that, as opposed to not
• 10:28
  Speaker 1: just pointing it out but putting it in the report.
• 10:31
  Speaker 2: I don't really, I mean, I superficially understand the difference
• 10:35
  Speaker 2: between those two, But why is that an important distinction?
• 10:38
  Speaker 1: Because if you have ten nodules that you flag and
• 10:42
  Speaker 1: you don't put it in the report, then radiologists will
• 10:45
  Speaker 1: have to put in the report.
• 10:47
  Speaker 2: Uh huh, and the ten nodules that it flags. It's
• 10:50
  Speaker 2: saying like you're nodules but they're fine, yeah, and then
• 10:53
  Speaker 2: making the radio just do that just drudgery of typing
• 10:57
  Speaker 2: basically exactly that one. I mean, maybe it's not easy
• 11:01
  Speaker 2: to solve for legal or bureaucratic reasons. Sounds easy to
• 11:05
  Speaker 2: solve that part of it.
• 11:06
  Speaker 1: It shouldn't be difficult it shouldn't be diffult. But you know,
• 11:09
  Speaker 1: you got to see your second question about trust. So
• 11:12
  Speaker 1: trust is something that obviously takes time. But you know,
• 11:18
  Speaker 1: radiologists aren't sort of people that have shied away from technology. Yeah,
• 11:24
  Speaker 1: we embraced technology. That's how we've grown over the years.
• 11:29
  Speaker 2: I made it only exists because of technology, Like presumably
• 11:31
  Speaker 2: you couldn't have had radiologists before you had X rays,
• 11:34
  Speaker 2: which are what one hundred and not two hundred years old?
• 11:38
  Speaker 2: One hundred hundred plus years old?
• 11:40
  Speaker 1: Yeah, I know that old a little awe twenty five. Yeah,
• 11:43
  Speaker 1: so if we haven't embraced it, it's probably because it
• 11:48
  Speaker 1: hasn't moved the needle for us. And if it moves
• 11:53
  Speaker 1: the needle, then it would be self evident. So a
• 11:57
  Speaker 1: big example we have from our own technological development was
• 12:01
  Speaker 1: the time place where you know, medical imaging. They were
• 12:06
  Speaker 1: like films that you would put up on the view boy,
• 12:09
  Speaker 1: like the kind of thing. And then it became digitized,
• 12:13
  Speaker 1: so you no longer did that picture archiving systems sure,
• 12:19
  Speaker 1: and what we call packs or when it became that
• 12:23
  Speaker 1: it was adopted left right center, and nobody was complaining
• 12:27
  Speaker 1: about it and nobody was asking for evidence it was
• 12:31
  Speaker 1: so self evident.
• 12:32
  Speaker 2: Yes, although I mean if you extrapolate that out, it
• 12:36
  Speaker 2: doesn't have the risk of taking your job in the
• 12:39
  Speaker 2: way that AI theoretically might for radiologists at some point, right,
• 12:44
  Speaker 2: that seems like a potentially meaningful difference.
• 12:48
  Speaker 1: It is, But I don't think that anybody is saying
• 12:50
  Speaker 1: no to AI because they are afraid of losing their job.
• 12:56
  Speaker 2: Uh, why are they saying no?
• 12:58
  Speaker 1: They're saying no because it's not currently being helpful. So
• 13:02
  Speaker 1: if you're going to be helpful, you have to either
• 13:05
  Speaker 1: make radiologists do less or do things faster.
• 13:10
  Speaker 2: And is there any domain yet where AI is helping
• 13:13
  Speaker 2: radiologists to either do less or do things.
• 13:15
  Speaker 1: Faster as it stands are right now? No? Yeah?
• 13:22
  Speaker 2: Interesting is there? This is kind of a question for
• 13:25
  Speaker 2: the end, but it's I can't wait because I'm so interested, Like,
• 13:29
  Speaker 2: do you feel like there is some broader lesson? Like,
• 13:33
  Speaker 2: it's extremely interesting to me that this domain that seems
• 13:37
  Speaker 2: like it should be so good for AI, it's still
• 13:40
  Speaker 2: not helping. Can you make any inference that goes beyond
• 13:44
  Speaker 2: medicine based on that fact?
• 13:46
  Speaker 1: That's historically any technology that has come through that is
• 13:51
  Speaker 1: automated it hasn't required the person overseeing the automation to
• 13:58
  Speaker 1: kind of share the process that leads the automation, the
• 14:02
  Speaker 1: scientific calculator and the lock tables. There's a point in
• 14:06
  Speaker 1: time people using block tables at a point in time stopped.
