Player is loading...
Detection of flares by decrease in physical activity, collected using wearables activity trackers, in rheumatoid arthiritis or axia spondyloarthiritis: an application of Machine learning analyses in r
Hervé Servy
Friday, Feb. 1, 2019 · 2:17 p.m. · 20m 56s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
uh i think is there something can they see also to sponsor travel and live may uh as
00:00:05
i said uh um and founder of the sierra but if it is not the topic today
00:00:09
i'm pleased to be member of that you are just for for big that recommendations uh which is managed by too
00:00:15
much you rest i can load the sec low this a which is my mentor since many years um
00:00:23
just a few disclosure just usually use that or raise yields that i'm
00:00:26
presenting to you today uh i are coming from a study that
00:00:31
has been funded by d. b. m. s. pfizer and arrange have care
00:00:35
which has got to machine learning and artificial intelligence a team
00:00:42
um if you want to know murray are of the free articles uh i
00:00:46
don't speak about the posters the first poster i've been pretty shot easier
00:00:50
to do than seventeen about this first reviewed and that the article i was a free one which shows that if you want to know more
00:00:58
the objective of the study uh that we call act
00:01:02
connect last two s. says is there is
00:01:05
i would say link uh i wish an association between the patient reported flair
00:01:11
and what we hope say ever we've in actually teach rack your
00:01:17
for patient with the air re annex a us or that rate is
00:01:21
the major that was quite simple we take a very simple activity tracker or
00:01:26
a french one for sure you can condone others that i why did
00:01:31
we choose this one is reverse think is because we have got some
00:01:35
air efficient wishes advise that does not need to be taken off
00:01:41
because we we know that we we could have some and issue or an difficulties so we decided to take a
00:01:47
kind of watch which has got one ear battery autonomy so like that we were sure that assumes a
00:01:53
passion has got the what she has no reason to take it off every night h. two are learned to
00:01:59
charge to load the battery so that one that was the first think um it's also every week
00:02:06
we ask for a question to the patients about all was o. d. experience some for a year
00:02:12
so on one side we have got a activity measure of by our spot watch another
00:02:17
side we have got some patient reported would come up on a weekly basis
00:02:23
so i give use our uh exactly what i explain uh the way we asked says the fray
00:02:28
r. y. as one single question has a disease
00:02:31
infrared during the last past week no
00:02:35
yes want to free a shot for a year or more than free day uh and then we
00:02:41
had allies okay that the frequency and it's impossible a link uh with a different time parameter
00:02:49
just to give you an overview of the populations as you can see is quite interesting if you look at the desk and shade
00:02:54
herbicides value you see that these patients are quite low disease activities
00:02:59
are not supposed to have got for her i would say
00:03:02
and a lot of them out and uh be allergic so some quite
00:03:05
well control that patients are zero one hundred um uh seventy
00:03:10
patient has been recluse in paris hospital in various uh our era
00:03:15
run metallurgical um considerations i'm a four five different considerations
00:03:22
the first things we re porch it was the first uh article
00:03:26
is yes it is possible meanings a patient accepts to to
00:03:30
to we are a tractor for at least for enough because it was not so easy because we have got some
00:03:36
uh first ritual from other studio reporting difficulties
00:03:40
a tune bought the patients and if you could do to make them accept is that they are
00:03:45
mortgage or like something that has been given by the doctor so we have more than eighty eight
00:03:51
percent of adherents meetings that the number of day where uh you have got to watch
00:03:57
nobody is the one that's supposed to tech of the watch but we
00:04:00
face also something apps is to watch need should be um
00:04:05
we need to upload the data from the watch so they need to learn h. in applications
00:04:09
on their mobile phone and some forget so we lost some data for this reason so
00:04:16
rafferty twelve percent of the day are missing eyes are because they take of the watch and forget to take it back
00:04:21
and i don't know after i don't know are um an activity where they don't want to have a watch
00:04:26
what ah or maybe because there is a um data on
00:04:31
transmission that you should the first thinks it is
00:04:35
more uh i was a click or a report a is is that's the reason for their indeed when
00:04:41
you have four as the patience on what i don't
00:04:44
seventy patients seventy four patients report at least once
00:04:48
for a up during the last free mouth again this patient reported flare and what is interesting is
00:04:56
we ask these questions did you experience for their uh every weeks
00:05:01
twenty eight percent of the weekly reports abroad the patients work
00:05:06
productively answer about yes i've got afraid remember we
00:05:09
have got some patients which have years of experience of the disease so we should pause in mail
00:05:15
in the body what is afraid so i give you um another a graphic
00:05:21
is to show that what was interesting it was not long for
00:05:24
their does not the long waves that are there are physician now it's
00:05:28
also some very short free of one day two days free days
