0:00

This podcast is sponsored by

0:02

Queens University Belfast, one

0:04

of the UK's leading research universities.

0:07

This historic Russell Group University has over 99%

0:09

of research assessed as

0:12

world leading or internationally excellent

0:14

in REF 2021. Their

0:16

researchers are making critical breakthroughs

0:19

in food, ionic liquids, cohesive

0:21

societies, and data-driven healthcare. Visit

0:24

qub.ac.uk and see how they are

0:26

bringing research to reality. This podcast

0:28

is supported by the Icon School

0:31

of Medicine at Mount Sinai, the

0:33

academic arm of the Mount Sinai

0:35

Health System in New York City

0:38

and one of America's leading research medical

0:40

schools. How will advances

0:42

in artificial intelligence transform medical

0:45

research and medical care? And

0:47

what will this mean for patients? To find

0:50

out, we invite you to read a special

0:52

supplement to Science Magazine prepared by

0:54

Icon Mount Sinai in partnership

0:56

with science. Just visit our

0:59

website at www.science.org and search

1:01

for the frontiers of medical research

1:03

dash artificial intelligence, the

1:06

Icon School of Medicine at Mount Sinai.

1:09

We find a way. This is a

1:11

science podcast for January 5th, 2024.

1:13

I'm Sarah Krespe.

1:20

Welcome to a new year at the Science

1:22

Podcast. We start out with a look back

1:24

at 2023 and

1:26

the top online news stories with

1:29

editor David Grimm. There'll be cats,

1:31

but also electric cement and mind

1:33

reading. Next, can a machine tell

1:35

a tuberculosis cough from other kinds

1:38

of coughs? Researcher Manuja

1:40

Sharma joins me to talk about her

1:42

work collecting a cough data set to

1:44

prove this kind of discrimination is possible

1:47

with just a smartphone. Every

1:53

year, our online news team publishes hundreds of stories

1:55

from all sectors of science. We talk about cats,

1:57

but we also talk about the science. hear

2:00

about lots of physical science, science and

2:02

kind of news of the weird on

2:05

occasion. And at the

2:07

end of the year, Dave, our online news editor,

2:09

rounds it all up with some

2:11

staff favorites and fan favorites,

2:13

reader favorites. So Dave, what

2:15

have you brought us today? What could I guess that there's

2:17

a cat story in the mix? Yeah, you

2:19

know, sir, every year I feel like we've got a

2:22

cat or a dog story that may be mandated

2:24

at this point by me. But

2:28

actually, it's usually because, you know,

2:30

a lot of these stories are some of our

2:32

most popular of year and that always ends up

2:34

either being a cat or a dog

2:36

story. Yeah, we have to watch out for the internet.

2:38

Right. Influencing everybody to become

2:41

cat content sites. That's right. Okay, so

2:43

I'm going to go with cats first

2:45

then. Okay. This one is

2:47

about the number of facial expressions that cats

2:49

make. Did you write the story? I

2:51

did not write this story. Actually, it was a remarkable

2:53

year because we had a lot of cat and dog

2:56

stories that I did not write. What's

2:58

cool about this story, which is true, a lot of

3:00

our top 10 list, it not only was a popular

3:02

story, but it's also an exclusive story. So we were

3:04

the ones to break the story and

3:06

the only ones who had the story. And

3:08

so this is a story about facial expressions and

3:10

cats and how many they make and how they

3:13

actually make a remarkable number. Researchers

3:16

have found that cats make

3:18

276 facial expressions. That does seem like a

3:20

lot. I don't know how many people make.

3:23

I don't know how many dogs make. I don't really have

3:25

any point of reference, but that cat makes

3:27

faces at me all the time. Right. Was

3:29

it surprising? I mean, I guess they're social

3:32

with us, but is that enough

3:34

to give them such a diverse repertoire

3:37

of facial expressions? Well, the thing I really

3:39

like about the study is people kind of

3:41

think of cats as sort of these antisocial

3:43

creatures, which is clearly not the case. I

3:45

mean, they live in our homes, they sleep

3:47

in our beds, and also they have to

3:49

often interact with other cats. They don't

3:52

have language like we do. So

3:54

how are they communicating affection, anger,

3:58

other types of communication if they a

4:00

language and it turns out, you know, the things that cat

4:03

lovers notice with their cats, the squinty

4:05

eyes, the whiskers that sometimes go in

4:07

different directions, the ears that go in

4:09

many different directions. The dangerous direction

4:12

back. The dangerous direction, right? And

4:14

all of these can combine in

4:16

various different combinations to produce all

4:19

of these different expressions. What

4:21

does the researchers do to figure out

4:23

this quantity of cat expressions? Well, I

4:25

was kind of jealous. They spent a

4:27

lot of time in a cat cafe

4:29

in Los Angeles recording, video recording all

4:31

of these cats interacting with each other.

