Player is loading...
Developine fine-grained actigraphy's for rheumathoid arthiritis patients from a single accelerometer using machine learning, sensors
Javier Andreau Perez
Friday, Feb. 1, 2019 · 1:51 p.m. · 26m 14s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
thank you much for the introduction uh first you know the thanks thomas and again for
00:00:04
this wonderful a morning after them we're having with so many interesting topics and
00:00:08
involving yeah into remote way try this which is very very important and a very interesting topic
00:00:15
so today um some the talk i'm i'm gonna draw um i wanted to use today and
00:00:20
it's about a change a little bit the title that you might see in the in your agendas so the say
00:00:26
something like a les sheen the potential of were able sensors
00:00:30
will continue ceramic between using machinery or did learning
00:00:34
um this is something so today um i want to just come based uh missus and i'm
00:00:40
i don't because my my feelings computational intel inside the wanna be like very deep into
00:00:44
but the mathematics and the need to read your computational signs of my algorithms but i want to discuss with you and be able get my
00:00:51
presentation will be let's say a paved the way torso discussion on on
00:00:57
on these methods and application of a in our it so
00:01:02
first let me introduce my institution first 'cause now as i'm the only one i think it only speaker from united
00:01:08
kingdom so you might probably uh you know just to predict why gonna talk about so i'm from essex
00:01:15
so this is a our university which is uh one of the biggest campus in the u.
00:01:19
k. it's kind of a wonderland with a lot of a green areas and like console
00:01:23
so uh many a nobel laureates did i mean they're frown actually came out from now
00:01:29
university in economy exulted you're also in politics and peace and we also we have
00:01:35
i'm a human right champions in india and you know
00:01:39
we also export so muff financial economic screws
00:01:44
and uh i mean maybe you're watching the news lately you probably have seen this gentleman here which is now discussing backseat
00:01:50
in the british parliament which uh is shutting down everybody with that order order so well he's our chancellor
00:01:58
so well uh i hope you are enjoying a little bit the the show of a breaks it
00:02:04
and i mean i think the probably uh without in essex that they would be in union
00:02:08
was quite boring so then that's why reality be creating does that limit the moment so
00:02:15
well so i want to uh start my talk we the these image
00:02:20
have fun internet made for you what's uh made by a patient
00:02:24
that says uh my medication hell's and me to not be convene confined and to my backed
00:02:30
which is quite sad because i mean your life is depending on on appeal or
00:02:34
or a specific medici so we come back to the slide on my conclusion
00:02:39
so let me uh give you a little bit of a site and r. s. u. ms summary
00:02:45
about um what i found in military at you when i was uh doing these ah study
00:02:52
so i found that uh uh and everyone try try this
00:02:56
has a huge impact on patient mobility and believing activities
00:03:01
so assume i see some um what highly cited papers uh the shawl that these uh
00:03:07
massive degrees of powers when patients are doing cause well and the doing
00:03:11
household or or or working activities and um
00:03:16
some activity metrics attested to stand stand to seat they have a high correlation with
00:03:21
some symptoms all for a remote where tried the such as paying an extension
00:03:27
uh then see to staten some specific activities which are quite random
00:03:32
are quite specific they're used to um let's say to study
00:03:36
and the preparation of the season some papers such as in this case was sick to stand
00:03:41
i'm also activities are related to have various a specific symptom of uh
00:03:48
l. r. a. which is the morning stiffness so we got when we have been to use with patients on
00:03:52
many of them that were commenting so this internet have which affects more my labs the morning stiffness
00:03:57
'cause every time i wake up i can do pretty pretty much anything so um
00:04:02
and that's something actually that he eats it connects to your to carry on reading
00:04:06
as well so i mean how how tight you are whole wrestlers you are
00:04:09
and also also the script or so of a jew activity
00:04:13
levels um also fatty was factory reported in studies
00:04:18
and i mean let me draw the attention to these natural paper these
00:04:22
last night's natural paper to mention actually that was quite interesting
00:04:26
then there are for movements and also some on uh they learned
00:04:30
that asian between some lights sunlight and darkness they affect
00:04:34
a lot this into magic experience of the page right so that was a huge make analysis that was probably signature
00:04:41
so right now i gonna come out to uh explain these the story had a doing these research so
00:04:47
uh this was so i worked there was a financially supported data
00:04:52
bases farmer d. s. k. uh in collaboration with apple