00:05:32
and the question was okay the patients reported but okay yep but is it real in the receipts with the film
00:05:41
and we look at also the physical activity as we say not so
00:05:45
a very interesting reserved except that they don't not walk enough
00:05:49
we know that uh even considering that this patient was in paris and we
00:05:54
know that the patient we're leaving the big c. t. r. walking more
00:05:57
than the patient leaving in your rounds out so even for this patient should are living in central paris were surprised at the our little
00:06:04
below but the number of step it a wise quite the same between air and s. paper
00:06:10
we try to look at the direct sure victoria was a physical activity me was the one who are working
00:06:16
little uh what don't walk a lot better day was one work
00:06:20
walking i if we finally to link between to be a
00:06:24
lie and we can expect uh between uh and also little
00:06:28
link with your logic okay we don't know why i'm
00:06:34
the first results we have got is we use some classical wow
00:06:38
was same statistical model what are we gonna make said model
00:06:41
and we found a link which are significant uh for
00:06:46
the lounge a flyer uh not the short flare
00:06:50
the only issue is that the hours is a clean cut difference meanings and the difference in term
00:06:54
of numbers that the other day is not so high some meaning yes we have got to sing is
00:07:01
each day can but is it easy to don't you fine we live um walking so honestly
00:07:08
it was not very good result ah even if we conclude
00:07:13
in the paper and which is mentioned below is that
00:07:17
there really is a decreasing physical activity for the patients which
00:07:21
report to have got to flare so we say
00:07:24
for air have got a pragmatic all unpacked meeting
00:07:29
it's not just or a feeling unfair reason impact and also
00:07:34
is there a second thanks a is is that's the link yeah existed
00:07:39
but it's not so easy how to manage much we rapidly oh
00:07:43
discover where not very happy with the result because we walk at
00:07:46
the day level why do we walk at the day level
00:07:50
why do we look at the number of that great big we'll get an idea
00:07:56
friend that patient onto that two hundred patients and knighted eight uh our first idea was
00:08:02
to give the reviewed community i mean you choose a statistician but when he
00:08:06
or um when we arrive with forty median point to manage you say
00:08:11
stop getting pretty is a aggregated data at the day little
00:08:15
because it was not easy to manage an heiress as to what you want so that
00:08:20
was the first limitation on those that probably has the root cause of the ship
00:08:25
but it was before what we did a few months after this study started into doesn't sixteen
00:08:31
the first reason has been printed on seventeen the real interesting result uh i've
00:08:36
been fridges and seventeen inches and eighteen is you use machine learning
00:08:41
meeting we use varies something very different which is i don't want to go to
00:08:46
a a two deep is what we call beige and classifications is something
00:08:50
even after the rain forest i think is that what you have to know as machine learning is not
00:08:56
magic you need to prepare the data and that's what can change again meeting we
00:09:01
prepare the data and what we'll do it also is we normalised to bootstrap
00:09:06
uh each patient's to incest meeting each patient was is go just on their meaning we know
00:09:12
uh so we we allies we get to the machine uh not of
00:09:16
numbers that the difference between uh what is the average tape yeah
00:09:20
dave for these patients so i think that's one of the things that
00:09:24
change a game and also we proceed smashing we've also data
00:09:29
at a a stepper mean you'd no longer agree geisha nerve that that's tape
00:09:33
today so we give the full detail or says the route data
00:09:37
that's one thing which interesting was mentioning is you feed the machine was the maximum
00:09:42
of that we we use the software that is called curiosities coming from a
00:09:48
ranch uh orange for a reason as to why uh i made a lot
00:09:52
of research ten years ago and machine learning and artificial intelligence uh and
00:09:57
taking i don't know why because uh ranges managing some data that i'm i'm on
00:10:02
the network and they want to predict a lot of things about all
00:10:05
to move to the data and things like that so that got a lot of our great and that got us after that is quite interesting
00:10:12
what we did is because we don't have a lot so unimportant quantity of data what we made is we
00:10:18
met some ira rearrangement of the data by on ten uh and set which has a sensitivity and
00:10:25
specificity and also we test what is happening when we look at the menu twenty would look at that
00:10:31
they are when we look at the pier for our when you look at the pay update
00:10:35
just a small things according to the vision and according to my
00:10:39
friend or it was absolutely not necessarily to look at the
00:10:43
data to mini prevent a physician save the number of tape today's
00:10:47
enough the things are not happening at the power that
00:10:52
and i let you discovers a result ah the arts is a machine learning productions
00:10:59
so you see that for example or uh for the patients who say they have got flare on eight hundred
00:11:05
on a cheap uh k.s the nation has been agree
00:11:09
agree with the patients so roughly it's ninety six
00:11:13
sensitivity in pairs and sensitivity and ninety seven space if e.