4:33

They actually got 194 minutes of cat

4:38

facial expressions. And that's how they

4:40

got this very large number. You know,

4:42

Sarah, you had mentioned earlier, how

4:44

do we know how this compares? And

4:46

we know that chip NZs, obviously very

4:48

closely related to us, produce 357 facial

4:50

expressions. So it's actually not too far

4:53

from chip NZs. But the numbers surprisingly

4:55

haven't been tallied for humans or for

4:57

dogs. So we don't know how that

4:59

compares to some of these other animals.

5:01

We should film some people. That's right.

5:05

Seems like that would be pretty straightforward to do. That's

5:09

our mandatory cat story. Now

5:11

we're going to talk about cement batteries.

5:13

I don't usually think

5:16

of cement as something that can hold

5:18

charge, right? Or can pass charge like

5:20

a big insulator. So

5:23

how can you turn cement into a battery? Cement

5:26

is not a very good conductor.

5:28

But what the study did

5:30

was it started with not what

5:32

can cement do, but imagine if

5:34

it could do these things, right?

5:36

So imagine if the foundation of

5:38

your home, which is probably native

5:41

cement, could store power. The foundation

5:43

could power the home or roads

5:45

made of cement. What if those could

5:48

transfer that power to cars in the

5:50

roadway and therefore be a source of

5:52

power for, let's say, electric vehicles, or

5:55

cement storing the power that's

5:58

generated by green technology? like

6:00

wind and solar. So that's

6:03

the possibilities. That's the possibilities.

6:05

Of the future. So

6:07

what do we do to get there? We start

6:09

with something really cool, which I'd never heard of. It's

6:12

a powdered form of carbon known

6:14

as carbon black. And what's cool

6:16

is we're talking about a very

6:18

futuristic technology that's reliant on a

6:20

technology that goes back a long

6:22

time. Since antiquity, carbon black has

6:24

been used to make black pigment.

6:26

But what you can do, what

6:28

the researchers did, was they found

6:30

when they mixed this with cement,

6:33

particles of carbon black repel

6:35

water. So they come together and

6:37

they actually form these long interconnected

6:39

tendrils within the hardening cement that

6:41

can act like a network of

6:44

wires. Maybe you see where we're

6:46

going here. Now if you've got wires, you've

6:48

got something that can all of a

6:50

sudden conduct electricity. Yeah, so you can

6:53

move and sort charge. And so then

6:55

you can store power into cement. Exactly.

6:58

How big is this? I'm assuming it's

7:00

not the size of my garage

7:02

floor. No, it's not. And there was

7:05

no powering of homes or cars done

7:07

in the study. It was actually just

7:09

powering up a few LED lights. And

7:11

the idea is now that they were able to show that

7:14

it actually works on a small scale, can

7:16

they power it up to a larger scale

7:18

to do some of these Gee-Whiz things we've been

7:20

talking about. Very interesting. All right,

7:22

Dave, we're gonna move on to the

7:25

next story, which is about how mad

7:27

scientists can help the public better understand

7:30

the practice of science rather than

7:32

the results of science. Who

7:34

is your favorite mad scientist? I already have one picked

7:36

out. My favorite mad

7:39

scientist is from my favorite movie, Back to

7:41

the Future. So it's gotta be Doc Brown.

7:43

Doc Brown. Doc Brown. Doc

7:45

Brown, unfortunately not in this story, but who

7:47

is your favorite mad scientist, Sarah? Okay, so

7:49

I actually have a tie between young Frankenstein.

7:52

Okay, that's a good one. And with a

7:54

man with two brains, have you seen that

7:56

movie? Steve Martin. Yes, either way,

7:58

there's a brain transfer. plant involved with

8:00

my mad scientist, my top mad scientist.

