Player is loading...
Digital RA - prediction and therapy by artificial intelligence?
Jacob Van Laar
Friday, Feb. 1, 2019 · 10:45 a.m. · 21m 09s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
thank you very much a constant dollars for inviting me to this wonderful meeting
00:00:04
uh you do have the digital los for what's a hot i think
00:00:09
and i think we are privileged here because this will be the first off i think many meetings
00:00:14
in the future on this topic this is really going to be groundbreaking in my a few
00:00:20
um so we've been involved in a year project and um it left
00:00:25
on the how to predicts disease flies in patients with rheumatoid arthritis
00:00:30
using routinely collect it a clinical data and this really
00:00:35
i think brings home the message that was uh just uh brought up the importance of data collection
00:00:42
um should do next
00:00:47
it's not ah sir you okay so uh it's just one disclosure i do
00:00:54
also have some other consultants these activities i didn't uh mentioned then
00:00:59
my main disclosure is that i'm really a clinical roommate always just and even though don't from my say on the
00:01:05
best i'm not an a. i. expert so i work
00:01:08
in a team with a i experts and statisticians
00:01:12
i really do not understand too much of what they do behind the computer
00:01:18
but i do know the general principles and of course i
00:01:21
can hopefully explain the results of our efforts to you
00:01:26
so this is for those of you were not remote all this just to their
00:01:29
rheumatoid arthritis is the most common inflammatory joint
00:01:33
disease it's a relapse ingrid meeting
00:01:36
this is in many patients and it's very difficult to predict less and is really
00:01:42
has a negative impact on treatment because you would ideally like to prevent flats
00:01:47
and tape or your medication your anti rheumatic medication accordingly
00:01:52
it's also a chronic disease so you really need to manage this disease long
00:01:57
term i need to be aware of potential long term side effects
00:02:02
um and then it's important to note that disease activity there's no golden stand that we are using
00:02:07
all sorts of imaging technologies these days to try to come up with the gold standard
00:02:12
but the gold standard is still the clinical assessments by
00:02:15
a doctor or trains notice of joint activity
00:02:20
and the reporting of the patients off disease activities
00:02:22
are very crude old fashions uh tools
00:02:27
that word developed by the way by a paid for real in a
00:02:30
nine making about twenty years ago so for about twenty years
00:02:33
we've been using this old fashioned clinical to and this serves as best better than
00:02:38
all sorts of novel in the imaging technology it's it's worth remembering that
00:02:45
and this is what we think we could uh see in the
00:02:48
future when we have a proper to a computer tool
00:02:52
that can serve as a um clinical decision support system
00:02:57
so patient comes to the clinic disease activity is assessed and then
00:03:02
the data from the clinical disease activity are entered into the a patient record of the patients
00:03:08
and the computer comes back with a kind of a risk score which pretext
00:03:13
the risk of a flare let's say in the coming three months
00:03:16
and then you can maybe treat your patience accordingly and up the
00:03:21
um the medication if the risk of a flat is high
00:03:25
or maybe you can tape or the medication if the risk of larry slow or continue with the same those
00:03:31
and i we think this could really have a a clinical use in our packed this
00:03:37
and then we can give to patients tailored advice i don't think the computer
00:03:41
as was said previously will ever replace a doctor doctors have been around for thousands of
00:03:46
years and human beings need another human being good to communicate that's the basic fact
00:03:53
so our hypothesis was that a flare of disease can
00:03:57
be predicted based on information about disease activity
00:04:00
about patient disease characteristics and treatment already present in
00:04:05
the uh clinical uh electronic patient files
00:04:10
so these are the potential scenarios that can be um the use for this production
00:04:15
model you could start tapering patients on full dose of lotus is activity
00:04:20
or you could stop tapering impatience you all ready tapered and in this case biology goals
00:04:25
are mentioned because these are expensive trucks uh about ten thousand years per year
00:04:31
and so if you can tailor to doze back to it may also the talk to work cost savings
00:04:37
and the last and there is that you could increase the dose and patience you already tape it a biological
00:04:43
but haven't an acceptable risk of flaring and then of course the question is what is
00:04:47
an unacceptable risk of flaring and i think that's very patient and doctor specific
00:04:54
and you can also use this to plan the next visit if the risk of less very low you could say to
00:05:00
the patient will come back in six months or twelve months because you risk of a flare is very low
00:05:05
and again essays uh the patient travel time and also cost for the hospital
00:05:11
so this is what we did we have a a big uh research data
00:05:15