00:04:56
so they have a problem so the problem was that they invented a new truck
00:05:01
uh for our patients that uh it helps it helps them so to carry on with
00:05:06
activities so in mind or some of the symptoms that uh are a patient half
00:05:11
and uh unfortunately when they wait when they when they went to the f. d. a. to prove these missing
00:05:17
uh if they came back with them like i mean how uh you can actually claim dungeon draco actually works
00:05:23
when you are working at the sees that you know you even to have moving can have
00:05:28
song impact on your can have some let's say uh that's only influence in how the pace is uh feels
00:05:35
so then uh when you feel like somebody calls in the mornings for no detector from a
00:05:39
comical college in the morning it means that you know they have got a big problem
00:05:43
because they expend a lot of means on uh being able to
00:05:47
set divided to to get the clearance of uh these uh
00:05:51
a a track and whether they couldn't because of these uh quite is that we're getting from
00:05:56
that get so then they decided we need something which we provide us some more information
00:06:02
so more to not all uh how the patient is doing continuously only nobody
00:06:07
got a lot basis and with a very short time speed using
00:06:11
in very small chunks of they don't make it a very small ventilation those qualities
00:06:16
uh so then apple uh the developed uh and not for for that i thought
00:06:23
we each uh eats uh it it may need track what you're doing
00:06:27
but it doesn't work with any sensors so far is just
00:06:30
pretty much for which you know you you you can just a patient that fill out and uh
00:06:37
so um i mean also been like for instance if you
00:06:41
uh connected to song sensor such as the fig beaten
00:06:44
some other smart watch you can get some staying count information which gives you a little bit up
00:06:49
uh that's a a and intensity of the activity of the patient across time
00:06:54
whether they needed something it was much more express so they need an active
00:06:59
an activity that tells them so what's the patient was doing in every single
00:07:04
day doing how much time they were doing that and what we're doing
00:07:08
so then they will they will hoping to get some not take a few which is like this is the signal from an accelerometer
00:07:13
which miscellaneous show can thing of your movements and then they tag these colours for different activities
00:07:18
are happening during the day that you are doing all the patients doing during the day
00:07:22
so there was a kind of channels as you can see it's a very high in l. c. noisy
00:07:26
signal that's zero meter uh it's it's something actually which is yes inertial information it's not like
00:07:32
uh let's yeah yeah or which contains more information uh apology wranglers uh the l.
00:07:37
c. d.s and so on so uh but coming back to this problem
00:07:41
so a demeaning and of being able to uh uh detect these
00:07:46
activities so was part of a nice idea they have
00:07:51
that they want to provide and objective ascends alright so the beginning they wanted to quantify the whole match the something that they can do
00:07:58
that with just with the scene i mean just computing the the amount of uh energy that accent on the console such as
00:08:04
uh if is the is the movement was bigger so how much amount of sedentary movement that's something that could be extracted but
00:08:11
when you go into what's actually so for instance how much the person has been blind date
00:08:15
so how many c. to stand stand to see activities has performed so all of that was a very
00:08:20
a a difficult that but dell uh all the difficulty it enables a very interesting
00:08:26
a new layer of assets and which is the one that it's related to the house so if you're able to know how many
00:08:33
times a person actually has the honesty to stand was stand to see it and you can have those chance of working activity
00:08:39
that person has performed in today and you are able to estimate for instance the workings be the candies
00:08:44
all this right time and even just perform some a a specific c. to stand stand to seat metrics about
00:08:50
how the person how long take to uh perform that's activity during the day and across these on
00:08:55
so these these metric of activities and and joe mobility that relate using that you say that you to uh
00:09:02
some conditions such as morning steve may so when saturn seem to stand time to join steve interface
00:09:08
a number of uh continues working periods to tighten and some pain
00:09:13
uh how many times you wake up actually if you think that you know it's three o'clock
00:09:16
in the morning and you find that at some point the person is working that
00:09:19
means that the person is having a very bad press quality so that's something which is
00:09:23
symptomatic experience for the adaptation and so was the same here for s. quality so
00:09:32
right so then we focus mainly so i'm gonna go into the algorithms that power this part of the objective assassin
00:09:39
right so they would beat basic occupational only to reno and mobility study which uh we were