8:03

Not to date myself, but I saw

8:05

that movie in years. Nice. I'm

8:07

going to say these brain transplants, they don't

8:09

really seem ethical. They don't. And nor

8:11

does probably some of the stuff that Doc

8:14

Brown was doing from the other mad scientists

8:16

discussing the story, Professor Frank from

8:18

the Simpsons. Oh yeah. Cave

8:21

Johnson from the video game portal.

8:23

So we got mad scientists in

8:26

all types of medium TVs, video

8:28

games, movies, everything. And

8:30

some people would say that gives people a

8:32

bad impression of what science is like. It

8:35

does. And actually the point of this story, yeah,

8:37

or the point of the researchers that

8:40

we will get to eventually, uh, that we interviewed

8:42

for the story is to actually do the opposite

8:44

to show not that mad scientists

8:46

are actually good, but to actually show that

8:48

discussing this stuff can actually be good for

8:50

science. It can help the public understand things

8:53

like ethical oversight and review boards, all

8:55

the things that we have in place

8:58

to ensure that science is

9:00

conducted not only sort of

9:02

feasibly, but also ethically

9:05

humanely. Yeah. So this

9:07

is, this had to do with a panel that was conducted

9:09

at a research conference

9:12

and then they had an audience of 450 participants that

9:14

were sort of asked

9:16

all these questions, you know, if mad scientists

9:18

did X, so developed a freeze ray that

9:21

could potentially maybe save people's lives. You

9:23

knew we might be freezing a child to

9:25

prevent him or her from crossing the street.

9:28

But what if the freeze ray like peeled

9:30

their skin off, you know, like be too

9:32

gruesome, but you know, like, how do you

9:34

sort of, how do you sort of deal

9:36

with ethical implications of testing and developing a

9:38

technology like that? Yeah. So this is one

9:40

of my favorite stories of the year because

9:43

we don't often get to do pop

9:45

culture in our stories. We also often

9:47

don't get to run a still from

9:49

the symptoms in our stories. And this

9:51

one has one of those as well.

9:53

Yes. And I know that's a severe

9:55

hardship for you. So

9:58

I like this story because it makes the IRB. fun.

10:00

Yes, exactly. I think our ethical review

10:02

boards need a shiny new reputation. That's

10:04

right. We pride ourselves on writing these

10:07

stories that we're trying to

10:09

translate a scientific discovery to the general public

10:11

so they understand what does it mean and

10:13

why is it important. And this is sort

10:15

of a part of science we don't often

10:17

talk about the review boards, but we also

10:19

want to give the public a sense for

10:21

how these work. And this is a fun

10:23

way to do it. Definitely. All right.

10:25

Next, we have a story with some sound.

10:28

And the question these researchers, I think we're asking

10:30

was, can you take recordings from the brain and

10:33

figure out what is happening in there? Like

10:36

what people are listening to or seeing?

10:38

And in this case, the

10:40

researchers try to reconstruct a song from

10:43

brain recordings. I'm going to play

10:45

the song, the reconstructed song now.

11:00

Okay, Dave, would you have

11:02

been able to identify what song this is? Probably

11:05

not. And for a couple of reasons. One,

11:07

as you can tell from the recording, or

11:09

maybe not tell from the recording, it's a

11:11

little, it doesn't sound like

11:13

any song you've probably heard.

11:15

But the other reason is that the

11:18

song is actually another brick on the

11:20

wall part one, which is the

11:22

lesser known of the another brick

11:25

on the wall songs from Pink

11:27

Floyd, the more much more famous

11:29

version being part two. So

11:31

listeners out there should feel bad if they

11:34

couldn't get it. But it is still

11:36

very remarkable, because this is basically the

11:39

case of researchers taking

11:41

brain readings from a number of participants

11:43

who had heard a variety

11:45

of songs, but another brick on the wall part one

11:47

being one of them, and training

11:49

an AI to sort of

11:52

be able to see could it piece

11:54

these recordings together? Could it interpret them?