• 14:10
  Speaker 1: But when they stopped, they weren't checking the calculator's output, right,
• 14:15
  Speaker 1: They weren't like, you know, okay, oh, let's just make
• 14:17
  Speaker 1: sure that that six figure calculation it did was correct.
• 14:22
  Speaker 1: And that's what you need to do in radiology, Like
• 14:25
  Speaker 1: if it's going to be AI propelled through, but if
• 14:28
  Speaker 1: every time AI is saying something, I have to scratch
• 14:32
  Speaker 1: my head and say are you right or wrong? Then
• 14:37
  Speaker 1: that's not quite automation. And so if AI is not
• 14:41
  Speaker 1: nine nine nine point nine nine percent good, it's not useful.
• 14:46
  Speaker 1: It cannot be ninety percent because there's ninety percent and
• 14:50
  Speaker 1: there's one ten chance that it could be wrong. I
• 14:53
  Speaker 1: don't know what that one in ten is going to
• 14:55
  Speaker 1: come from. I still have to treat every single of
• 14:59
  Speaker 1: its output as potential one in ten, which doesn't help me.
• 15:04
  Speaker 2: You could imagine some world that is not our world,
• 15:07
  Speaker 2: that is purely governed by reason, where the AI just
• 15:11
  Speaker 2: has to be better than the radiologist. It doesn't have
• 15:13
  Speaker 2: to be ninety nine point nine to nine percent, right,
• 15:17
  Speaker 2: because you know, radiologists are imperfect like all human beings, right,
• 15:22
  Speaker 2: and like all machines, like all probabilistic machines. But we
• 15:25
  Speaker 2: don't live in that world. Right For radiologists to trust AI,
• 15:29
  Speaker 2: the AI has to be way, way better than radiologists.
• 15:32
  Speaker 2: To me, the analogy there is driverless cars. Driveralist cars
• 15:34
  Speaker 2: are plainly already safer than humans, but they're not safer
• 15:38
  Speaker 2: enough for everybody to trust them everywhere. Does that seem
• 15:42
  Speaker 2: relevant in radiology to you?
• 15:44
  Speaker 1: So you know when you say that AI should be
• 15:49
  Speaker 1: better than radiologists, you know, in terms of diagnosis, better
• 15:55
  Speaker 1: is not something that can be easily measured like you
• 16:00
  Speaker 1: would measure faster. If somebody is faster, they're faster, and
• 16:06
  Speaker 1: there's a timing and they're faster. Into radiology, somebody can
• 16:10
  Speaker 1: be better, the better at a certain thing and worse
• 16:14
  Speaker 1: than another thing.
• 16:15
  Speaker 2: I mean, let me, let me give a naive answer
• 16:18
  Speaker 2: and you can tell me why it's not, why it's wrong.
• 16:21
  Speaker 2: Fewer false positives and fewer false negatives.
• 16:25
  Speaker 1: Oh, I mean, if you can do that, great.
• 16:29
  Speaker 2: Like that seems better to me.
• 16:31
  Speaker 1: But yeah, yeah, that is definitely better. Okay, fewer false positives,
• 16:36
  Speaker 1: a few false negatives, but I would also like to
• 16:38
  Speaker 1: see if it can do that. The disease is a spectrum.
• 16:42
  Speaker 1: You know, Disease is not just a monolith. It's it's
• 16:50
  Speaker 1: certain things are quite so obviously diseased and so obviously
• 16:53
  Speaker 1: not disease. So where is it performing the entire range
• 16:58
  Speaker 1: of the spectrum. And if it is, then there's no
• 17:01
  Speaker 1: need for the radiologists.
• 17:03
  Speaker 2: As you're talking about that, I get the sense that
• 17:05
  Speaker 2: you feel like that is some distant world that may
• 17:07
  Speaker 2: never happens, that a correct inference.
• 17:10
  Speaker 1: It could happen. But you know, there are a few
• 17:12
  Speaker 1: things to bear in mind. First is I think there's
• 17:16
  Speaker 1: this idea that AI has these superhuman skills which somehow
• 17:24
  Speaker 1: humans can't figure something out and AI will figure it out.
• 17:27
  Speaker 2: In some domains, this appears to be the case some domains.
• 17:30
  Speaker 1: Yes, you know, this is not.
• 17:32
  Speaker 2: The structure of proteins, notably.
