Latest / What's Your Problem? / Can Your Phone Tell When You're Getting Sick?

Transcript

Can Your Phone Tell When You're Getting Sick? · What's Your Problem?

0:15
Speaker 1: Pushkin. There are a lot of reasons that I'm excited
0:22
Speaker 1: about today's show. I'm going to tell you three right now.
0:26
Speaker 1: Number One, the show is about this whole dimension of
0:30
Speaker 1: medicine that I essentially didn't know existed, acoustic biomarkers, basically
0:38
Speaker 1: using a person's voice to assess their health. Second thing
0:42
Speaker 1: I'm excited about the show is about the intersection of
0:45
Speaker 1: AI and healthcare, one of my top say five intersections.
0:51
Speaker 1: Love that intersection. And three, today's guest doctor Yael Bensusan,
0:57
Speaker 1: gave me what was truly the best excuse that anyone
1:01
Speaker 1: has ever given me for canceling an interview at the
1:04
Speaker 1: last minute.
1:05
Speaker 2: Yeah, so I'm really sorry for having to cancel on
1:07
Speaker 2: you yesterday.
1:09
Speaker 1: What was the what was the surgery you had to
1:12
Speaker 1: do yesterday?
1:13
Speaker 2: So yesterday we call it airway surgery, where I take
1:17
Speaker 2: a base to the oar and I have to open
1:20
Speaker 2: up their windpipe or their trachia because they're a scar
1:23
Speaker 2: tissue that's blocking them from breathing. So I have to
1:26
Speaker 2: go with a laser and cut the scar tissue out
1:30
Speaker 2: and then take a balloon and open up their windpipe
1:32
Speaker 2: so that they can wake up and breathe better, and
1:36
Speaker 2: that translates to a different sound when they're breathing.
1:38
Speaker 3: So when they're not.
1:39
Speaker 2: Breathing because of the scar tissue, we can sound like,
1:43
Speaker 2: you know, very noisy breathing. We call it the Darth
1:45
Speaker 2: Vader breathing. And then when wake up they wake up
1:49
Speaker 2: from surgery and they're done, they have silent breathing, which
1:51
Speaker 2: means that I know that I did a good job.
2:00
Speaker 1: I'm Jacob Goldstein and this is What's your Problem the
2:02
Speaker 1: show where I talk to people. We're trying to make
2:05
Speaker 1: technological progress. Doctor Yeah. Eld and Susan run the Health
2:09
Speaker 1: Voice Center at the University of South Florida, and she
2:13
Speaker 1: is also leading a team of researchers that's building a
2:16
Speaker 1: giant database of human voices and breadths and health information.
2:22
Speaker 1: Her problem is this, how do you record the voices
2:25
Speaker 1: of thousands of people without violating patient privacy laws while
2:29
Speaker 1: building a giant public database that could someday allow your
2:34
Speaker 1: phone to warn you, based solely on your voice that
2:38
Speaker 1: you may be getting sick. Yeah El told me that
2:41
Speaker 1: she got into this field in part because she used
2:43
Speaker 1: to be a singer.
2:46
Speaker 2: So I growing up, you know, I always was in
2:50
Speaker 2: a very musical family. I took singing lessons when I
2:53
Speaker 2: was a kid, and then I started singing more professionally
2:57
Speaker 2: around eighteen years old, and I had a short but
3:01
Speaker 2: exciting singing career. I wrote pop folk music. We had
3:05
Speaker 2: a bend, and we toured. We had an album out
3:08
Speaker 2: in two thousand and twelve. Yeah, and I mean it
3:12
Speaker 2: was a lot of fun. And actually the reason I
3:15
Speaker 2: was able to have that short and exciting career was
3:19
Speaker 2: because I met a speech pathologist when I was fifteen.
3:21
Speaker 2: So I was taking singing classes and one day my
3:25
Speaker 2: teacher looked at me and she said, there's something wrong
3:28
Speaker 2: with your voice.
3:28
Speaker 3: Go get checked.
3:31
Speaker 2: And I met a laryngologist who put his camera down
3:35
Speaker 2: and she said, you have nodules on your vocal cords
3:38
Speaker 2: and you might not be able to sing again if
3:39
Speaker 2: you don't take this seriously. And I went to see
3:42
Speaker 2: a speech pathologist. I did rehabilitation with my voice for
3:45
Speaker 2: six months, and I was able to sing again. And
3:47
Speaker 2: I mean that's what led me to then become a
3:49
Speaker 2: speech pathologist growing up, and then eventually go to med
3:54
Speaker 2: school and then decide to become a laryngologist.
3:56
Speaker 3: So it was kind of all interconnected.
3:59
Speaker 1: So I know that your research now and most of
4:04
Speaker 1: what I'm really interested to talk with you about is
4:08
Speaker 1: is around acoustic biomarkers. So just to start, I mean,
4:13
Speaker 1: what's an acoustic biomarker?
4:16
Speaker 2: Very good question. So what is a biomarker? First, A
4:19
Speaker 2: biomarker is something that indicates the presence of a disease, right,
4:25
Speaker 2: So if you think about a biomarker for a cancer,
4:28
Speaker 2: so different cancers have different types of biomarker. For example,
4:31
Speaker 2: for ovarian cancer, we're looking for a specific thing you know,
4:35
Speaker 2: called ca in your blood. For different types of cancers,
4:38
Speaker 2: they could take a blood draw and find a specific biomarker.