11:56

Could it sort of spit out what

11:59

these participants that actually heard and

12:01

while it's not a perfect rendering

12:04

of the song, it's also pretty

12:06

remarkable. There are some elements there. There

12:08

are some elements there, especially if you

12:10

know the song. Of course we

12:13

can't play the song because copyright, so

12:15

go Google it. Oh no, too

12:17

bad, too bad. But yeah, I

12:19

do hear elements if you listen

12:21

to them one after the other that are

12:23

similar. So how do you

12:25

get recordings like this from somebody's brain? Yeah,

12:27

right. Well, this was not easy. I mean,

12:30

this is from something that was done almost

12:32

a decade ago. There was

12:34

electrodes inserted into the brains of people

12:36

with epilepsy. The purpose of

12:38

the study was to record brain activity during a seizure,

12:41

but there was also part of the study that was,

12:43

you know, they were playing music. And

12:45

so many years later, scientists took these

12:47

readings for this new study to try

12:49

to figure out, can we reconstruct this

12:51

song? But it's not just that. It's

12:54

because that's maybe too fun for a scientific

12:56

study, right? My reading is important, but maybe not

12:58

the goal here. What was really cool about the

13:00

study was not just the reconstruction of the song,

13:02

but that the researchers were actually able to pinpoint

13:05

a new area of the brain that

13:07

seems to be involved in the perception

13:10

of musical rhythm. And the

13:12

other cool thing is the application. So you could say,

13:14

well, why are we even doing this? But you can

13:16

imagine we have obviously a lot

13:18

of people in the world that can't speak as

13:20

results of strokes or injuries or degenerative diseases. And

13:22

so if there was a way to sort of

13:24

be able to read their minds as it were

13:26

and be able to sort of kind of spit

13:29

out what they're trying to do now, obviously we're

13:31

not there with this study, but

13:33

studies like this are the first steps

13:35

in being able to do something like

13:37

that. Okay. All right, Dave. Last

13:40

story we're going to talk about, you have to check out the rest of

13:42

the top set online after this one. And this

13:44

one, I think, I feel must have garnered a

13:47

lot of interest from the headline alone. So I'm

13:49

going to just read. Crocodiles

13:51

are alarmingly attuned to

13:53

the cries of human infants. Key

13:56

word there being alarmingly, right? Exactly.

14:00

All right, so this seems like it might have been

14:02

a difficult study to do. How did

14:04

researchers decide that Crocs had

14:06

a preference for stressed out babies? Well,

14:09

Sarah, I want to let you know

14:11

that no babies were harmed in the

14:13

making of this study. Instead, researchers went

14:15

to a zoo in Morocco that houses

14:18

more than 300 Nile crocodiles. And

14:22

what they did was they played on loudspeakers

14:24

a number of different

14:27

cries, some from human babies,

14:29

some from other primate

14:31

babies. And the

14:33

babies cry sort of differed in what sparked

14:35

the cry. Some of these cries were just

14:38

like, they were kind of upset because they were

14:40

getting a bath, you know, and they didn't want to get

14:42

a bath. But some of them were much more frantic cries,

14:44

you know, they're getting a shot and they don't really know

14:46

what's going on. So there's a lot more panic in

14:48

the cry. And the question was, would

14:50

the Crocs respond, really responded

14:53

all to these cries? And would they

14:55

respond especially well to a cry of

14:58

an infant in distress? The

15:00

big reveal is yeah, they're alarmingly

15:03

attuned to the cries. Alarmingly

15:05

attuned, right. Okay. But

15:07

I was also surprised that people weren't so

15:10

good at this. Like even with babies, not

15:12

like bonobos or whatever, but like chimpanzees. No,

15:14

people weren't good at it with babies, human

15:16

infants. Right. That was the other thing.

15:18

When the researchers played these sounds for people,

15:21

they knew that they were all cries,

15:23

but they couldn't really make that distinction

15:25

between like, is the baby especially kind

15:27

of panicked, you know, worried

15:29

in this particular sound or is

15:31

this kind of just kind of a more normal run

15:33

in the middle cry? Should I play them? For

15:36

me, it's hair raising. I can play them though.