platform in it where where all routine clinical data are automatically entered
00:05:21
and we extract this data from patients with the diagnosis
00:05:24
of rheumatoid arthritis you it started biological treatments
00:05:28
and to became eligible for tapering or stopping biological
00:05:32
if they're disease activity uh met certain criteria
00:05:36
and as an outcome measure we useful as now this is another uh a little uh difficulty because
00:05:42
the uh this definition of flare is very subjective
00:05:45
and only recently has an international task force
00:05:49
come up with some kind of definition for flare but mind you
00:05:52
this is again dependent on a lot of subjective parameters
00:05:58
and this is what we had we uh found seventeen hundred seventy
00:06:01
eight r. i. patience in our uh recess take the platform
00:06:05
the ones with biological or five hundred and eighty eight the ones eligible for tapering were three hundred to fourteen
00:06:12
and the patients with sufficient number of data on these parameters were three hundred
00:06:17
and mind you there were some limitations because a bass this is active the score was not done
00:06:23
it all patients at all this is in fact in only forty percent of the patient visits
00:06:28
and state on smoking where difficult to retrieve and that's
00:06:32
because our electronic the uh a data file
00:06:36
allows doctors to find ten spots to enter the data so this is crazy this is a national
00:06:42
software developer who made i think a a basic error to allow too many options in the database so
00:06:49
it's very difficult to extract data from distant places and we decided to basically not do it
00:06:55
and also data on be a miser body mass index we're
00:06:58
missing also a lot and that also always puzzles me
00:07:02
that doctors do not always measure length or height and weight anymore or blood pressure
00:07:09
i don't know whether this is a unique experience itself but i'm i'm just if i suspect it's a generational thing
00:07:15
so we we decide it's too would do a little cup competition this was
00:07:19
a joint effort with a siemens health in yes from holland and germany
00:07:24
and uh we have a whisk it it it acts parker well sing with a last uh uh by statistician
00:07:31
and the expert in joint modelling and this is what joint modelling is but essentially it uses
00:07:36
data from the past and the present to predict the future so it's not only
00:07:42
you're doing a kind of cross sectional data analysis of the patients it it also
00:07:47
takes all the data from the parts of this patients into the future
00:07:52
and you can do sophisticated joint modelling to predict a disease course in the future
00:07:58
the um the downside of the system is that it's a it's a laborious
00:08:03
and it doesn't allow you to enter too many parameters so there's a maximum
00:08:08
kind of number of parameters that you can enter in the system
00:08:11
on the other hand we had um a awesome yeah from a siemens who is another was good from
00:08:17
um is pa country always a lot of uh which 'cause they're um
00:08:22
who moved to germany and he is an expert in machine learning
00:08:25
and see you use independence as so he didn't independent study using the same database
00:08:31
using the difference technique to come up with different similar kind of uh parameters
00:08:38
and this is uh what the models look like on the right
00:08:41
you see the kind of the the test uh models results
00:08:45
i would show the variation if you select certain uh uh um sub populations
00:08:51
from the whole population so in a sense you're looking at how robust
00:08:55
is my model to predict um the future um you know
00:08:59
if i do a a similar study it another population
00:09:03
and on the left you see a table with all the parameters that were used in as i
00:09:06
mentioned that were more parameters used in machine learning because that essentially allows you to do anything
00:09:13
with any sort of data and on the right you see the parameters that
00:09:17
were selected by parker was by the way experienced rheumatoid arthritis research or
00:09:22
so he knows which parameters are importance where's awesome yeah it's
00:09:26
another remote arthritis specialist to use it as she learned
00:09:29
or data scientists the just and the steps anything in in in the been a looks at the outcome
00:09:37
so this is just to give you a flavour of wood what what could happen in in daily life
00:09:42
um on the uh left access to y. axes use easy disease activity spores
00:09:47
grey means that the patient is using a biological white means that there's no why logical no truck
00:09:53
and the red line and reflects the rest of a flare in the patients uh what
00:09:59
you could see in the middle is that in this patient the biological was tapered
00:10:03
and you can say that those was type it to zero but in fact the machine or the
00:10:09
the model would predict the fly all ready before you decide it's too uh wait and say
00:10:15
so what this means is that uh what this tells you you can essentially
00:10:19
allow the machine to protect the risk of a recent flare and instructs the doctor not to
00:10:25
um you know a wait and see but instead and change the medication and aspirations
00:10:33