00:09:46
using this active e. x. e. sensor which was developed by some colleagues in new castle
00:09:51
decide use for that you could buy back it's a huge uh uh data collection system at the moment which is
00:09:57
getting a lot of movie data from many different people around a a u. k. i think was internationally
00:10:04
for the sensory its its position in the uh uh eh l. five position and uh
00:10:10
you know this the the pacing with the these it's of course the
00:10:13
sensory it's waterproof and i mean has all the clinical regulations to
00:10:17
uh be able to use it in patients so eagles hall may just where the
00:10:21
sensor and and and you know we gave them also these uh uh
00:10:25
i'm a i'm a i forms actually these uh up in deference so he's
00:10:29
talking also what he's doing and that i have come to that
00:10:32
because of the censored and when the when we this at the study was around ninety pounds
00:10:37
and uh now just we're working at as well with that because i think the technology it it
00:10:41
was like that two years ago it technologies a boring very quickly at the moment so
00:10:46
uh aren't which is a company that we have a lot of contact you know in in in cooperation with boss has developed a new sensor
00:10:52
which actually into the acceleration alexi x. electoral matters and and has a lot of
00:10:57
good afternoon is uh with batteries you can be taken last for a
00:11:00
lot of time so we just it costs around twelve six friends so this
00:11:05
is fantastic not so we can have a sensor that could be actually
00:11:08
it uses a prescription to patients so i mean we're not that far away on
00:11:12
the technology and the hard work about uh these things that we're talking today
00:11:16
so uh like coming back a little bit on the symptoms so we got a these uh uh
00:11:23
different conditions of the patients that this was my affect uh the mobility of the patients
00:11:30
so uh how they performed activity such as uh some of the ones that were reported
00:11:35
uh that that they had more impact word any he eighteen the stiffness a hands on the
00:11:40
re spain because you need hands to some volume doing different activities and so on
00:11:45
so if we look at the the mobility signals let's say uh we feed them feature
00:11:51
again you know a lot uh these uh noise yes or only the signals and
00:11:54
we get these patterns where we can see that you know the activities they haven't somehow
00:11:58
determine pattern uh uh that you know they have a different pattern you know
00:12:03
of all on the on that surrounded the signal but we can as you can see here we have
00:12:08
uh you know we put actually together to uh let's just additional expansion
00:12:12
how the data from the healthy subjects and how they don't they are patients will look
00:12:17
so then uh as you may see here uh the data from the patients would help the subject was much more moody
00:12:23
news i mean the way that accelerometer captured activities while they
00:12:27
are great because depending on the seat not symptomatic expediency
00:12:30
this data was actually much more to genius so sometimes you know you might you might say that you know
00:12:35
some activities might look like something else likes you know for instance for these patients like looking like
00:12:39
uh is doing such as our right to save the activity just by looking at the at the diesel expectation so um
00:12:47
well at the end we collected fee tonight patients that where you know in their houses and they were
00:12:52
uh collecting data i with the sex around with us attach that file station using the mobile phone
00:12:58
so then now we're gonna go into the problems of uh these
00:13:01
products so uh we have a very huge imbalance data set
00:13:05
because uh i mean of course you know all the time seemed to standing isn't interested in your
00:13:09
your your daily so you're doing standing you're doing working increases your line down because she was
00:13:15
fleetingly some of us that we could see a a and c. just than and and
00:13:19
this is selected is that you just perform in a very small times doing today
00:13:24
so then you have a huge and balance they this is you have a lot of information from so activities and very few involved
00:13:31
this is a huge problem actually in in twenty modular soliciting another challenge to these
00:13:37
another challenge is uh how many times so what what how much time do
00:13:41
you need to be able to discover does activity on the accelerometer data
00:13:46
so for instance if you have your walking we think that um if you're working for say ten seconds this is a proper working activity
00:13:53
if you are standing for let's say twelve like your properly standing right
00:13:57
uh however you know some other than stations so activity such
00:14:00
as c. to stand stand to see that these are things that happen actually fast so then you might just need maybe
00:14:07
three seconds three seconds windows three seconds to five seconds to be able to detect those activities
00:14:13
so then we have also the slightly more like how are we able to
00:14:18