4:41
Speaker 2: It's an indicator of a disease. An acoustic biomarkers is
4:45
Speaker 2: something that can indicate a presence of a disease, but
4:47
Speaker 2: that you can hear. So that's the definition of an
4:50
Speaker 2: acoustic biomarker. So I always say, you know, when you
4:54
Speaker 2: have people in your family that are not well, you
4:57
Speaker 2: will always notice first and you'll say you don't sound
5:00
Speaker 2: good right, or you sound funny. And I have the
5:05
Speaker 2: luxury to know that because I'm a voice doctor. So
5:08
Speaker 2: then people will bring me their family members or people
5:10
Speaker 2: will come saying, I don't know what's wrong with me,
5:13
Speaker 2: but my wife told me to come because my voice
5:16
Speaker 2: is not good. And sometimes it's because their vocal cords
5:20
Speaker 2: are not working, but a lot of times it's because
5:23
Speaker 2: they can have a neurological issue or a cardiac issue
5:26
Speaker 2: that is affecting their voice.
5:28
Speaker 1: So, more broadly, what's going on with AI and acoustic biomarkers.
5:34
Speaker 2: Yeah, so, so many exciting things are going on. I
5:37
Speaker 2: think that's the first answer. There are so many startups,
5:42
Speaker 2: so many companies, industry researchers, academic researchers that are working
5:46
Speaker 2: and looking into voice AI. And the reason is it's
5:50
Speaker 2: really cheap to collect. Right to think about this, If
5:54
Speaker 2: you have a phone, it's really cheap to collect Compared
5:56
Speaker 2: to this.
5:57
Speaker 1: You don't have to pick a blood sample. You have
5:59
Speaker 1: exactly just you've got the phone. You've got the device
6:02
Speaker 1: literally in your hand already. All you have to do
6:05
Speaker 1: is talk, and you're talking already.
6:07
Speaker 2: And you're talking already, so it's cheap to that's why
6:09
Speaker 2: pharmaceutical industries are also very interested, and there's a lot
6:12
Speaker 2: of pharmaceutical projects around it. So there are a lot
6:16
Speaker 2: of projects that are going on and the state or
6:21
Speaker 2: the The current landscape is that there's tons of people
6:24
Speaker 2: working on very similar things and very interesting and various disease.
6:29
Speaker 2: So I always I kind of categorize them in three
6:31
Speaker 2: categories of diseases that are being studied. One is the
6:36
Speaker 2: disease that affects the voice box. Okay, so vocal court paralysis, absolutely,
6:41
Speaker 2: it's intuitive. There's going to be vocal biomarkers in that
6:44
Speaker 2: voice box cancer, right, that's easy. Then there's a voice
6:48
Speaker 2: and speech affecting disorders, so disorders that don't affect the
6:52
Speaker 2: voice box, but that have an impact on the voice
6:55
Speaker 2: and the speech. Parkinson is one of them, right, Alzheimer's
6:59
Speaker 2: is one of them. A stroke somebody having a stroke,
7:02
Speaker 2: they don't have a problem with their voice box, but
7:04
Speaker 2: their speech is going to be altered. So these are
7:06
Speaker 2: voice and speech affecting conditions. So lots of work is
7:09
Speaker 2: being done in that field. And the third one is
7:12
Speaker 2: diseases that you don't think would affect speech, but still
7:15
Speaker 2: people are doing research on that. So there was a
7:17
Speaker 2: really interesting study on diabetes. They're saying that there was
7:21
Speaker 2: a group that published that they could diagnose people that
7:24
Speaker 2: were diabetic versus non diabetics based on their speech and this.
7:28
Speaker 1: So this third group is one presumably where there's at
7:34
Speaker 1: least the potential for AI to detect differences that even
7:39
Speaker 1: experts like you cannot detect, right, I mean, is that
7:43
Speaker 1: what's going on there? What?
7:45
Speaker 2: So AI is not magical, you know, I think it's
7:48
Speaker 2: It does a lot of things. But what AI does
7:50
Speaker 2: that the layperson doesn't do is that it can analyze
7:53
Speaker 2: a lot more data faster.
7:55
Speaker 1: Yeah.
7:56
Speaker 2: Right, So AI has the possibility, if you have a
7:59
Speaker 2: large data set, to then find small differences in these
8:04
Speaker 2: data sets that we don't have. I mean, I would
8:06
Speaker 2: have to listen to, you know, thousands and thousands of
8:09
Speaker 2: voices and compare them statistically.
8:11
Speaker 1: It might it might, right. It might also be able
8:13
Speaker 1: to detect differences that are not even audible.
8:17
Speaker 2: It could exactly. I can give it an example. There's
8:20
Speaker 2: a company looking at atrial fibrillation, and I cannot validate
8:25
Speaker 2: their data because that's one of the limitations that we're
8:27
Speaker 2: going to talk about. But obviously their data set is
8:30
Speaker 2: not public. But they're saying that they can diagnose atrial
8:33
Speaker 2: fibrillation based on the voice. And their explanation is that
8:36
Speaker 2: our voice vibrates to the sound of our heartbeats.
8:40
Speaker 1: Big if true? Fun if true?
8:43
Speaker 2: I mean you know, again, the limitation here is that
8:45
Speaker 2: it's there's a lot of things you can't validate. But
8:48
Speaker 2: they say that they've been validating it with EKGs and
8:52
Speaker 2: that they can see it. They can hear a difference
8:54
Speaker 2: in the voice between patient patients with a.