15:38

Yeah. Well, we got, this

15:40

is all about the sound files, right? Well, let's

15:43

do distressed baby. And

15:47

then complaining, but not really upset

15:50

baby. I

15:54

can tell the difference, but I don't know. You

15:56

can tell the difference, right? So you would make

15:59

a great Nile crocodile. Okay,

16:02

so when we say that crocodiles responded

16:04

to these cries, what was the response

16:06

like? What did they do? Well, there

16:08

were speakers again playing these cries and the

16:11

crocs tended to approach them much more likely

16:13

to approach the speakers when they were playing

16:15

these cries, sometimes fight the speakers. Alarmingly,

16:18

Sarah, they would sometimes fight

16:20

the speakers. So this is

16:23

a pretty substantial response, these

16:25

crocs said. So the next

16:27

leap in logic is that they were trying to eat

16:29

the baby? I mean, could they have been trying to

16:31

help the baby? Right. And you

16:33

know, let's give these crocs the benefit of the

16:35

doubt. One of the sources we spoke to in

16:37

the story said, you know, we

16:40

know that they're responding, but we don't really know

16:42

what they're trying to do. Because actually,

16:45

we know that these particular species

16:47

of crocodiles, that they respond to

16:49

the distress calls from their own

16:51

young. And so possibly, this is

16:53

just a sort of a nurturing

16:56

reaction, you know, they're concerned, you know, it

16:58

doesn't quite explain the biting of the

17:00

speakers. Well, hey,

17:02

don't some crocodiles put babies in their

17:04

mouth and carry them around? I don't

17:06

know. But that's possible. Yeah, I guess

17:09

that may be true. So okay, so

17:11

I'm going to go for altruistic. Okay,

17:13

crocodile in this. All right, Dave,

17:15

thanks so much for bringing us some really

17:17

fun stories this year. There are five more

17:19

phase on the site. There's five

17:21

more. So we got, you know, we've got our

17:24

top 10 list. So we've discussed five, there's a

17:26

bunch of other ones that are very popular, very

17:28

loved by either our staff or our readers or

17:30

both. So be sure to go to the site

17:33

and check the rest of those stories out. All

17:35

right, thanks so much, Dave. Thanks, Sarah. Dave Grimm

17:37

is the online news editor for science. Visit

17:40

science.org/podcast for a link to the list. Up

17:43

next, listening for tuberculosis coughs

17:45

with minuetia Sharma. Before

17:54

we get to the next part of the show, I'd like

17:56

you to consider subscribing to news from

17:59

science. Every week, we share

18:01

stories from our news site, News from Science.

18:04

Science journalists and editors kindly come on here

18:06

and tell a story for our ears that

18:09

they've been spending sometimes weeks or even

18:11

months reporting and writing. If

18:13

we were counting, our award-winning journalists publish

18:16

as many as 20 stories a

18:18

week, from tracking policy to

18:20

investigations, international science news,

18:22

and yes, when we find

18:24

new secrets about mummies, we report on that too.

18:27

It's an unbelievably valuable service. If

18:30

you were here with us during early COVID days,

18:33

you must have heard how plugged in

18:35

and devoted our news team truly is.

18:38

Please consider supporting non-profit science

18:40

journalism by becoming a

18:42

subscriber for around 50 cents

18:44

a week. To subscribe,

18:46

go to science.org/news, scroll

18:49

down a little bit, and click subscribe on

18:51

the right side. What can

18:53

you tell from a cough besides, oh, that person is unwell?

19:07

This week in Science Advances, Manuja Sharma and colleagues describe

19:10

a method of screening for

19:13

tuberculosis infection using coughs at a smartphone. Hi,

19:17

Manuja. Welcome to the Science Podcast. Hi, Sarah.

19:19

Nice to meet you. Thanks for

19:22

having me. Thank you. This is super

19:24

interesting. Tuberculosis is a big killer. More than

19:26

10 million people per year die from this disease and

19:28

it's on the rise lately. Manuja, what do we do

19:30

now to screen for active TB? What

19:33

kind of screens are there? The

19:36

current goal standards for TB diagnosis include sputum

19:38

culture or gene experts. Molecular

19:42

tests and they are basically

19:44

done on the sputum that's collected

19:46

from patients after they cough

19:49

and that is sent for either a

19:51

PCR analysis or a treatment. culture

20:00

to see if there is any TB

20:02

bacteria growth in it. These two

20:05

are the most common ones though

20:07

they are not available at every

20:09

level of clinics. Right, so not

20:11

everybody's gonna have the PCR machine.

20:13

Yes, these are not really available

20:15

at the peripheral health centers and

20:17

also it's not very easy to

20:19

collect the sputum especially for kids.