• 17:34
  Speaker 1: Sure. But the biggest problem, and I think this problem
• 17:38
  Speaker 1: recurs quite consistently, is human biology, human pathology. It's not
• 17:47
  Speaker 1: very imaginative. Let's just take something very simple, how disease
• 17:53
  Speaker 1: presents itself right, how disease announces itself, So so got
• 17:59
  Speaker 1: clinical symptoms, it's not very I maginative. Like if you
• 18:03
  Speaker 1: have shoulder pain, for instance, shoulder pain, it could be
• 18:08
  Speaker 1: from muscle problems in your shoulder, right. It could also
• 18:13
  Speaker 1: be because one of your organs in your abdomen has
• 18:17
  Speaker 1: this burst, and it could also be a very uncommon
• 18:24
  Speaker 1: manifestation of a heart attack. So there's something about the
• 18:27
  Speaker 1: body that just doesn't really have much imagination. So you're
• 18:31
  Speaker 1: also saying shoulder pain, three very very different things. Now,
• 18:36
  Speaker 1: of course that's just a symptom. Imaging supposed to clarify
• 18:41
  Speaker 1: things even further, and it does just a large extent,
• 18:45
  Speaker 1: but it also fails. So at some point biology has
• 18:50
  Speaker 1: its own limitation. Biology in the way it announces itself
• 18:55
  Speaker 1: as a certain limitation. And no matter how much you're
• 19:00
  Speaker 1: going to try and decipher that things will look similar.
• 19:04
  Speaker 1: Diseases will look similar, and disease and not disease will similar. Yes,
• 19:10
  Speaker 1: So the reason many reasons why we you know, you
• 19:13
  Speaker 1: wouldn't give it all to AI is because you have
• 19:16
  Speaker 1: to have the second check. And even if that second
• 19:20
  Speaker 1: check isn't going to make much difference, the mere fact
• 19:23
  Speaker 1: of having the second check means that you don't have
• 19:26
  Speaker 1: autonomy in a domain where things are less than hun
• 19:30
  Speaker 1: percent perfect, and the human in the loop is for
• 19:35
  Speaker 1: multiple things, not just simply to make the process more accurate,
• 19:39
  Speaker 1: but to have somebody accountable.
• 19:45
  Speaker 2: We'll be back in just a minute. Tell me about
• 19:58
  Speaker 2: doing X rays at Everest base Camp.
• 20:04
  Speaker 1: Yeah, so, so that is one of a big, big
• 20:10
  Speaker 1: use case for AI. A few years ago we took
• 20:13
  Speaker 1: an X ray machine, which is a portant extra machine
• 20:16
  Speaker 1: with AI capabilities, to the Everest clinic in base Camp,
• 20:22
  Speaker 1: where they didn't have extras before that and then certainly
• 20:26
  Speaker 1: don't have radeologists. And one of the big questions there
• 20:29
  Speaker 1: is does the person have what's known as high stud
• 20:33
  Speaker 1: pro medema. It's it's it's significant because in a high
• 20:37
  Speaker 1: STU pro redeemer, you need to get the person down.
• 20:40
  Speaker 1: That's the only treatment. It's not just enough giving them medication.
• 20:44
  Speaker 1: So what AI has been able to do in that
• 20:47
  Speaker 1: realm is to give assist two people who aren't radiologists.
• 20:53
  Speaker 1: And that's a big use case for AI in multiple
• 20:56
  Speaker 1: domains where you don't have radiologists and sometimes I just
• 20:59
  Speaker 1: don't have physicians.
• 21:01
  Speaker 2: So I read that you did a sabbatical not too
• 21:05
  Speaker 2: long ago at a university in India, and I'm sure
• 21:09
  Speaker 2: more generally if whether during that sabbatical or otherwise, you
• 21:14
  Speaker 2: you saw firsthand either people using AI and settings where
• 21:18
  Speaker 2: there are not many radiologists, or you saw potential for
• 21:21
  Speaker 2: that use.