8:56
Speaker 1: Fib atrial fibrillation. It puts you at risk for a stroke, right,
9:00
Speaker 1: it can go undiagnosed. So like, if if this works,
9:04
Speaker 1: that would be very helpful to many people, right, absolutely, absolutely.
9:09
Speaker 1: So you're mentioning like that's super interesting. It's it's interesting
9:13
Speaker 1: more generally. So, so you're building a giant database, right,
9:18
Speaker 1: and I find that interesting for a lot of reasons.
9:21
Speaker 1: It happens. I don't have you come across the work
9:24
Speaker 1: of faith A Lee. Absolutely, yeses, So I talked to
9:27
Speaker 1: faith A Lee for this show not long ago. Wow. Right,
9:30
Speaker 1: she's like nerd famous, right yeah, And so you know,
9:36
Speaker 1: as you know, she built this giant database of images
9:40
Speaker 1: about ten years ago a little more now called image net.
9:44
Speaker 1: And that was that giant database was what allowed these
9:50
Speaker 1: early machine learning models AI models to you know, start
9:55
Speaker 1: recognizing images, right, and so the database was this necessary tool,
10:02
Speaker 1: necessary thing for the AI to really work, right, And
10:06
Speaker 1: so are you building the acoustic biomarker version of that?
10:13
Speaker 2: So the first the short answer is yes, but I'd
10:16
Speaker 2: like to start by saying that I am not building
10:19
Speaker 2: it's our distortion.
10:20
Speaker 1: Yes, yes, are you all are?
10:23
Speaker 3: Actually, I'll just.
10:24
Speaker 2: First start by recognizing here that it's it's a it's
10:27
Speaker 2: a huge team. So we're the Bridge to Way I
10:29
Speaker 2: Voice Constortium is a team of fifty investigators across the
10:33
Speaker 2: US and Canada. We're funded by the NIH through the
10:36
Speaker 2: Bridge to Way I program and the goal absolutely this
10:40
Speaker 2: is the first time I hear the analogy to the
10:41
Speaker 2: image net database.
10:43
Speaker 3: I like it.
10:43
Speaker 2: I usually give the example of the genomic database, the
10:47
Speaker 2: Human Genome Project, huge.
10:49
Speaker 1: Project, more famous, more famous, they're.
10:51
Speaker 3: Both very famous. But I like this analogy.
10:53
Speaker 1: Well. Image net is maybe a little bit closer of
10:56
Speaker 1: an analogy, but maybe less Yeah yeah, sexy, yeah.
10:59
Speaker 2: Well, but I mean it's interesting because the genome project
11:02
Speaker 2: has also very interesting ethical particularities like voice, right, the
11:06
Speaker 2: image has a little bit less of the ethical constraints.
11:08
Speaker 3: For is, when we talk about whole genome.
11:10
Speaker 2: Sequencing or genomics data people kind of understand that voice
11:15
Speaker 2: has similar concerns in terms of process.
11:18
Speaker 1: We want to get to the concerns, but I want
11:20
Speaker 1: to first talk about what you're doing and and then
11:23
Speaker 1: we can talk about you know, not doing anything wrong. Yeah.
11:28
Speaker 1: So broadly, if it becomes the thing you hope it
11:32
Speaker 1: will be, what, what is it going to be? What
11:35
Speaker 1: is the bridge to AI voice database going to be?
11:39
Speaker 2: So it's going to be this large database of thousands
11:42
Speaker 2: of human voices linked to other health information that are
11:47
Speaker 2: going to be available to researchers and potentially people other
11:52
Speaker 2: than researchers as well, to be able to make discoveries, right,
11:56
Speaker 2: to learn to use a voice AI, to train you know,
12:00
Speaker 2: the next generation of people on how to learn to
12:03
Speaker 2: build models on voice AI, to help pharmaceutical companies develop
12:07
Speaker 2: products or learn even to to develop products, right, And
12:11
Speaker 2: the other really important thing is to teach people what
12:15
Speaker 2: type of standards we need right right now, a lot
12:18
Speaker 2: of different projects, there's really a lack of standards. People
12:21
Speaker 2: collect voice in different ways. That's why it's really hard
12:25
Speaker 2: to pull data together. So our dream was really to say, like, hey,
12:30
Speaker 2: you want to do voice research, here's a manual, my friend, right,
12:34
Speaker 2: like here is how you collect the voice to make
12:36
Speaker 2: it accurate. This is the protocols with the task that
12:39
Speaker 2: we think, based on our studies, give the best biomarkers. Right,
12:44
Speaker 2: These are the type of biomarkers you can look for
12:46
Speaker 2: and this is the data you can train, so really
12:50
Speaker 2: create a manual of operations also for people to be
12:52
Speaker 2: able to make discoveries, and that's the goal to have
12:56
Speaker 2: the most impact on patient care.
12:58
Speaker 1: So what are the biomarkers? What are you asking people
13:00
Speaker 1: to do? What are you collecting?
13:02
Speaker 2: So I separate things between. So there are respiratory biomarkers,
13:08
Speaker 2: voice biome markers, speech biomarkers, and linguistics biomarkers, and they're
13:13
Speaker 2: all different. So let's go about why these are different.
13:16
Speaker 2: So respiratory is easy, right, So we ask people to breathe,
13:20
Speaker 2: to cough, to take big breaths in and that has
13:24
Speaker 2: a lot of information on our pulmonary capacity, on how
13:27
Speaker 2: our windpipe is shaped. Okay, that's respiratory. Then voice and
13:31
Speaker 2: speech what's the difference. So voice is really the sound
13:34
Speaker 2: that we make when our vocal cords come together. So
13:38
Speaker 2: when we say, like birds can voice, but they can't speak.