be able to recognise this activity when they have different feelings
00:14:23
right so then just to recap the technical challenge that we had we had only one track
00:14:28
selected no matter when coming to this problem where we only have one fax accelerometer
00:14:32
is because we want to longitudinal study so then we won the battery to last longer very low cost
00:14:39
if you have many different sensors you have data scold you have many different
00:14:43
sensory modalities and manufactured under the guise consuming a lot of buttons
00:14:48
so then then we have the problem that we have to focus on the lowest reality of production which is around three seconds
00:14:54
all all fourteen of activity going like this to do able to discover bolton
00:14:59
so then we have a total of twelve activities and of course as i mentioned before i had very highly unbalance
00:15:06
right now i'm gonna go into the yeah so we have we have you
00:15:10
know some of the speakers today we have mentioned about the leader never
00:15:14
some others as mentions well around them for so just to live with your also discussing
00:15:19
and uh this is a very popular methods that you know are actually is some sample learning methods are
00:15:24
the ones that in complete these inches of cattle they're the ones that top always in the
00:15:29
in the in the different ranking and it learning is also very interesting uh i'm actually learning technique because you
00:15:34
know yellow should learn features so you don't have to structural features of learning very highly discriminative subspace
00:15:41
so problems i would say so uh these machine learning algorithms requires in general terms
00:15:48
very large training set so the more actually classes that you want to predict
00:15:52
the more that that you're gonna need so it's that that's pretty much to the remote
00:15:55
them for you know when we working in a division tell you some much right
00:16:00
so uh they usually do not scale well if you just pluck the data do
00:16:05
noisy data and is more later into uh do the learning algorithm straight away
00:16:10
uh they have a lot of problem with over feeding the something that hasn't mentioned today because you know these models are quite complex
00:16:16
so they try to a model things in a very complex way with very different
00:16:20
uh many different interpretations and and convolutional operations that are very complex mathematical come
00:16:26
so then you know if you have you have very small data it's it's an older feeding probably new model
00:16:31
is expressing is is focusing on your data and your real training data too much and you want your
00:16:36
your model you much in learning to generalised to every single data that you could collect later on
00:16:42
uh plus the problem we have with these issues that we don't balance and and these logical sequence of data
00:16:48
this is something i would come back later by inactivity so it's it doesn't make sense
00:16:53
to predict for instance a line activity after you have the honesty to stand
00:16:57
right so then this is something which are we come back later bessie another specific propping up at developing active
00:17:04
uh but so but i'm going to the learning architecture just briefly uh to get an idea about how do
00:17:10
we uh let's say we we because something some of the part of the uh artificial intelligence to
00:17:16
uh let's say expertise to be able to find these uh architecture that we learn from data
00:17:21
so we got these uh accelerometer training data so we have in
00:17:26
a a mature learning we also you know we have the training they certainly someone
00:17:30
is not familiar with much learning we have the training set and the text
00:17:34
the training set it's the amount of data that you used to teach too much in learning
00:17:39
whatever whatever you want to do so to teach to for instance to be
00:17:42
able to the recognising the specific objectives that s. d. v. d.s
00:17:46
or to discover cancer or to segment uh it's uh uh
00:17:51
yeah i'm videos or or or the initial are excellent c. d.'s
00:17:56
and uh and and then we use the testing the testing is something which you haven't used for training at all
00:18:02
the testing is something that just comes uh and in the jordan yeah we
00:18:06
shouldn't have singles testing because otherwise you know if that's testing data
00:18:09
uh it was including training you really like shaking because you model everybody knows about that's what
00:18:14
you want to be something which hasn't been seen before and that's just testing architecture
00:18:18
so what's right for any of the picture is made of to miss it blocks the first these data making
00:18:23
so we using some noble techniques which is a subspace learning which uh on top of that
00:18:29
that funding for specifies and this one is maybe what we're doing is that we are
00:18:33
uh because in the data is very noisy so we want to uh
00:18:36
develop will want to transform our data into a new subspace
00:18:40
where does this can nation with these recognition will be easier how we can remove all those components that
00:18:45