• 21:23
  Speaker 1: So India is a very interesting country because normally when
• 21:28
  Speaker 1: you talk about technology, there's a drift that happens between
• 21:34
  Speaker 1: high income countries to low medical income countries, where the
• 21:38
  Speaker 1: I is kind of the opposite, where its first use
• 21:41
  Speaker 1: cases really have been in low middle income countries and
• 21:46
  Speaker 1: fighting things like tuberculosis. So tuberculosis is still very rampant
• 21:52
  Speaker 1: in Africa in South Asia such as India, and there's
• 21:59
  Speaker 1: only so far you can go with detecting tuberculosis using
• 22:03
  Speaker 1: the setoscope or looking at the Spoutin further, you have
• 22:08
  Speaker 1: to chess sexually, but it's possibly enough to do a
• 22:11
  Speaker 1: chest sextuary. That chest sextuary needs to be read and
• 22:15
  Speaker 1: read therein then. So imagine you're going around a village
• 22:18
  Speaker 1: in India with a portable chest sextualy mounted in a
• 22:22
  Speaker 1: bus or something, and you take an image and then
• 22:29
  Speaker 1: you send it back to the city and by the
• 22:32
  Speaker 1: time it's read three weeks later, you're like, oh, you
• 22:35
  Speaker 1: know this person is positive, and now you can't find
• 22:38
  Speaker 1: this person. And now this person you know, has TV
• 22:43
  Speaker 1: and may very well be infecting other people, but you
• 22:45
  Speaker 1: can't find this person because it's three weeks later. Imagine
• 22:50
  Speaker 1: if you can give the answer there and then, and
• 22:54
  Speaker 1: that's what AI is doing. So AI embedded in the
• 22:58
  Speaker 1: chest textually is giving the answer there and then, and
• 23:02
  Speaker 1: it's making a huge impact.
• 23:04
  Speaker 2: And so is that happening right now?
• 23:07
  Speaker 1: AI?
• 23:08
  Speaker 2: Yeah, places where there are no radiologists, people are using
• 23:11
  Speaker 2: AI to diagnose TB or rule out KIV.
• 23:15
  Speaker 1: It's it's been happening for the last I would say,
• 23:19
  Speaker 1: at least for the last eight years.
• 23:21
  Speaker 2: And how good is it?
• 23:22
  Speaker 1: How?
• 23:22
  Speaker 2: How how good is it? How good is the AI?
• 23:24
  Speaker 1: Excellent? Excellent? Answering the question with high negative predictive values.
• 23:32
  Speaker 1: So when the AI says no TV, there's no TV.
• 23:37
  Speaker 2: So there are no false negatives they've got They've turned
• 23:40
  Speaker 2: the dial way in that direction. How how is the
• 23:42
  Speaker 2: false positive rate.
• 23:43
  Speaker 1: A false positives? But it's not the only thing. The
• 23:46
  Speaker 1: next thing is they have to have a blood tiss
• 23:48
  Speaker 1: to see if they have a TV or not.
• 23:50
  Speaker 2: How does the false positive rate compare to the false
• 23:53
  Speaker 2: positive rate of the median radiologist.
• 23:56
  Speaker 1: So I would say that if you have a very
• 24:00
  Speaker 1: high end radiologist, like a huge.
• 24:03
  Speaker 2: Expense, that's why I gave you median median.
• 24:06
  Speaker 1: So it's probably no other median. But it's not as
• 24:10
  Speaker 1: good as the expert, or it's just good at the expert.
• 24:13
  Speaker 2: So it's so it's pretty good. It's as good as
• 24:15
  Speaker 2: like an okay doctor. Yeah, it is okay okay, which
• 24:19
  Speaker 2: in a place where there are no doctors or certainly
• 24:21
  Speaker 2: not enough to read the X rays, it is great.
• 24:23
  Speaker 1: Yeah. But you know, if you ask any physician, are
• 24:27
  Speaker 1: you above average? Most of them will say they're above average.
• 24:30
  Speaker 2: It's a problem at some level. It's a problem, right
• 24:33
  Speaker 2: the lake Wobegonen. The fact is medicine. So you're affiliated
• 24:38
  Speaker 2: with the University of Pennsylvania. I assume that means you
• 24:41
  Speaker 2: you teach, you work with residents and NED students. Is
• 24:44
  Speaker 2: that right?
• 24:44
  Speaker 1: That's correct.
• 24:46
  Speaker 2: So I'm curious how younger physicians, physicians in training, how
• 24:52
  Speaker 2: their relationship to AI is different or is it different
• 24:56
  Speaker 2: than yours.
• 24:57
  Speaker 1: So that's a very good question, and I think that
• 25:01
  Speaker 1: gets sounds to the training that is required in the
• 25:04
  Speaker 1: era AI. And you hear all sorts of things about
• 25:10
  Speaker 1: AI and radiologists, AI and physicians, et cetera, et cetera.