13:43
Speaker 2: If you have a bird that speaks, then you'll be very.
13:46
Speaker 1: Rich or you have a parent.
13:52
Speaker 2: So when we when we do voice tasks, we ask
13:55
Speaker 2: patients to say E or.
13:57
Speaker 3: Ah or I.
13:59
Speaker 1: Get the difference.
14:01
Speaker 2: Birds and voice biomarkers will be impacted when our voice
14:06
Speaker 2: box is changed or our resp is changed. Right, So
14:11
Speaker 2: somebody with pneumonia probably cannot hold a note for very long,
14:15
Speaker 2: So that's voice biomarkers. When we talk about speech biomarkers,
14:18
Speaker 2: then you go into articulation. So some people, for example,
14:22
Speaker 2: who have neurological deficits or their mouth is not working correctly,
14:25
Speaker 2: they're going to have trouble articulating. They're going to have
14:28
Speaker 2: trouble saying some words. So these are biomarkers we can extract.
14:32
Speaker 2: And then lastly there's linguistic biomarkers. So what type of
14:36
Speaker 2: words are people using, what type of semantic how fast
14:41
Speaker 2: do they speak for example? These are all different types
14:44
Speaker 2: of biomarkers that.
14:46
Speaker 3: That we can extract.
14:47
Speaker 2: So to give you a very tangible example, I was
14:49
Speaker 2: reading a paper from a group looking at biomarkers of depression,
14:54
Speaker 2: and rate of speech was one of the important biomarkers
14:58
Speaker 2: they found. So people who are sad or depressed will
15:01
Speaker 2: speak at a slower pace, so words per second is smaller.
15:06
Speaker 2: So that's simple when you think about it, it's a
15:08
Speaker 2: simple by marker, right, So that's to give up tangible examples.
15:13
Speaker 2: So in terms of I think I didn't answer your
15:15
Speaker 2: question fully, So what are we asking patients? So we
15:18
Speaker 2: ask people to do all these tasks so coughing, breathing,
15:22
Speaker 2: a e. Then we make them read those validated passages,
15:27
Speaker 2: and we also ask open questions. And then when we
15:31
Speaker 2: ask open questions, we have to ask about questions that
15:34
Speaker 2: make them emotional and some that don't make them emotional,
15:37
Speaker 2: because if you trigger emotion, that causes a bias on
15:40
Speaker 2: how your voice will sound.
15:42
Speaker 1: What what question do you ask to make people emotional?
15:46
Speaker 3: So it's really interesting.
15:47
Speaker 2: So at first we would ask, you know, our first
15:50
Speaker 2: question was, you know, can you talk to me about
15:53
Speaker 2: something that makes you sad? It could be somebody that
15:56
Speaker 2: died in your family or you know, So that was
15:58
Speaker 2: our prompt. And then our question without emotion was tell
16:03
Speaker 2: us about your disease and.
16:07
Speaker 1: Only a doctor. What'd think that's for that emotional question?
16:09
Speaker 3: Exactly?
16:10
Speaker 2: I mean, but it's like when you think about it,
16:12
Speaker 2: like Our consortium is like tons of experts that put
16:14
Speaker 2: their minds together to develop.
16:16
Speaker 1: Tell me about having Parkinson's. That's the unemotional question we're
16:19
Speaker 1: going to ask.
16:19
Speaker 3: And then we I mean, we like, why are you here?
16:22
Speaker 2: I think it was not that obvious, but it's like,
16:24
Speaker 2: tell us about why you're here to see your doctor today.
16:27
Speaker 2: And then analyzing the data, because we do pilots, right,
16:30
Speaker 2: we audit our data. We realized that people were starting
16:33
Speaker 2: to tear up, like we had people crying while talking
16:36
Speaker 2: about why they were coming to the doctor today, which is.
16:38
Speaker 1: Supposed to be the example of unemotional.
16:40
Speaker 3: Sure, correct, So we had to change that.
16:45
Speaker 1: Yes, interesting, So okay, this is great. So you're getting
16:50
Speaker 1: a lot of auditory information from every patient. What other
16:55
Speaker 1: information you're getting from each person? So much?
16:58
Speaker 2: So to give you an idea, our full protocol is
17:00
Speaker 2: about one.
17:01
Speaker 1: Hour okay, so of the patient with the patient.
17:05
Speaker 2: With an ipassion it's an iPad, So everything is based
17:08
Speaker 2: on an iPad and there's a helper right now or
17:10
Speaker 2: research assistant. So we collect data. We collect very extensive
17:14
Speaker 2: demographics in terms of you know, age, race, geographical location.
17:19
Speaker 2: We collect language, So what language do you speak? How
17:22
Speaker 2: many languages is do you speak, what languages do you write?
17:26
Speaker 2: You know, what part of the world are you from?
17:28
Speaker 2: That's really important. Then we collect about disabilities. Are you
17:32
Speaker 2: hearing compared are you visually impaired? Because that makes a
17:34
Speaker 2: change in your voice, your smoking status, your hydration status,
17:40
Speaker 2: your fatigue status, because that's so we're we kind of
17:43
Speaker 2: thought about anything that could affect voice, right, your socio
17:47
Speaker 2: economical status because if you think about it, that's going
17:51
Speaker 2: to affect you know, your linguistics as well. And then
17:55
Speaker 2: so other that extensive demographics, then we collect confounders, so
17:59
Speaker 2: we think about anything that could change your voice. Do
18:02
Speaker 2: you have allergies? Do you do you have dental issues?