00:11:18
c. t. so it's quite amazing uh numbers we
00:11:22
discover is that if we feed the mission with their
00:11:25
day it decreased to fifty so there um
00:11:30
i would say that the bracket is around the house which remember two hours
00:11:34
so as soon as you are higher than two hours the machine is load is losing is
00:11:39
magic and features so i'll then we try to intensify that's so that's i give
00:11:44
you the different uh sensitivity s. b. c. v. c. d. also the cabin for the
00:11:48
one where front of kappa and that uh we try to understand the gnashing because
00:11:54
we have got physician in the team and is it okay but ah doesn't mention take these decisions and
00:12:00
that's where we'll uh some of them i would try to re presents okay uh trends of uh
00:12:06
move nancy between the fray and no fair week honestly nothing interesting to
00:12:13
but this and there is one reports that was interesting case
00:12:18
we had the machine to because cats is very good indeed you give
00:12:23
some parameter to the software and combine the parameter like resuming to
00:12:29
and they're events um and you play with a one by another i don't if i
00:12:33
buy a new set of the variable and we ask you have shown me
00:12:39
what tells a parameter me to what are the time of the week is that are some
00:12:45
trigger to take a decisions so more is the blue t. is the more important duties
00:12:52
i want to begin to say you have an id want to uh uh if you have
00:12:56
if you think about your patience what could be user very blue moment on this graph
00:13:04
uh_huh
00:13:06
your vision with it
00:13:11
the moments where you can come up with a disease moment where if you have got a flare
00:13:16
you can change the way you activity outperforming i explain monday morning it was not significant
00:13:23
because as i said we have got some patients who happened to be allergic which are quite quite control so they walk
00:13:29
c. brings their children to school so more the morning the cops are
00:13:33
facing the relies so they do what they have to do
00:13:38
that is you have got a flare saturday afternoon is a moments
00:13:42
where the ways they walk it's the way the move
00:13:46
is changing that the same for ways they remember every to a wednesday afternoon
00:13:52
that also moments that the matching you don't to fight if you have got a
00:13:56
free every switch that will be the moments where you will do less
00:14:00
then another wins they have to know where we have no flare so
00:14:04
it was very interesting that at the end smashing sound what's
00:14:09
oh to say you don't you find the patient perspective on the ways they leave with these erin
00:14:14
de after two years where is a learned to live with the disease and don't stop
00:14:20
walking because i've got a little friars know all to cope with
00:14:25
and that was one of air interesting uh ethics
00:14:30
so our course as a corporation up with this study and
00:14:33
his ran allies demonstrates that imagine learning can be use
00:14:37
on re metallurgy on physical activity data and we found
00:14:42
that patient reported for again inspiration reported flay
00:14:45
are extremely envious wait notices a six percent sensitivity
00:14:49
and soon as seven to ninety seventh
00:14:53
sorry don't match the ninety seven beers sensitivity a strong willing to physical
00:14:58
activity and very some pattern that out a patient's a lot also
00:15:03
on the video levels or could be identified to predict at for a
00:15:07
one thinks about productions the machine has been about to classify
00:15:12
meaning the machine looks all did you move these last week and tell you have you been in
00:15:21
free at all not the machine is to not able to predict that and those are
00:15:27
very interesting question there is a limitation for sure for this study is
00:15:32
when the patient is quite big for a connected device in mentioning that is too small we have come some iron missions
00:15:39
we know that varies a little risk off over featuring but we think we manage it
00:15:44
because we have got a very good guy who know many mad that maybe there
00:15:49
is a population section meanings one we have gotten actually which track uh so and um
00:15:53
also we where we accept to have an activity take care but we wear
00:15:57
owner of a knife on or something does as as that's been very the c. s. p. uh questions and also sure is
00:16:04
a question of what is a fair the it's not for its question for physician uh and that's the question was
00:16:11