• 25:15
  Speaker 1: But the theme that seems to be very consistent is
• 25:18
  Speaker 1: that if you're a real expert, like if you know
• 25:23
  Speaker 1: your stuff, if you know medicine and side out, if
• 25:25
  Speaker 1: you know radiology inside out, AI can definitely make you better,
• 25:32
  Speaker 1: uh huh. But if you're not, if you're starting off
• 25:36
  Speaker 1: your nervous your I wouldn't use the word mediocre, but
• 25:41
  Speaker 1: I'd say inexperienced, there is a big chance AI can
• 25:44
  Speaker 1: make you worse. Huh. And that's because of a phenomenon
• 25:49
  Speaker 1: called automation bias where if you don't know enough. You know.
• 25:56
  Speaker 1: There's the thing about knowledge, which is that there's a
• 26:01
  Speaker 1: certain point. I think it's the Dunark Krugig or Donald
• 26:04
  Speaker 1: Krug come over with the name.
• 26:06
  Speaker 2: Dunning Kruger Rug. People who aren't very good overestimate their abilities.
• 26:12
  Speaker 2: That's the basic that is that right.
• 26:15
  Speaker 1: But you know, as you go along that curve, there
• 26:18
  Speaker 1: comes a point where you know so much that you
• 26:22
  Speaker 1: start knowing where you don't know, like you're actually familiar
• 26:26
  Speaker 1: with your ignorance, right, And at that point, at that
• 26:31
  Speaker 1: level of expertise, you can use AI intelligently because you
• 26:36
  Speaker 1: won't use AI to answer questions that you already know,
• 26:41
  Speaker 1: noting that AI can be wrong in those. Rather, you
• 26:44
  Speaker 1: know so much that you now you know what you
• 26:47
  Speaker 1: don't know, and you can use AI to fill in
• 26:49
  Speaker 1: that gap. What happens on the other extreme, the experience extreme,
• 26:54
  Speaker 1: is that you don't have that knowledge, so you don't
• 26:57
  Speaker 1: know when AI is wrong, and you start relinquishing more
• 27:01
  Speaker 1: and more things to AI because you never had the
• 27:03
  Speaker 1: point where there was no AI, where you were making
• 27:06
  Speaker 1: those decisions on yourself, and you could end up nevi
• 27:10
  Speaker 1: getting to the point where you could question AI.
• 27:13
  Speaker 2: So this is the trap residents, doctors training to be
• 27:17
  Speaker 2: radiologists could fall into, could.
• 27:19
  Speaker 1: Fall into, could fall into. And I don't think we've
• 27:22
  Speaker 1: kind of figured out quite what to do with that.
• 27:25
  Speaker 2: Uh huh. That is an interesting problem that seems to generalize, right,
• 27:29
  Speaker 2: You could imagine that trap in all sorts of fields
• 27:32
  Speaker 2: where I mean even you know, people worry that kind
• 27:37
  Speaker 2: of entry level jobs that senior people view as drudgery,
• 27:42
  Speaker 2: but that are in fact the way you learn the
• 27:44
  Speaker 2: field you kind of get this sort of intuition by
• 27:46
  Speaker 2: doing a thing ten thousand times. Those to the extent
• 27:50
  Speaker 2: they're being automated, seem to present that kind of risk
• 27:52
  Speaker 2: in many fields potentially. But so what do you do
• 27:55
  Speaker 2: about it? You were responsible for training radiologists. How do
• 27:58
  Speaker 2: you help them not fall into this trap?
• 28:02
  Speaker 1: So you do what Both My kids are undergrunds and
• 28:07
  Speaker 1: they're toll very specifically not to use chat gut. They said,
• 28:13
  Speaker 1: if we catch you, you're not going to get your grades.
• 28:17
  Speaker 1: And better than that is just to examine them in
• 28:21
  Speaker 1: scenarios where they can't access it. So I think, I
• 28:24
  Speaker 1: think you know, we're in a trial and error situation.
• 28:28
  Speaker 1: We're in an experimental area. We haven't quite got any
• 28:33
  Speaker 1: precedents for this in the exact manner. But in order
• 28:38
  Speaker 1: to leverage AI, you've got to be in a scenario
• 28:42
  Speaker 1: where you're a complete physician yourself, and you can't be
• 28:47
  Speaker 1: a complete physician if you start using AI early or
• 28:49
  Speaker 1: too early, And so that's that's going to require in
• 28:54
  Speaker 1: the form of training to force people or at least
• 28:58
  Speaker 1: say to people, well, if you use AI, that's fine,
• 29:01
  Speaker 1: but your final exams will be without AI. So get
• 29:05
  Speaker 1: used to doing it without AI.