18:05
Speaker 2: Do you wear braces? So everybody gets a basic test
18:09
Speaker 2: about if they are depressed. So no matter what disease
18:13
Speaker 2: you have, you kind of get the basic tests for
18:15
Speaker 2: all the other disease to measure if it's possible that
18:18
Speaker 2: you have concurrent diseases at the same time.
18:21
Speaker 1: Because presumably because people are in fact complex, and there
18:24
Speaker 1: are many people who have depression and Parkinson's and you
18:27
Speaker 1: want to understand what's going on there.
18:30
Speaker 2: I mean, most people are complex, right, It's really rare
18:33
Speaker 2: to have and people that go to the doctor are
18:36
Speaker 2: not twenty year old and healthy. Right, most of the
18:39
Speaker 2: people who will use our technology or will benefit from
18:42
Speaker 2: these database will be your typical sixty year old chronic
18:45
Speaker 2: disease patient that comes into the doctor and they're not
18:48
Speaker 2: they don't have a sterile bill of health.
18:50
Speaker 1: How many people do you want to have in the database? Like,
18:53
Speaker 1: is there a final number you're going for?
18:55
Speaker 2: So at the beginning, we were aiming for thirty thousand,
19:00
Speaker 2: which is extremely it's extremely ambitious, I think to be fair,
19:03
Speaker 2: I mean, if after four years we get to ten thousand,
19:06
Speaker 2: I think it'll be a huge success. Okay, And you
19:10
Speaker 2: know the data collection. I think what we're learning is
19:13
Speaker 2: that data collection is very resource intensive. To have good
19:17
Speaker 2: data is very resource intensive.
19:20
Speaker 1: So what happened that made you realize that thirty thousand
19:25
Speaker 1: was maybe harder than you thought?
19:28
Speaker 2: So? I think we thought that we wanted to collect
19:31
Speaker 2: as much data as possible, and our original plan was
19:34
Speaker 2: to collect a lot shorter protocols, you know, like shorter clips.
19:40
Speaker 2: But as we started working with patients, we realized that
19:44
Speaker 2: by getting more data from the same patients, we can
19:48
Speaker 2: actually have a lot more information and it provides a
19:51
Speaker 2: lot of interesting you know biomarkers. So we're focusing more
19:55
Speaker 2: on getting more data from a smaller amount of patients
19:59
Speaker 2: and really with the right data, kind of right data
20:02
Speaker 2: with a lot of clinical information attached to it.
20:08
Speaker 1: After the break, what the world will look like in
20:10
Speaker 1: a few years if everything goes well. So this is
20:22
Speaker 1: a big project that yeah, Elle and her colleagues are
20:25
Speaker 1: embarked on. It's a four year project. They're about a
20:27
Speaker 1: year in and there will be interim data releases along
20:31
Speaker 1: the way. So I asked her, how long will it
20:34
Speaker 1: take for this project to advance the state of the
20:37
Speaker 1: science in acoustic biomarkers.
20:39
Speaker 2: Yeah, I would say to say at the end of
20:43
Speaker 2: the four years would be a probably the best answer.
20:47
Speaker 2: I think at the end of the four years. But
20:49
Speaker 2: I think that you know, you can just say, oh,
20:51
Speaker 2: we'll just start training models at the end of the
20:53
Speaker 2: four year once we have all the data. Right, It's
20:55
Speaker 2: not just about you know, building one model that I'll
20:57
Speaker 2: answer your question, is about continuously training models to understand
21:02
Speaker 2: which biomarkers to extract the then build products that walk.
21:06
Speaker 1: So, so, if things go well, what will this world
21:13
Speaker 1: look like in whatever five years?
21:17
Speaker 2: Yes, So, I mean there's there's a few things that
21:21
Speaker 2: this can help with in general, voice biomarkers. Let's not
21:24
Speaker 2: talk about just our project. Diagnosis is one thing, right,
21:28
Speaker 2: early diagnosis, but that's probably the hardest thing, Huh. Screening
21:35
Speaker 2: is most more important. So when we think about screening,
21:39
Speaker 2: it means you, let's say you live really far you
21:41
Speaker 2: don't have access to a doctor, but your doctor has
21:43
Speaker 2: an iPhone and you can talk into the iPhone and
21:46
Speaker 2: it can say, hey, something's wrong. You know, you need
21:49
Speaker 2: a neurological specialist, for example. So to help screen and triage.
21:53
Speaker 2: I think this probably we're looking at in the next
21:55
Speaker 2: five years, something definitely possible. The other product that I
22:00
Speaker 2: think will be very possible within five years is tracking
22:03
Speaker 2: of diseases. If you want to monitor the evolution of
22:07
Speaker 2: parkinson or how people respond to drugs. That's why pharmaceutical
22:11
Speaker 2: companies are very interested.
22:13
Speaker 1: Right. So the acoustic biomarker is not just a binary
22:16
Speaker 1: signal of disease, no disease. It can tell you a
22:19
Speaker 1: lot about the status of disease. Is it getting better,
22:23
Speaker 1: is it getting worse?