what do we do with that which didn't not find um the answer to this question is
00:16:17
what do we know that these patients is probably at the flyer
00:16:22
last week and the week before what should we do
00:16:25
it sounds as esther gee that would be required to to to
00:16:29
be sure is it good to pass to these patients come
00:16:32
back or should we do some i don't know some we should increase the court you could uh either for week
00:16:39
so that's another things to know a patient is suffering from one thinks it's another story to
00:16:44
know what should we do what is good at jerry oh sure things uh um again
00:16:52
i would say is the most important thing is the reserves
00:16:57
sean made to collaborations me in between physicians between
00:17:00
engineer between mathematics and that's one very be take away
00:17:04
from from his experience than other things is
00:17:08
we have to be ready to leave with some models that we are not
00:17:11
able to understand it's impossible to understand that you would you could only
00:17:17
have some tap on sensitivity test but you have to forget about being able to understand
00:17:22
exactly all the machine take the decision because the machine browse probably one thousand parameter
00:17:27
of where when we make a recommendation and uh i don't know it's ten parameters and
00:17:32
very also come question about uh data very last conclusions as as i said
00:17:38
it was very interesting is that the study comes fear that there is an object to the consequence
00:17:43
of patient reported fair and machine learning can be able to to detect this fear and
00:17:51
we have also is interesting is you can see in the different lubrication a is we first
00:17:57
um we bought that there is a link but which is not clinically for is not a big difference
00:18:04
but then when you know like that that uh with different choose you you change again and you begin to to see the things
00:18:10
very differently and that's i think so but again the question which
00:18:15
is to painting is okay what should we do now
00:18:18
with this kind of come and how to use it in a daily life also do we
00:18:24
the product is all to use it and buttons or can you got right or is that another question thank you read engine
00:18:32
thank you very much for your it's wonderful torque um we have time for one quick question for every servers so
00:18:39
someone one question that i didn't include him so
00:18:47
uh_huh
00:18:52
topic of up a religion and that's the cajun you do any work for uh for casting for years and uh
00:18:59
because they might they might be relevant for a mighty fine and nasty
00:19:03
and are taking some decisions issue rovers no i'm getting um
00:19:09
there is a project right nah inference where we try to use uh these principle
00:19:14
or ah m. because what we so is that the machine is headboard chew
00:19:19
feel how to say all the patient is fitting and we have got such
00:19:25
many patients say one to start logic the life is changing rapidly
00:19:31
and we will see the project that will start is the very first week
00:19:36
of the patients after a pure logic initiations measure energy which tracker is
00:19:42
a mock your of the six morph response to be logic
00:19:47
and that's a very promising because like that will be able to predict and so to expose
00:19:53
a to reduce exposition of the thirty patients roughly which the not uh uh projects
00:19:59
an answer to be logic and also a leverage a window of opportunity that is quite
00:20:05
common in this disease where if we know that to be allergic is not producing
00:20:10
good research we could maybe think change more rapidly that them is also
00:20:15
we will be there will be but there is a lot of technical challenge because all
00:20:19
these data at system sorry need a lot of data and the question is
00:20:23
the sample size meeting how many patients i would maybe needs something like four hundred
00:20:27
or five hundred patients to train the machine and that's it's them men challenge
00:20:33
i think you're right thank you so pets time to move on and with the next turk yes
00:20:38
i'm pleased to welcome uh my ugly and from framework germany she's a an expert in a
00:20:44
computer science advisory or to my sister and i will go deeper in beep learning and uh with
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
AI and Radiology
Dr Amine Korchi
Feb. 18, 2019 · 11:05 a.m.
851 views
Brain MRI Analysis and Machine Learning for Diagnosis of Neurodegeneration
Dr. Ahmed Abdulkadir, University of Pennsylvania
Feb. 6, 2020 · 10:04 a.m.
163 views