20:21

Oh yeah. So that is another

20:23

thing that adds complexity. So availability

20:25

of tests and also doing those

20:27

some of the tests are not

20:29

that easy. There's a

20:31

need here and your study actually aimed

20:34

to use coughing sounds as a screening

20:36

tool. You know how people

20:38

attempted to use coughing sounds before

20:40

for disease diagnosis? Looking at cough,

20:42

both the cough counts as well as

20:44

the cough features goes way back. Cough

20:46

count is more popular you know when

20:49

you go into the clinic, doctors ask

20:51

you know have you been coughing, how

20:53

often are you coughing? There are studies

20:55

showing that those subjective methods are not

20:57

that accurate. People really don't remember you know

20:59

how often they are coughing and so we need

21:01

counters to help with that. Beyond

21:04

just the frequency of coughing it's

21:06

also the features of the cough

21:08

that's been of interest lately. There

21:10

have been studies done it even

21:12

in 1990s to understand you know

21:14

different kind of pulmonary diseases and

21:17

trying to understand if the cough sounds different

21:19

and what they have seen is

21:21

there is some differences because of

21:24

how the cough is produced and

21:26

essentially because of that difference in

21:28

the source of sound, we get

21:30

different kind of frequency

21:32

mapping of coughs and that's

21:34

been recently studied, COVID-19 made

21:36

it quite popular and so

21:39

there have been a lot of interest in

21:42

the field of tuberculosis also because cough is

21:44

one of the most primary symptoms that's used

21:46

for symptom screening. So if we can add

21:48

more objective analysis in there, can we add

21:51

more analysis to the features and try to

21:53

understand from the sound of the cough if

21:55

it sounds different from any other pulmonary disease

21:57

that would be an easy to use screening.

22:00

tools at peripheral health centers. So what makes

22:02

it such a hard problem to solve,

22:04

you know, using cough sounds as

22:06

data for a screening tool? Well,

22:09

the first is actually establishing the

22:11

fact that there is signature in

22:13

cough. That's not yet established. So

22:15

we are still at that research

22:17

question whether there is signal and

22:19

cough, disease-specific signal and cough. The

22:22

first hurdle is actually getting access

22:24

to the dataset of

22:26

subjects coughing because they have TB

22:28

and subjects naturally coughing because they

22:30

have some other health issues and

22:33

trying to understand if we see

22:35

a difference there. How were you

22:37

able to capture coughs from people

22:39

with tuberculosis? So this dataset was

22:41

collected at Kenya Medical Research Institute

22:43

in Nairobi where patients were coming

22:46

in with TB symptoms and they

22:48

were screened with TB symptoms using

22:50

the standard screening protocol. And

22:52

then they were asked whether they want

22:54

to sit for a two-hour study in

22:57

a room where they'll be just, you

22:59

know, sitting in passively coughing and not

23:01

asked to produce cough voluntary. So all

23:03

these subjects sat in a room and

23:06

their cough was recorded using three

23:08

different devices. And we made sure

23:10

that the room did not have

23:12

a lot of ambient noise and

23:14

also there was less talking in

23:16

the room so that we get

23:18

clean cough. And similarly, we had

23:20

a lot of subjects who did

23:22

not have TB and or who

23:24

were incorrectly screened for TB and

23:27

they were also enrolled in the study

23:29

to get some coughs from non-TB subjects.

23:31

How many coughs did you end up using in

23:34

your model? We have a cough dataset of around

23:36

33,000 coughs from 149 subjects. That's a lot of

23:40

coughs. Yeah, that's a lot of coughs.

23:43

We ended up using for the balanced

23:45

dataset for training and testing. We used

23:47

around 21,000 coughs. And we're

23:50

also releasing that dataset with the state book.

23:52

Right. So I actually asked you for cough

23:55

sounds and we can play them here. We

23:57

have some that are from TB patients. Yes.

24:02

And some that are from non-TB

24:04

patients. And

24:07

we can't tell the difference. Doctors can't

24:09

tell the difference. And actually, for

24:11

your study, you transform those piece

24:14

of audio into basically a visual,

24:16

you know, like a graph. And

24:18

that was important to get a computer

24:21

or machine learning to tell the difference between them.

24:24

Yes. Like you said, anyone with

24:26

normal hearing abilities will not be able to

24:28

differentiate between the two costs. And

24:30

so the idea was to look it into

24:32

a different domain, try to understand and see

24:35

how the frequency and energy is changing over

24:37

time. And can we see

24:39

any features that specific to TB

24:42

or other pulmonary health issue. And

24:44

so we transformed that into scale grams,

24:47

which is essentially a plot of

24:50

how the frequency is changing over

24:52

time. And we took

24:54

images of those plots and send

24:56

it through a ResNet 18 classifier

24:58

and looked at the results of the

25:00

Briney classifier, whether it's classifying is that TB

25:03

or not TB. And how good

25:05

was the classifier at making that distinction? So

25:07

we use cross-validation with our

25:09

dataset. And with that, we

25:12

got an accuracy of around 79%. This

25:16

obviously needs further validation using independent

25:18

test sets. And that's true for

25:21

every machine learning model we need

25:23

for the validation. So

25:25

one thing I noticed is that you used different

25:27

kinds of recorders to capture the costs.