22:24
Speaker 2: Evolution, especially if you train it on your own voice. Right,
22:28
Speaker 2: it's even easier to detect changes in somebody's voice as
22:32
Speaker 2: they progress, like your sory for example, or Alexa that
22:36
Speaker 2: learns listens to your voice. So that's going to be
22:39
Speaker 2: a really good tool for pharmaceutical companies. That's why they're
22:42
Speaker 2: investing in it, right, to see how you respond to
22:44
Speaker 2: a drug, how you respond to a treatment. And when
22:47
Speaker 2: you think about telehealth at home, right, so more and
22:51
Speaker 2: more we're going to talk about remote monitoring people. There
22:55
Speaker 2: were just too many people on this earth to all
22:57
Speaker 2: be in hospitals when we're sick.
22:59
Speaker 1: Well, and if you can stay out of the hospital
23:02
Speaker 1: when you're sick, that's better, Right, You don't want to
23:04
Speaker 1: go to the hospital unless you have to do yeah, or.
23:07
Speaker 2: That you're you know, your Lexa detects when your voice
23:10
Speaker 2: starts detailor rating and sends you a nurse before you
23:14
Speaker 2: need to go to the hospital.
23:15
Speaker 1: So there's a more general version of that one, right
23:19
Speaker 1: that you could imagine, which is you get your whatever,
23:22
Speaker 1: your iPhone, your Android phone, and you have a choice
23:25
Speaker 1: when you're setting up your phone, like do you want
23:27
Speaker 1: to opt into to the phone listening and to tell
23:32
Speaker 1: you if you need to go talk to your doctor, right,
23:34
Speaker 1: just like a very broad based thing that you could
23:39
Speaker 1: opt into, like I would probably opt into that. I mean,
23:43
Speaker 1: is that a thing that you think about?
23:45
Speaker 2: So, I mean yes, I'm sure that you know Apple
23:48
Speaker 2: is working on that already.
23:49
Speaker 3: They are, you know.
23:52
Speaker 2: The question is there has to be technology that's being
23:56
Speaker 2: developed as well to ensure privacy of not only you,
24:00
Speaker 2: but your environment. Right, because when it's your phone, then
24:03
Speaker 2: it's your environment as well.
24:05
Speaker 1: So you brought up privacy in that context, we can
24:08
Speaker 1: knock out private see in the context of the database
24:10
Speaker 1: as well. Here, how could it go wrong? Building a
24:17
Speaker 1: database of thousands of people's voices with tons of data
24:23
Speaker 1: about them sort of answers itself.
24:26
Speaker 2: Yeah, it can go wrong in many ways. And I
24:28
Speaker 2: just came out of like two hours of meetings of this.
24:31
Speaker 2: So add the Bridge to ay I program. We have
24:34
Speaker 2: a huge group of bioethicists and one of our big
24:38
Speaker 2: aim as a group is really to ensure patient privacy
24:41
Speaker 2: and to answer these questions of how do we protect
24:44
Speaker 2: patient privacy in the context of open data. Right, So,
24:47
Speaker 2: you are absolutely right, tons of things can go wrong.
24:50
Speaker 2: People can be potentially reidentified through their voice. So one
24:56
Speaker 2: of our biggest goals this year is determined what part
24:59
Speaker 2: of the voice is identifiable and which part is not okay,
25:03
Speaker 2: And all of this is based on the Hippo law.
25:05
Speaker 2: Hippo law is from the nineteen nineties.
25:08
Speaker 1: Hippola the that governs sharing and security of people's medical information.
25:13
Speaker 2: Correct protected health information PHI, we call that, and that
25:17
Speaker 2: law was made in nineteen nineties, right and back then
25:20
Speaker 2: they listed a list of things of what they called
25:23
Speaker 2: PHI or identifiers that cannot be shared openly and that
25:27
Speaker 2: should stay in the hospital. And voice prints are listed.
25:32
Speaker 2: When you go into what a definition of a voice
25:35
Speaker 2: print is, it's very nebulous. It's you know, we don't know.
25:39
Speaker 2: So because of that nebubularity.
25:42
Speaker 1: If I have that word, if that's nebulosity, I'm fred,
25:45
Speaker 1: I don't know.
25:49
Speaker 2: Because it's so nebulous. A lot of institutions, a lot
25:52
Speaker 2: of hospitals will say, well, you know, voice is not
25:56
Speaker 2: is not an identifier as long as you don't say hi,
26:00
Speaker 2: I'm John Doe and I live at four twenty five,
26:02
Speaker 2: blah blah blah. Other universities will say, no, no, no,
26:06
Speaker 2: voice is always an identifier. You can never really least
26:09
Speaker 2: voice data. So what our group is doing right now
26:13
Speaker 2: is really looking at why the hippo law says this,
26:16
Speaker 2: what are the actual legal implications of sharing voice? And
26:21
Speaker 2: we always grade it in terms of risk, Right, if
26:24
Speaker 2: I talked about all the things that we collect, you
26:27
Speaker 2: can think that the respiratory sounds are probably very safe
26:30
Speaker 2: to share versus a speech sample. As we say free speech,
26:37
Speaker 2: it's probably the most identifying if you have to grade it, right,
26:41
Speaker 2: And we're kind of looking at, well, where is the balance?
26:45
Speaker 2: How much can we release? And also we can transform
26:49
Speaker 2: the data, so for example, we can change the data,
26:52
Speaker 2: the audio data and what we call visual spectrograms.
26:56
Speaker 1: Like a waveform.
26:58
Speaker 2: Yeah, it's a sort of waveform that machine learning can use.