now this was like i'm sorry was made by a parker results a user data
00:10:37
scientists are miller's uh numbers and letters so this is too much i know
00:10:43
oh i borrow this five from him but this is the validation co what so we tested our model in
00:10:48
our own coordinate left and we then decided to do
00:10:51
a validation study in independence clinical trial population
00:10:56
so these patients were of very well monitored in contrast to the patient from the messy clinics in our hospitals
00:11:02
and this was a study where patients receive the biological
00:11:06
an idea where continued on the biological according
00:11:10
to usual care or what tapered and and hopefully
00:11:14
the tape it's completely off the medication
00:11:17
and this study showed that you can actually reduce biological in fifty
00:11:21
percent of the patients without major harm to the patients
00:11:26
and if the patient develops a flare you can start a biological again and then disease activity will go down so this
00:11:32
is one of the landmark studies in remote all okay we
00:11:35
showed that it you can in patients with stable disease
00:11:39
tape or biological but it comes at the price and the price is the
00:11:43
risk of flaring and this price needs to be discussed with the patient
00:11:48
uh_huh
00:11:50
and this was the initial results from the the the two models
00:11:54
of the um the the machine learning and a joint modelling
00:11:58
and so on the top u. c. d. um uh the results in the u. m. c. u. data
00:12:03
i'm pretty nice but if you apply to the data to be trance population the validation
00:12:09
pro what you found we found that the joint modelling approach was still good
00:12:13
that's the uh data for machine learning we're not so
00:12:16
good anymore however if we use the dress population
00:12:21
to train also the machine learning program the result of the machine learning became each week equally good
00:12:29
so we concluded that the external validation you'll it unexpected pasta
00:12:34
results for joint modelling when compared to the machine learning
00:12:37
and we think that this is a hypothesis that a joint modelling maybe better
00:12:42
with clean data such as a clinical trial with complete data set
00:12:46
worse machine learning you can handle can handle missing data
00:12:50
but it imp use these data uh at random or some kind of us algorithms
00:12:55
um so it may not be as good um if you use this in um in in this other population
00:13:03
so um this is the kind of the whole results of this uh study
00:13:08
so on the top left sorry on the top uh left you see the results from the clinical trial
00:13:14
and the top uh down you see the results from the uh machine
00:13:17
learning tapering and the joint mumbling tapering and what what this earth
00:13:22
slide shows is that you can indeed use these models to reduce the number of less
00:13:29
and also reduce the number of patients flattering when compared to
00:13:33
complete paper it so just blindly tapering all patients
00:13:38
will increase the risk of last and will also increase the number of
00:13:42
patients flaring and in fact you can use these tools to be
00:13:46
more specifically identify patients who are going to flare or not so it
00:13:52
is in in the sense could be a decision support system
00:13:56
so this could be the future you know all remote all the just these days are female
00:14:00
and so a male patient comes the dock doctor while i get the flair for arthritis if i stop
00:14:06
my medication doc and the room with all this is clips all the buttons is computers as no
00:14:13
so in conclusion i think that uh i hope that this little study
00:14:18
uh well i'm convinced you that um these technologies these techniques both the
00:14:24
standards sophisticated statistical analysis and machine learning
00:14:29
can be used to predict flares at least on the short term into in these patients
00:14:34
and maybe use in clinical support and we're now going to use these models in a
00:14:39
prospective clinical trial where will use the support system to guide clinical decision making
00:14:45
but i would like to and my talk with the most important message which
00:14:50
i think is it is that the strength of a model really depends
00:14:53
on your data if you don't have the data if you don't have your data
00:14:58
you know you can't really use machine learning or whatever technique
00:15:01
or even our to artificial intelligence to predict the future
00:15:04
and this will be i think the most important practical thing uh because
00:15:09
it requires a doctors to be trained and to be disciplined
00:15:13
and and uh this data in a coherent way and this is at least not in my department how doctors operate
00:15:20
and this this is how the nurses operates they use the follow protocols but doctors like
00:15:25
that the the freedom and they think that their intuition is better than the machine
00:15:30
so um we probably need to wait for the next generation of doctors
00:15:34
who is willing to accept this a technique thank you have
00:15:42
yep you said on thank you how much that's was wonderful and look into the future we have time for one question
00:15:53
yeah i do you know so you are very nice oh thank you so you
00:16:00
you you were saying uh you were saying that a good data is needed
