Dennis Wall: The changing face of autism diagnosis and treatment

Dennis Wall: The changing face of autism diagnosis and treatment


(bright upbeat music) – [Narrator] From the campus
of Stanford University, – [Narrator] People
are worried about data. They’re worried about their
privacy and their security. They should be. We need secure systems. – [Narrator] This is The
Future Of Everything. – [Narrator] But we can’t
have a system that closes that data off. It is too rich of a source of inspiration, innovation and discovery
for new things in medicine. – [Narrator] With your host Russ Altman. – Today on The Future Of
Everything, the future of autism. Now autism is a condition
many of us are familiar with, usually diagnosed in childhood. And it’s characterized by the
an unusual series of symptoms. Difficulties with social communication and interaction with other people, unusual sometimes repetitive or sometimes self-destructive
patterns of behavior and difficulty with language. It also has many other
co-occurring medical and psychiatric and
neurological signs and symptoms. Autism exists on a spectrum which includes very
high functioning people who even have great achievements to folks who are severely
disabled by the condition and require care 24 by seven. Now the causes of autism are complex and include both genetic and
perhaps environmental features. I should say that the
scientific community is clear that autism is not caused by vaccines. This is my public service announcement. And in fact there are
grave societal consequences to not using vaccines, including terrible infections that can cause very bad
morbidity and mortality. This ends the public service announcement. The treatments for autism are difficult and not always successful but it is clear that early detection and early intervention is
best when it is possible. Professor Dennis Wall is a professor of pediatrics,
biomedical data science, psychiatry and behaviors
science at Stanford University and he’s an expert at using data science and biology to understand,
detect and treat autism. Dennis, I was surprised to learn that the diagnosis of autism is a difficult and complex process, and one that is actually
particularly uncomfortable for patients with autism
and their parents. But you’ve done some work
to simplify this process. Can you tell us about that?
– Sure. First of all thank you for having me on the show it’s awesome. But yes I’ll explain some of the situation that’s in place right now. So the current clinical
standard for diagnosis is essentially a one-to-one relationship between clinician and child. It’s done at a clinic with the facility that has all the resources that they think they need
at this point in time to diagnose the child. This can take 10, sometimes
17 hours, this process. It can span two days. And naturally as the
incidence of autism has risen which has been dramatic
over the last ten years, let’s say maybe 600% rise in incidents. That waiting list grows
and grows and grows, creating a bottleneck in the ecosystem that’s really difficult to break through. – I can hardly think of any other disease that takes 17 hours of
evaluation to diagnose. Sorry I’m just distracted. – Its amazing series of assessments include academic psycho metrics,
behavioral rating scales, speech testing and then
autism specific testing. – So they’re working hard
during those 17 hours? – They’re working quite hard and it is quite extensive because a lot of times these kids are clinically challenging especially if in the windows of
developmental trajectories where meeting milestones are
different for whatever reason. – What are the other
things that they are trying to kind of rule out. In other words maybe it’s autism, what are some of the
other things it could be if a kid is having trouble along
the lines that I described. – Global developmental delay which is fairly diffuse diagnosis but nevertheless distinct from autism. – And would have different treatment. – May be and this is another thing that’s a little bit arguable in the field. The delivery of that… As far as I know and I’ve
looked at this quite a bit there’s really not a road map to say if this kind of thing is going on, deliver this kind of behavioral therapy. So they may overlap quite considerably. Other things that are diagnosed within this kind of assessment could be speech delay, just
general language issues, communication issues and so on. – But and not autism.
– And not autism. So distinct from autism. – So I derailed you I’m sorry. So you were saying about the how does this diagnosis work
and where can we improve it. – Naturally this is sort of true in other sectors of Medicine where you have clinical established
procedures and protocols for care and assessment. They’re done in hospitals. But with things like
psychiatric conditions and developmental delays, developmental pediatrics in particular, these things are behaviorally assessed. And the behavioral assessments
can be done in other ways. And so what we’ve proposed is
that we’d be able to move this out of the clinic into a mobile platform, essentially evaluation of videos provided by parents of their children
in their natural environments. This is beneficial for two reasons, one because it’s natural. They’re actually behaving the
way they normally do at home rather than an artificial environment like in a clinical setting. And two because its mobilized. It’s sort of ex-clinical, it sort of goes upstream of
and can be done in advance of any particular visit that they would do at a developmental medicine center. – So if I’m understanding
you you’re gonna take videos and it’s not gonna be 17 hours of video it’s gonna be a shorter amount in which the parents and the
kids are both gonna love this because they don’t have to leave home. But that seems to be too magical. So how much video do you need and how good is it gonna
be compared to 17 hours of highly trained
professionals yada-yada-yada. – So we’ve been walking through
a series of clinical tests to document the efficacy of this, so step back and say in the spirit of HAI actually and– – Human-centered Artificial Intelligence a new initiative at Stanford University. – What we’ve proposed is that we can not only evaluate on video but through machine learning methods. We can determine the likelihood of risk for an autism diagnosis, in a matter of about three minutes. – Great so take us through it. How does it work. – So what we proposed first is that starting at the very beginning, we had access to medical record data that represented the outcome from the administration of the exams that can take 10 to 17 hours for example. And we had a bolus of such data, enabling us to really apply
the lens of machine learning to ask very simple questions like can we reduce the complexity of the behavioral features