27:02
Speaker 2: We can extract acoustic features, right, like loudness, frequency, stuff like.
27:08
Speaker 1: That, and basically trying to figure out how to make
27:10
Speaker 1: a person be not identifiable based on their voice without
27:16
Speaker 1: messing up the database. Like that's the balance, right, Like
27:19
Speaker 1: if you monkey with their voice too much, then your
27:22
Speaker 1: monkey with the data, the database that we care the
27:24
Speaker 1: most about. Like that seems like a hard trade off.
27:28
Speaker 1: So if we go farther out into the future, you
27:31
Speaker 1: solve all these problems, you build your giant database, the
27:34
Speaker 1: models get really good. All of these things seem like
27:36
Speaker 1: things that may well happen. I'm curious about, you know,
27:43
Speaker 1: AI doing some chunk of what you do now. Right,
27:48
Speaker 1: we see this happening, say in radiology already. AI is
27:51
Speaker 1: clearly very good at doing some of the technical work
27:54
Speaker 1: that radiologists do in diagnosing scans of patients. Right, how
27:58
Speaker 1: do you think about the future of AI, you know,
28:02
Speaker 1: using acoustic biomarkers to make diagnosis in a way that
28:06
Speaker 1: is similar to what you do now as a human being.
28:10
Speaker 2: Yeah, So, I mean, I I don't think I'm going
28:11
Speaker 2: to lose my job yet because I would say that
28:14
Speaker 2: my primary goal as a doctor is not to necessarily
28:19
Speaker 2: do that, right, Like, my primary goal is, yes, to diagnose,
28:22
Speaker 2: but it's to treat patient. So for now, AI is
28:24
Speaker 2: not going to treat the patient. So I think what
28:26
Speaker 2: it's going to do is it's going to support a
28:28
Speaker 2: lot of the workforce. So for example, I'm an academic laryngologist.
28:35
Speaker 2: I'm a super a super specialist.
28:37
Speaker 3: Right.
28:37
Speaker 2: For people to get to me, they often see like
28:39
Speaker 2: five different doctors.
28:41
Speaker 1: So instead of going through for specialists who can't figure
28:43
Speaker 1: it out, you go to your primary care doctor or
28:45
Speaker 1: even you just talk to your phone, and your phone
28:49
Speaker 1: says you better talk to your primary care doctor, and
28:50
Speaker 1: your primary care doctor sends sends the patient directly.
28:54
Speaker 2: To you, correct, correct, right to say like hey. Because again,
28:58
Speaker 2: most of what we do for a very long time
29:00
Speaker 2: will will need a gold standard right diagnosis. So often
29:05
Speaker 2: you know it's it's a it's a biopsy, or it's
29:08
Speaker 2: a it's and imaging.
29:09
Speaker 3: You need a gold standard.
29:11
Speaker 1: The acoustic biomarker is not a clear enough diagnostic technique.
29:16
Speaker 1: You need something more reliable.
29:18
Speaker 2: So I don't think it's going to you know, no
29:20
Speaker 2: doctor will say, oh, well, based on this, this is
29:22
Speaker 2: your diagnostics, start chemotherapy. That's not where we're going. I
29:26
Speaker 2: wouldn't take chemotherapy based on an acoustic biomarker. But it's
29:32
Speaker 2: hopefully going to support a lot of primary care and
29:35
Speaker 2: access to care to get to the right person faster.
29:39
Speaker 1: Great anything else we should talk about.
29:44
Speaker 2: The one thing we didn't talk about, I guess I
29:46
Speaker 2: talk about this all day, so sometimes it's hard to
29:48
Speaker 2: remember what I've said in iowha haven't said. But the
29:52
Speaker 2: implication for probably all this new telehealth you know, online
29:57
Speaker 2: world that we live in, a lot of industries are
30:01
Speaker 2: already integrating tools. So, for example, Canary Speech is a
30:07
Speaker 2: startup that sold a product. I think they're working with
30:09
Speaker 2: teams to capture if there's signs in your voice of depression.
30:16
Speaker 1: Teams meaning Microsoft teams, like Microsoft's version of Zoom.
30:20
Speaker 2: Yeah, yeah, so I think And don't quote me on
30:22
Speaker 2: the particular. Maybe I'm you know, I'm not giving, but
30:24
Speaker 2: but I know there's a few startups that are starting
30:27
Speaker 2: to integrate products in Zoom or in teams to let
30:31
Speaker 2: employers know that, hey, your employee is not doing well
30:34
Speaker 2: based on his voice, for example, Right, and.
30:36
Speaker 1: What is your view of the efficacy of those?
30:41
Speaker 2: So, I mean, I I the easy the quick answer
30:45
Speaker 2: is it probably works partially. Yeah, But the question is
30:50
Speaker 2: not if it works full you're not. The question is
30:53
Speaker 2: does it make a difference? Right, So let's say let's.
30:57
Speaker 1: Do what they say it does. Is a question that
30:59
Speaker 1: matters to me, right, like does it are the claims valid?
31:03
Speaker 1: Seems like a reasonable starting Yeah.
31:05
Speaker 3: I think so.
31:05
Speaker 2: So. I just I just reviewed one an article of
31:08
Speaker 2: one of the startups Fantastic that's looking at like depression,
31:11
Speaker 2: and I mean their numbers look great. I do think
31:14
Speaker 2: that's that the results that a lot of these projects
31:16
Speaker 2: are getting are definitely positive and promising.