include noises and just keep the ones that are only related to the t. v. to me
00:18:51
right and then the hierarchical model that souls already shoes although some of the channels that yeah
00:18:58
and uh so how how do we actually sold this issue of over fitting and how are
00:19:04
we able to make predictions that somehow makes sense i mean we want to have activities
00:19:09
uh and uh the uh i mean we had to have production star part of our sector activity
00:19:14
and we have this volume balance set so uh we have this problem of
00:19:18
over fitting and this problem of valuable and said so what we'd eat
00:19:22
is to develop an architecture that it takes a day the concert and the strategy of divide and conquer
00:19:28
so we for years uh we develop some plastic farce of before you have seen that there are many different boxes
00:19:34
which they are specifics uh learning i mean the specific matching learning uh functions so each of these functions
00:19:40
they will take the that would have a subjective discriminating between these
00:19:46
a different uh type of uh activities and then we're going down to into the classification
00:19:52
uh i'm a hierarchy and then we're projecting because every classify or wearing matching
00:19:57
learning uh recognition algorithm will help with the confidence about how much
00:20:01
ga they them actually learning is able to recognise the
00:20:04
activity so then we just propagated a confidence downwards
00:20:09
and and and then so um uh what we do is that we penalise
00:20:14
because something classifications of the upper part of the three they are
00:20:17
uh we have to penalise it hiker because you know we'll getting on the specification which you will
00:20:22
be actually more wrong than if you might miss classify something at the bottom of the team
00:20:28
so with is the lighter concur with just the we sold for the challenge of having
00:20:32
that imbalance data said and then we're we're focusing our much learning seem very simple
00:20:38
or problems so then we don't have this massive classification problem from the beginning so we have smaller chunks of classification
00:20:44
and and we use them to build up these nasty classified that
00:20:48
we that we find these lease which are actually our preached
00:20:52
and then how how do we uh is um model our training somehow
00:20:57
then uh we can take the most out of that training so
00:21:02
we have a we did song a site analysis where we got ten subjects
00:21:07
and and then uh we where in a controlled setting where molly about how these uh uh controlling about how these subjects were
00:21:14
uh performing these activities and we end up having that we have seen that if we use up to the specific models
00:21:20
these will these models where i'm having a a a very bad performance a very uh
00:21:26
hoping actually they hide deviation and accuracy and they've measure they've misery
00:21:29
set that is if you get a deformation informative less problems
00:21:33
so then we try different training set this thing some of them we include this is something we we we include song
00:21:39
have the data some control later so then the learning much in the something called time a a transfer learning
00:21:44
they they are much e. learning were able to learn some concerts for on how activities were performed by the health individuals
00:21:50
and then match it to the uh uh are a patients so then the at the end imagine
00:21:55
learning has a general idea including the healthy and patients how activities weber from with x. m.
00:22:02
and then uh as i said uh oh yeah it's also something which a constant above much
00:22:07
e. learning so we want to have a sequence of activities predicted that makes sense
00:22:11
right the same engine before in that sample something on top of our productions we
00:22:14
are that uh air division intelligent agents which is a decision making i didn't
00:22:19
that we'll we'll make sure that both productions are they make sense so
00:22:24
for instance so after a c. stands what has to come existence
00:22:27
so then we're able to um let's situate our predictions by using an
00:22:32
agent on it division in values on top of our children
00:22:35
so it just direct caps on the machine learning some some other
00:22:39
souls so you may see a that are there that the
00:22:43
other was was well over ninety ninety five percent that's a very good actually the recognition a great for uh activities
00:22:50
and and the and so when they have a having a these maps of activity that uh the
00:22:55
the final wanted so just to point out the this was a paper that we published
00:22:59
uh based on these uh these work uh this but it was something that we
00:23:02
didn't want to to anywhere actually expecting to polish that at the end we
00:23:06
we decided to uh to publish it on is getting a lot of popularity disney has more than one thousand five hundred
00:23:12
uh readings and like this year and is another paper related also to uh
00:23:16
i've been tables which uh you want actually get the and warren
00:23:19
is is pretty much a explaining how to how how the feature of
00:23:23
labels will be so uh just wind up a g. s. k.