• 29:08
  Speaker 2: Yeah, I mean, I guess it would seem that in
• 29:11
  Speaker 2: radiology it's easier to give people an exam that is
• 29:14
  Speaker 2: like what you actually do as a radiologist than it
• 29:18
  Speaker 2: is in other domains of medicine.
• 29:20
  Speaker 1: Absolutely, And in fact, the American Border Radiology the one
• 29:24
  Speaker 1: that conducts the exam years ago, when I was doing
• 29:29
  Speaker 1: the exam, it was an oral exam, one on one
• 29:33
  Speaker 1: with the examiner. No AI know nothing, then just scans.
• 29:39
  Speaker 2: They just show your stands and say what is it?
• 29:41
  Speaker 1: Just show scans and on the other side of the
• 29:43
  Speaker 1: scan is somebody with an expressionless face. You don't know
• 29:48
  Speaker 1: when you're right, you don't know when you're wrong. The
• 29:50
  Speaker 1: expression doesn't change. And then they went virtual and they
• 29:54
  Speaker 1: gave questions and multiple choice questions. You still couldn't use AI,
• 29:59
  Speaker 1: but you had no person overlooking that. Now they're going
• 30:02
  Speaker 1: back to the old system, and I think that's.
• 30:03
  Speaker 2: A good thing, so that people can't cheat with AI,
• 30:06
  Speaker 2: so that.
• 30:06
  Speaker 1: People just learn to to people, Yes, the impul thing.
• 30:11
  Speaker 1: Most important thing is that you're you know, you're able
• 30:14
  Speaker 1: to look somebody in the eye and say, hey, you know,
• 30:17
  Speaker 1: I'm reading this scan. Here's what I think it is,
• 30:20
  Speaker 1: and that's the best way to make sure that people
• 30:22
  Speaker 1: are performing like humans. I mean humans have to be humans.
• 30:27
  Speaker 1: I mean to be in the loop, you have to
• 30:28
  Speaker 1: be an expert. How could you be in the loop
• 30:31
  Speaker 1: if you're not an expert.
• 30:34
  Speaker 2: So I want to talk about the future for a minute.
• 30:38
  Speaker 2: And you know, maybe there's like the future that you're
• 30:41
  Speaker 2: worried about in the future that you hope comes to
• 30:43
  Speaker 2: pass in radiology with respect to AI, Like, is there
• 30:48
  Speaker 2: do you worry? Like? Is there is there a future
• 30:51
  Speaker 2: you worry about in radiology?
• 30:53
  Speaker 1: So I'm not worried about my job being taken. I
• 30:57
  Speaker 1: think I'm doing to loose sleep about that.
• 31:01
  Speaker 2: Would you worry if, like you're a kid who's in
• 31:04
  Speaker 2: college wanted to be a radiologist.
• 31:06
  Speaker 1: I'd say I'd say that, you know, you should look
• 31:08
  Speaker 1: at ten years and be willing to pivot. So the
• 31:11
  Speaker 1: thing is what I worry about. I worry about humans
• 31:16
  Speaker 1: being stupid or more stupid than we actually are at
• 31:20
  Speaker 1: the moment, stupid er. And I think that it's important
• 31:24
  Speaker 1: that somehow we don't do much manual work and we
• 31:29
  Speaker 1: go to the gym every now and then, but with
• 31:32
  Speaker 1: the cognitive stuff, you can't just do that do what
• 31:36
  Speaker 1: we do with the biceps. With the cognitive work. We
• 31:39
  Speaker 1: have to keep at it otherwise it really will kind
• 31:42
  Speaker 1: of disappear. So I do worry about that cognitive disappearance.
• 31:48
  Speaker 2: People will stop thinking and get dumber because we have
• 31:51
  Speaker 2: stopped thinking exactly.
• 31:53
  Speaker 1: I really want to use AI to be smarter and
• 32:00
  Speaker 1: to be able to extend my cognitive abilities and to
• 32:06
  Speaker 1: see things that I don't see. And that's fun to
• 32:09
  Speaker 1: see things that I don't see, and to be able
• 32:11
  Speaker 1: to process large amounts of information that I can't process
• 32:15
  Speaker 1: because I don't have the time, I don't have the bandwidth,
• 32:18
  Speaker 1: and I don't have the computational power. And so I'd
• 32:21
  Speaker 1: like to see AI doing that, and I'd like to
• 32:24
  Speaker 1: be able to do that. You know, our knowledge is vast.