31:19
Speaker 1: Absolutely, we'll be back in a minute with the light.
31:28
Speaker 3: M h.
31:37
Speaker 1: Now, as promised, we're back with the lighting around. What
31:41
Speaker 1: was your band called?
31:43
Speaker 2: Ha, My chase stage name was Ella Bence Ella Bence
31:48
Speaker 2: because my last name is Ben Susan so that was
31:50
Speaker 2: too long.
31:51
Speaker 1: And yeah, al became Ella.
31:53
Speaker 3: Yeah, I could say my first name.
31:57
Speaker 1: What did you have a hit song in French? What
32:00
Speaker 1: was it called?
32:04
Speaker 2: I wouldn't call it a hit song. It was called
32:05
Speaker 2: annalis samp means I guess in English, it's like a
32:10
Speaker 2: one way flight.
32:12
Speaker 1: Can you sing a line?
32:13
Speaker 3: No, that's my previous life.
32:17
Speaker 1: Can you just say a line? Yeah?
32:22
Speaker 3: I was Uh, it's in French, though I know it'll
32:25
Speaker 3: sound great. Yeah, And nalisam.
32:33
Speaker 2: Means get me a one way flight for the other
32:36
Speaker 2: side of the world. I hope people are really happy there.
32:39
Speaker 1: Well you're here, now, you're you're you're in Tampa. Now
32:42
Speaker 1: did it work out as hoped?
32:44
Speaker 2: Oh? Yeah, I mean I have I have the best
32:46
Speaker 2: job in the world, you know. I get my my
32:49
Speaker 2: mom raised us me and my brother saying you guys
32:53
Speaker 2: need two jobs, one that make money, makes money and
32:56
Speaker 2: the other one that makes you really happy. And if
33:00
Speaker 2: you manage to have both in one job, then you'll
33:02
Speaker 2: have made it, you know. And I get to be
33:05
Speaker 2: a surgeon and work with voice and voice professional and
33:10
Speaker 2: you know, it's been my passion pretty much all my life.
33:13
Speaker 3: So yeah, do.
33:14
Speaker 1: You work with professional singers as a physician?
33:18
Speaker 2: Absolutely? I mean I love treating my professional singers, so yeah,
33:22
Speaker 2: I love that part.
33:22
Speaker 3: Of my job.
33:23
Speaker 1: Have you treated anybody famous? Yes?
33:26
Speaker 3: But I can't tell.
33:28
Speaker 1: What's Taylor Swift really like?
33:30
Speaker 3: Oh that I don't know. I wish No.
33:32
Speaker 2: She sounds fine, though she probably doesn't need a laryngologist.
33:35
Speaker 1: What's the best cure for a sore throat?
33:39
Speaker 2: Voice rest and advil? It takes the inflammation away.
33:43
Speaker 1: Uh huh? Advil? Just ibuprofen and don't talk and voice rest.
33:48
Speaker 3: Yes?
33:49
Speaker 1: Okay? Are you just always involuntarily diagnosing people based on
33:55
Speaker 1: their voice all the time? Okay?
33:57
Speaker 3: So funny funny fun fact.
33:58
Speaker 2: Two months ago, my girlfriend from residency called me. I
34:03
Speaker 2: hadn't spoken to her in like a year and a
34:05
Speaker 2: half and she called me and she said hi, And
34:07
Speaker 2: I said, you're pregnant?
34:09
Speaker 1: Really?
34:11
Speaker 2: And I could hear it because pregnancy gives you like this,
34:15
Speaker 2: you know, you get stuffy in a certain way in
34:18
Speaker 2: your nose, like we call it rhyanidis of pregnancy.
34:21
Speaker 3: And I knew her voice very well.
34:22
Speaker 2: She was my girlfriend for a long time, you know,
34:24
Speaker 2: we studied together, and she just I knew it, you know.
34:28
Speaker 2: And and I think she says, hey, how are you.
34:30
Speaker 2: I wanted to talk to you, and I'm like, you're pregnant,
34:32
Speaker 2: and She's like, how.
34:33
Speaker 3: Did you know? So?
34:35
Speaker 1: Yes, that's amazing people.
34:37
Speaker 2: I mean I was listening to the political debates, you know,
34:41
Speaker 2: and I'm like, ooh, this guy needs a laryngologist. I
34:43
Speaker 2: could I'm diagnosing people all the time.
34:46
Speaker 1: Well, they should give you a call. Yeah, absolutely, Okay,
34:53
Speaker 1: lovely to talk with you.
34:54
Speaker 2: It was one of the funnest interviews I've done.
35:00
Speaker 1: Yeah, Albin Susan runs the Health Voice Center at the
35:02
Speaker 1: University of South Florida. She's also a principal investigator on
35:06
Speaker 1: the Bridge to AI Voice Project. Today's show was produced
35:11
Speaker 1: by Gabriel Hunter Chang, edited by Lydia Jeane Kott, and
35:14
Speaker 1: engineered by Sarah Bruguer. Just a quick note, We're going
35:17
Speaker 1: to be taking a break for the next couple of weeks,
35:20
Speaker 1: but we will have an episode in our feed next
35:23
Speaker 1: week from our colleagues over at the Happiness Lab that
35:26
Speaker 1: is timed not coincidentally to World Happiness Day, which I'm
35:31
Speaker 1: informed is on March twentieth. I'm Jacob Goldstein and we'll
35:34
Speaker 1: be back soon with more episodes of What's Your Problem