25:30

And maybe I'm interested in this because I

25:32

work in audio. But I thought it was

25:34

interesting that smartphones that you use turned out

25:36

to be the best way to record these

25:39

costs. Is that surprising to you? It

25:41

was a little surprising because we used

25:44

three kinds of microphones. One was a

25:46

microphone similar to what's being used for

25:48

the spot cast recording a high-end microphone.

25:51

Then there was a smartphone microphone. And

25:53

then the third one was a conference

25:55

microphone that we see in conference rooms

25:57

on the table. The worst. the

26:00

worst. My hunch was that the

26:02

worst would be the conference microphone,

26:04

then the smartphone and then the

26:06

best recording but what we saw

26:08

was the smartphone did the best

26:10

and one of the reasons could

26:12

be that when the subjects are

26:14

coughing, the smartphone had automatic gain

26:16

control so it was able to

26:18

adjust the amplitude based on how high

26:20

the subject is coughing which was not

26:23

the case for the higher-end microphone and

26:26

we got a lot of pops out that

26:28

was saturated so essentially clipping so

26:31

that was I think one of the reasons

26:33

for not getting good coughs from there. Oh

26:35

that was some good audio chat. Yeah

26:37

for sure so

26:39

yeah if you have a microphone that doesn't compress the

26:42

sound when it gets too loud you end up with

26:44

yeah it's just a wall of sound that you can't

26:46

see any features in and so here you go the

26:49

smartphone is good for that. So another

26:51

question about your data real quick before we move on

26:53

to how you can turn this into an

26:55

application. This is unsourced information

26:57

but I do feel like people

27:00

from different places sneeze differently, do

27:02

they also cough differently place to place?

27:05

So that could be true so all

27:07

of this data was collected in Kenya

27:09

at a research facility there so it

27:11

was a study just done in that

27:14

demographic and we saw the difference there.

27:16

There could be differences in how people

27:18

are coughing for example here versus there

27:21

and that's just because you know it's

27:23

part of the sound people sound different

27:25

and so the coughs can also look

27:27

different and that's where the machine learning

27:30

comes in you train the model you

27:32

need more data and you also need

27:35

various transformations that you can do to map

27:37

models from one data set to the other

27:39

so yes there could be differences there and

27:41

that's something to work on when you know

27:44

the screening tools come out actually is used

27:46

in the public. Yeah so it'll be interesting

27:48

to find out how much coughs vary place

27:51

to place and how this can be adapted

27:53

to that if they do it

27:55

might be that these features are universal

27:57

regardless of you know your vocalization preference.

28:00

Do you know anything about the relationship

28:03

between what you're seeing in your scaleogram,

28:05

what you're hearing in the cough and what is going

28:07

on with the person? Why does it

28:10

sound different? One of the

28:12

theory is that TB impacts the

28:14

lung tissues and because

28:16

of those impact in those lung tissues,

28:19

the cough sounds differently.

28:22

And what we played, also played

28:24

around was different frequency ranges trying

28:26

to understand which frequency range was

28:29

having the most impact on the

28:31

model. And we saw that

28:34

data in the range of 10 to

28:36

4 kilohertz was impacting the model

28:38

the most. Okay, that's super interesting.

28:40

And so perhaps there's something about

28:42

the way the lung tissue has

28:44

changed that's like affecting that frequency

28:46

of the cough. Could be. Now,

28:49

if this smartphone is good enough

28:51

to record your sound for your

28:53

data set, is it also good

28:55

enough to take sound from people

28:58

and use it to diagnose or

29:00

at least say this person might have

29:02

TB based on their cough? Yes,

29:04

and that's what the next steps for

29:06

the projects would be. And even

29:08

the model that we have built is a lightweight model

29:10

that can go on the smartphone. The

29:13

idea would be to capture the sound of the mic on the

29:15

phone and run it on the phone and

29:17

then see the analysis in the clinic. So

29:19

there is a challenge that comes in with

29:21

a screening tool based on packs of coughs.