• 32:28
  Speaker 1: A lot of things have changed over the years. We
• 32:30
  Speaker 1: used to memorize the bones of the hand, right, the
• 32:34
  Speaker 1: bones of the hand. There was a nice pneumonic and
• 32:36
  Speaker 1: they're always very kind of your cheesy pneumonic. So I
• 32:38
  Speaker 1: can't say it. It's not safe for work.
• 32:41
  Speaker 2: I want you to say it now.
• 32:44
  Speaker 1: It involves a lot of names and the other acts
• 32:47
  Speaker 1: that skateboard lunate tricot piece of form. You probably don't
• 32:53
  Speaker 1: need to learn that. You probably don't need to memorize
• 32:55
  Speaker 1: all of that. So there's definitely some change in how
• 33:00
  Speaker 1: we do things in sort of we can't be doing
• 33:04
  Speaker 1: what we were doing in the seventies, eighties and nineties. Now.
• 33:07
  Speaker 1: Having said that, there's a call, there's a call of
• 33:11
  Speaker 1: being what a physician is. There's a core level of
• 33:14
  Speaker 1: totality of knowledge that's important to have.
• 33:22
  Speaker 2: We'll be back in a minute with the lightning round.
• 33:25
  Speaker 1: M Okay, now we're.
• 33:36
  Speaker 2: Going to do the lightning round. The questions are going
• 33:38
  Speaker 2: to be somewhat different. What was the hardest thing about
• 33:40
  Speaker 2: writing a play?
• 33:42
  Speaker 1: It was accepting that nobody would like it. Did anybody
• 33:48
  Speaker 1: like it? That was pleasantly surprised. And so it was
• 33:54
  Speaker 1: about a doctor in a rural setting that I was
• 33:58
  Speaker 1: trying to stay away from technology from information and got
• 34:04
  Speaker 1: confronted with a snake bite patient who ended up dying.
• 34:09
  Speaker 1: And it was in rural India where they still blame
• 34:12
  Speaker 1: doctors when things go bad. They treat doctors like God
• 34:15
  Speaker 1: when things go well, but they treat doctors like shit
• 34:18
  Speaker 1: when things go bad.
• 34:20
  Speaker 2: Do you think AI will replace writers or radiologists.
• 34:24
  Speaker 1: First, it will replace mediocre writers before it replaces mediocre ideologists.
• 34:30
  Speaker 1: You know. The thing is that do you read the
• 34:32
  Speaker 1: New York Times or the economists? Yes? Do you know
• 34:37
  Speaker 1: who the author is?
• 34:37
  Speaker 2: And the economist only for the oh bits because I
• 34:41
  Speaker 2: love her so much.
• 34:42
  Speaker 1: Yeah, I have trouble because they all sound the same.
• 34:46
  Speaker 1: They all sound the same, but by design, by design,
• 34:49
  Speaker 1: by design, because there's a method of writing that they
• 34:52
  Speaker 1: all encompass. And AI definitely should take over that. I remember,
• 34:59
  Speaker 1: you know, I remember reading people like Christinations, when you
• 35:03
  Speaker 1: would read something a boy Christinations or salam ma Reshti
• 35:07
  Speaker 1: George orwell.
• 35:08
  Speaker 2: Great you were naming great writers.
• 35:11
  Speaker 1: You know that the pros came from those people. Yeah,
• 35:15
  Speaker 1: because it was unique and for better or worse, it
• 35:18
  Speaker 1: had a characteristic And maybe I will take over that too.
• 35:22
  Speaker 1: Maybe chap Chept will able to take that too. But
• 35:25
  Speaker 1: if it comes to uniformity, if it should, definitely chew
• 35:29
  Speaker 1: that up and for breakfast, and I hope it does.
• 35:33
  Speaker 2: What's the most beautiful scan you've ever seen?
• 35:38
  Speaker 1: I'd say that the most beautiful scan, unfortunately is of
• 35:40
  Speaker 1: a fashion that had multiple pathologies. So it was beautiful
• 35:43
  Speaker 1: on an aesthetic sense, but not so much for the
• 35:46
  Speaker 1: patient's sense. It was somebody that had a stroke from
• 35:51
  Speaker 1: a palmary emblass from a clot was called a paradoxical
• 35:56
  Speaker 1: emblass crossing from the right side of the heart to
• 36:00
  Speaker 1: the left side, and as it was crossing over, it
• 36:03
  Speaker 1: was caught in the act crossing over.