00:23:28
uh finish apple uh be very successful so they use these metrics that
00:23:32
uh we passed into them with these activities this general activities for
00:23:36
for uh you know the police also in american culture from at all in the in the national conference
00:23:41
so they end up finding that by using these uh these uh these activities these
00:23:45
markers of activities in their in their in their probably in the proposal
00:23:48
so they there was there was a height difference between uh the the the courts of just
00:23:53
our uh having a posse wanda was that we're using the direct they want to
00:23:59
so take a message recognise activity accurate impress easily you know with the single eczema depending on
00:24:05
and is by its need to do this i need of defining a specific learning architecture math models
00:24:10
and uh as i say actually and uh i think we're rich mentioning
00:24:14
today is that when why would you really needs its quality data
00:24:19
so lots of uh amount of data some huge amount of data sometimes
00:24:23
it couldn't be decided that data is null quality no also is not
00:24:27
good enough so quite is pretty much the essence of it um
00:24:32
the features how does this technology for l. information and and says
00:24:35
management's so not just for a supporting me and uh
00:24:39
um i'm a foreigner companies on being able to uh a test the problems but
00:24:44
also knowing how do we can tell the the patients to get on
00:24:48
with their with the season even having some very integration uh we can detect
00:24:51
when someone uh can develop these uh uh the season later stage
00:24:57
so what we need and we discussed also we did yelled yesterday and
00:25:00
maria so what we need is more uh involvements of clinicians
00:25:04
uh thomas because he has a very nice environment at home with you know uh in computer scientists and i but
00:25:09
i think the many more teachers they need to join also on workers are on site
00:25:14
uh data scientists to be able to incorporate that knowledge that you have as well
00:25:17
and the expedient on the designing process of the mighty learning and yeah
00:25:22
so thank you very much a messy none can dress yes so anywhere
00:25:29
thank you very much for for this very nice to work um record shop of times or i think
00:25:35
we don't have time for question no but uh we interested people can capture per at the
00:25:39
coffee break or we'll move on and um i think your um i'm on up to welcome our
00:25:47
every sorry uh who has spent ten years so we're working for i. b. m. microsoft and
00:25:53
then found don't sign of your son or your which is our d. policy are oh so is an
00:25:59
expert in clinical trials and he's also member of sir you are pass force for the river
00:26:04
or about the use of a big daytime remote protein is talk will be about uh for their
Conference Program
Demystification the digital health world
Thomas Hügle
Feb. 1, 2019 · 9:41 a.m.
6530 views
Digital Needs, fears and possible solutions in rheumatology - From a patients' perspective
Janos G. Kozma
Feb. 1, 2019 · 10:28 a.m.
1296 views
Digital RA - prediction and therapy by artificial intelligence?
Jacob Van Laar
Feb. 1, 2019 · 10:45 a.m.
1405 views
Digital lessons learned from musculoskeletal radiology
Patrick Omoumi
Feb. 1, 2019 · 11:37 a.m.
1896 views
Big data: possibilities and threats for complex disease genetics
Zoltan Kutalik
Feb. 1, 2019 · 12:04 p.m.
1349 views
Developine fine-grained actigraphy's for rheumathoid arthiritis patients from a single accelerometer using machine learning, sensors
Javier Andreau Perez
Feb. 1, 2019 · 1:51 p.m.
1180 views
Deep learning meets rheumatology: new methods to predict disease progression in arthritis
Maria Hügle
Feb. 1, 2019 · 2:38 p.m.
1399 views
Digital detection of drug adverse events in rheumatology - a whole population analysis
Paul Hasler
Feb. 1, 2019 · 2:57 p.m.
1146 views
Effect of patient app use on quality of care in rheumatic disease
Yomei Shaw
Feb. 1, 2019 · 3:16 p.m.
1194 views
Automated image analysis in RA
Richard Ljuhar
Feb. 1, 2019 · 4:04 p.m.
1184 views
ELECTOR - eHealth in rheumatology. Patients with rheumatoid arthritis acquire sustainable skills for home monitoring
Marie Skougaard
Feb. 1, 2019 · 4:27 p.m.
1141 views
Return of experience: what are the digital tools available to train HCPs
Xavier Abadie
Feb. 1, 2019 · 4:50 p.m.
1105 views
Differential diagnosis assessment in ambulatory care with the automated medical history-taking device "Diaana", supported by a pilot randomized study
Adrien Schwitzguebel
Feb. 1, 2019 · 5:31 p.m.
1255 views
Recommended talks
Ada inside
Vincent Zimmer, Ada Health, Berlin
June 15, 2018 · 12:39 p.m.
598 views