00:16:05
so very simple question what he's good data
00:16:11
so okay as we can go who is do we need in and and would
00:16:14
most likely as the same question that's on the talk on the big data
00:16:18
what big that that means how big is big how big is diffusion so
00:16:23
all i know that uh with machine learning and we can test
00:16:26
like twenty different machine learning test and so on and we have a
00:16:29
better uh information letterman to talk with the your sister um
00:16:34
but but at one point we can only feet over learn from small that i said
00:16:39
and that was something we have some disappointment between what the machine will tell us
00:16:44
and the dining data set versus a test or reality that's a big yeah
00:16:48
so what good data means okay so good taken means first of all data
00:16:54
that means no missing data so when i talk about missing data in this group we know that's for
00:17:01
example smoking is a protector of non responsiveness to
00:17:04
biological so or it's more the disease severity
00:17:08
we don't have data on smoking and this is what what we found
00:17:11
out that we are very bad in collecting data on smoking
00:17:15
this is a bit we we're missing a key protector so that's the first thing that we
00:17:20
need the second thing i think is um well regular data collection is important because
00:17:26
both the joint modelling and the machine learning not only use data from the present but also data from the past
00:17:32
so it makes sense to regularly collects data or you could use
00:17:37
maybe wearable seen the futures to to have reliable faithful objective
00:17:42
uh data collection because our data collection on bass for example still subjective based on the uh
00:17:48
up to be a doctor or nurse who collects the data so this is i think
00:17:52
a process that will require a continued interaction between
00:17:57
d. machine learn a soda machine learning uh expert's uh data scientists and doctors and then you
00:18:03
need to come up with the model that works best at least in your clinic
00:18:07
but in a may need to be a tested and retested in each clinic separately i thank you
00:18:15
thank you for uh for this great uh presentation knees in
00:18:18
question and scenes despite the fact that things many years
00:18:23
you are in many are sent to fix association from march or
00:18:27
call set of data to connect i think that that is
00:18:30
still amazing to see that forty percent of the dance and you just a meeting which is supposed to be the
00:18:35
called standard measure of air re how do you think we can change the game i'm i'm not a
00:18:41
doctor how do you think it's time to promote some kind of near schemata reduced that would be
00:18:47
in their due to you to connect this kind of data you think it's
00:18:50
a social or any question meaning does the healthcare agency should be
00:18:55
get these data are things that i would get to be born more important indeed
00:19:00
we see what i think is the solution because i added to this province
00:19:05
since ninety three t. news anyway we see in every country that army seemed
00:19:11
at the physicians side i yeah i think it's both a behavioural problem
00:19:16
a behavioural uh uh behaviour of doctors and nurses um so um you know i've i'm i've had
00:19:23
i've been head of my department for five over five years and it's very difficult to
00:19:28
change the culture in in in in in the doctor population and
00:19:32
even with this data they remain very spectacle you know we're we're not going to work with this kind of
00:19:38
uh buttons you know what did they think you know i'm i'm living in a different uh planted
00:19:43
but this is going to be the near future i'm sure and i hope
00:19:46
that insurers will pick this up because they have a stake in this
00:19:50
it's in the interests if the costs are cuts and also in the interest
00:19:54
of hospital boards you know if the cost of trucks can be cut
00:19:58
so i think we need to kind of bypass the culture of existing medical culture and make sure it's imposed
00:20:05
by insurers or hospital boards like in the not the country where you if you want to prescribe genetic
00:20:11
you need to report a lot of data are not the country yes ah and if
00:20:15
you want to keep the patient under be allergic you need to regularly report
00:20:19
a certain number of data yes but i know this is the case
00:20:22
in building for example they have a national registry brutality registry
00:20:26
and we met always just will not get reimbursed if they don't at data on bass and another
00:20:32
activity scores in the national database it's very annoying for them
00:20:36
because it cost them fifty more minutes if they're
00:20:39
not i. t. specialist it takes them half an hour more right and they're paid by the hour
00:20:44
so it's a it's a clash between different interests different stakeholders
00:20:49
okay thank you yeah i think what a a matter of time we'll have a coffee break and little physical
00:20:55
am ex uh exercises required so it's in the second floor you can take the elevator all want
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
Nutrition, metabolic health, cancer and NCDs
Elio Riboli, Imperial College London, England
Sept. 28, 2017 · 2:27 p.m.
573 views