• 36:06
  Speaker 2: You saw not just the effect but the cause exactly happening. Yeah,
• 36:09
  Speaker 2: I've seen you mention a book you have called Atlas
• 36:12
  Speaker 2: of Normal rent Invariants that May Simulate disease. Tell me
• 36:16
  Speaker 2: about that book.
• 36:17
  Speaker 1: It's a book that I started reading when I was
• 36:20
  Speaker 1: a resident twenty years ago and I still haven't finished. Keats,
• 36:26
  Speaker 1: not the poet, Keats, the iradiologist kids compiled a whole
• 36:30
  Speaker 1: range of normal on X rays, and normal is really
• 36:37
  Speaker 1: the most difficult thing to learn.
• 36:41
  Speaker 2: Say a little bit more like why is normal so hard?
• 36:44
  Speaker 2: It is interesting and not immediately obvious. Why that is.
• 36:49
  Speaker 1: So because there's so many variations of normal. It's almost
• 36:54
  Speaker 1: like what Tolstoy said about happy family all being seen
• 37:01
  Speaker 1: and every unhappy family being unhappy in their own way.
• 37:05
  Speaker 1: You reverse that disease is a clear disease kind of
• 37:11
  Speaker 1: is very similar homogenous to some extent, whereas normal has
• 37:19
  Speaker 1: a lot more variability.
• 37:21
  Speaker 2: What's your view of healthy people getting a full body scan.
• 37:26
  Speaker 1: I don't bother.
• 37:28
  Speaker 2: Yeah, why because you will end up.
• 37:31
  Speaker 1: Calling overcalling normal, over diagnosing normal. I mean, yeah, there'll
• 37:36
  Speaker 1: always be the success story of one person that whose
• 37:40
  Speaker 1: life was saved, and genuinely so. But for every success story,
• 37:43
  Speaker 1: they're about ninety nine where they'll be just taken down,
• 37:47
  Speaker 1: you know, a rabbit hole for no reason whatsoever and
• 37:51
  Speaker 1: reassured at the end of it and felt like there's
• 37:54
  Speaker 1: some sort of cancer survi not they. I wouldn't bother.
• 37:58
  Speaker 1: But I also think that it's it's it's a decision
• 38:01
  Speaker 1: made by the individual. Some people are risk averse and
• 38:05
  Speaker 1: you have to just take what yours restaurerances.
• 38:08
  Speaker 2: As well anything else we should talk about, well, I think.
• 38:13
  Speaker 1: I think you've covered a lot. I think that people
• 38:16
  Speaker 1: view AI with both. On one hand, you've got this
• 38:23
  Speaker 1: view largely coming from Silicon Valley, and I understand why
• 38:26
  Speaker 1: they need to make money with view of AI has
• 38:30
  Speaker 1: just been this sort of transforming technology. And on the
• 38:35
  Speaker 1: other hand, there's there's a fear and skepticism of it.
• 38:39
  Speaker 1: And I think the truth is somewhere in between. That is,
• 38:43
  Speaker 1: it is a technology that needs nurturing, and it's a
• 38:49
  Speaker 1: very personal technology. I think it'll end up being very
• 38:52
  Speaker 1: personal to the person nurturing it. No two people will
• 38:57
  Speaker 1: view AI in the same way.
• 38:59
  Speaker 2: I appreciate your time. Thank you for being so generous
• 39:02
  Speaker 2: with your time. Thank you very much, sort of. Jaw
• 39:11
  Speaker 2: is a radiologist at the University of Pennsylvania. Please email
• 39:15
  Speaker 2: us at problem at pushkin dot fm. Tell us what
• 39:19
  Speaker 2: kinds of shows we should do more of or less
• 39:20
  Speaker 2: of our particular people you think would be good on
• 39:22
  Speaker 2: the show. You can also find me on x and
• 39:25
  Speaker 2: on LinkedIn. Our show is produced by Gabriel Hunter Chang
• 39:28
  Speaker 2: and Trina Menino. Our editor is Lydia Jean Kott, and
• 39:31
  Speaker 2: our engineer is Sarah Bruguier. I'm Jacob Goldstein and we'll
• 39:35
  Speaker 2: be back next week with another episode of What's Your
• 39:37
  Speaker 2: PALP