29:23

It is when the patient comes in, we'll

29:26

have to wait for the patient to cough

29:28

and then do the analysis. There have been

29:30

studies using forced coughs that you the

29:32

patient comes into the clinic and you ask them to

29:35

cough and then you take those coughs

29:37

and analyze that. But what we

29:39

saw in a study that the passive

29:42

and natural coughs differ a

29:44

lot and model trained

29:46

on passive cough is not translatable

29:48

to voluntary cough. Yeah.

29:51

So there are some challenges there,

29:53

but we think generally patients having

29:55

TB do cough a lot and

29:57

so maybe if not in the

29:59

clinic. they could be given that app on

30:01

the phone if they have a phone or they

30:03

could be given a phone to take home for

30:06

a day and it can record night coughs and

30:08

then that can be used later on to

30:10

analyze and screen for TB. Okay.

30:13

Are there other diseases, other human body

30:15

sounds that we could listen for,

30:18

analyze and help with disease diagnosis?

30:20

There has been studies done in

30:22

general on overall voice analysis because

30:24

you know how people sound worse

30:27

when they are down with so.

30:29

So we do think different sounds

30:31

can be important, not just cough,

30:33

just in general sound analysis of

30:36

how the person is talking or

30:38

even like the wheezing sound that

30:40

you know at times is produced.

30:42

So all of that can also

30:45

help in analysis. But for this purpose of

30:47

the study, we only looked at cough and

30:49

one of the reason was cough is

30:51

used in symptoms screening for tuberculosis and

30:54

we wanted to see if there's any

30:56

way to make that symptoms screening more

30:58

robust. So if

31:00

people in the clinic or you know at home

31:02

have this app and they come in

31:04

and they say, oh my app says I have TB,

31:07

what does the clinician do differently? Do

31:09

they still you know do those next

31:11

tests but this kind of helps

31:14

rule in or rule out TB quicker

31:16

and easier? So it's essentially

31:18

a screening tool to help facilitate

31:20

the presence screening tool which is

31:23

just you know asking patients whether

31:25

they have fever, cough or night

31:27

sweats. So just adding

31:29

on to that with objective analysis

31:31

of cough and not just patient

31:34

level, it can also be used

31:36

to understand hotspots of TB transmission

31:38

because it is one of the

31:40

most infectious diseases causing death right

31:42

after COVID-19. So it's

31:44

very critical to understand where the spread

31:46

of TB is. So if a

31:48

lot of people in the community have the app and

31:51

we get the data, we are getting

31:53

a lot of TB sound like coughs

31:55

in this area. So that could help

31:57

bringing in interventions to prevent that transfer.

32:00

transmission. So we think it

32:02

can be used at two levels, individual

32:04

and community levels. That's wonderful. What's

32:06

next for this research? What else do

32:09

you want to learn about coughs or

32:11

how to do screening with them? So

32:13

the first step is validating this result.

32:15

We have done a cross-validation study and

32:17

we want to evaluate our models on

32:19

an independent test set. And

32:22

with that, we still have

32:24

ongoing cough recordings happening in

32:26

the area. So we want

32:28

to switch now to

32:30

smartphone and do all the cough collection

32:32

on that and try to see if

32:34

we can use the model

32:37

to do some real-time prediction and how

32:39

that behaves. Yeah, that's great. Thank

32:42

you so much, Manuja. Thank you, Sarah, for

32:44

having me here. Manuja Sharma was

32:46

a PhD student in the Department

32:48

of Electrical and Computer Engineering at

32:50

the University of Washington when she

32:53

worked on this paper. You can find a

32:55

link to the Science Advances paper at

32:57

science.org/podcast. And

33:00

that concludes this edition of the Science

33:02

Podcast. If you have any comments or

33:04

suggestions, write to us

33:07

at sciencepodcast at aaaf.org.

33:09

To find us on podcasting apps,

33:11

search for Science Magazine, or

33:14

you can listen to the show on

33:16

our website, science.org/podcast. The

33:18

show was edited by me, Sarah Crespi,

33:20

and Teva McLean with production help from

33:22

Megan Tuck at Pataji. Jeffrey

33:25

Cook composed the music on

33:27

behalf of Science and its publisher, AAAS.

33:29

Thanks for joining us.