needed essentially to arrive at a diagnosis
without any loss of accuracy. – So maybe in those 17
hours there were things that are more valuable and less valuable. – And not surprisingly there are. There’s quite a bit of mutual information among the characteristics
that are measured, there’s quite a bit of
correlation among the features and so therefore machine learning is very simple in this case. Linear classification approaches are straightforwardly showing
us that the feature space can be reduced dramatically. – Which means less
questions have to be asked. – Which means less
questions need to be asked. And we get a model out of it. So we not only get the feature set but we get a model to run the features, so that we can produce a
score that fits a distribution which is unlike the binary classifications provided today by the way. It’s not just a yes/no, it’s
an actual probability score that says you have confidence
in this classification, the classification is
trending towards severe, medium or mild severity autism. So it becomes beneficial
in a number of ways. We went on to propose that
we could see these features in video, naturalized video
in the homes of the child. – So once you know the
smaller number of questions that need to be asked, you can say are there ways to get the answers to these
questions more efficiently than a 17 hour visit
to the Medical Center. – And so we went through
the documentation of that. The third thing that we
proposed after having confirmed that we can in fact see
those features in videos that are one to two
minutes in length even. We went on to say that we
could get those videos assessed by non experts. Individuals who have not
been traditionally trained for the diagnosis of autism
or not developmental– – So these are not highly
paid psychiatrists? – They are not and this
paper we just published in the special topics in PLOS Medicine which came out in November 27th, showing how we can not only
observe the features and videos but get non-experts to
feature-tag these features, so that we can run the models in a matter of about 4 minutes on average. This is The Future Of Everything I’m Russ Altman I’m speaking
with Dr. Dennis Wall about the diagnosis of autism. I’m sure we’re gonna get
into other topics as well. So that sounds pretty cool. So let me just make sure I got this. The three-minute video. Is this just a video of the kid or do you prompt the parent
to like give some challenges of some kind so that you get the kid to do some standard tasks? – We do provide some basic
sort of structure and feature, some suggestions like
child at play at mealtime, with Legos kinds of things. Some basic ideas. But generally speaking we don’t need to have a tremendous amount
of consistent structure as long as the child
is doing natural things in their natural setting. – And then the second thing you said is that you were able to
find relative non-experts who watch these three minute videos and I guess fill out a kind
of form of what they see. What kind of things are they noting down in the observation of the of the child? – Eye contact, response
to name, joint attention, social smiling, things like that. If I smile you smile that sort of thing. – Normally that’s what I do. That great. So does it work? And are all of the medical
centers closing down their diagnostic which I’m gonna guess are also lucrative. Where are we in all of this? – Yes it absolutely works. The sensitivity and specificity
that we were able to show in this paper that I mentioned is 94/78 respectively. So really at a bar that’s
clinically acceptable. Next step is and this is
actually being taken forward by a company that I started called Cognoa is to work through the
path of regulatory approval to get this thing listed
as a medical device for diagnostic purposes and it will be the first of its kind. – Would you actually use it for diagnosis or would you use it to triage the kids who should go get a
more detailed evaluation or do you think that detailed evaluation might be on the way out? – I do think that there are children who are absolutely clinically challenging and so triage is a very good way to go with this kind of procedure because it can say these
is yeses and these is nos and these is the kids in the middle that really are difficult
for whatever reason. They’re too young, they’re
in a bilingual household, they’re delayed for other
reasons we’re not quite sure yet. So we do need to see a specialist to be able to suss out what is going on. But one other just to
comment on this question which is great one. I do think general pediatricians are the first line of defense. And generally speaking, they’re not performing
efficient developmental delay of ethos even though
they’re recommended to do so by the American Pediatrics, roughly 33% are doing it. – And my understanding
is that this is critical because early detection is key. – And so the mechanism like this a vehicle that’s mobilized and powered
by efficient machine learning that can render a very
robust diagnostic outcome can be used by pediatricians to enable them to act more efficiently within their seven-minute well-baby checks or well-child checks, to either refer with confidence
or diagnose on the spot. And that’s very much what I
hope will become the future for this particular system, enabling pediatricians to
take action essentially. – Fantastic, so I know that you also care about the the journey of the child after diagnosis and treatment. And you’ve done some really fun things in particular with the Google glass which we all thought was dead but it turns out it’s not so dead. So in what ways could a pair of glasses possibly be useful for
a child with autism? – It’s another great question and I guess commenting on that, I do think the diagnostic
procedures are really effective and if we’re gonna do a
better job at diagnosis than we’re doing today and
drive down the average age from four and a half to
something less like two, the last thing we wanna do
is create another bottleneck on the other side in the healthcare system and so therapeutic access has to be there. And that’s why we pivoted our attention towards developing mobilized mechanisms because like with diagnosis,
access to therapy is delayed. There’s waiting lists and
those are only getting longer. – Can you paint a picture of the traditional treatment protocols? – The traditional treatment
protocols are one to one again, clinical practitioner to a child, typically with with what’s
called discrete trial training where they’re showing the child’s sort of images of emotions. So facial expressions. Here’s an individual who’s
angry with flash cards on paper and doing it over and over and over again until they master the angry emotion and move to the happy emotion. – So this is addressing the
deficit in social communication? Reading faces that’s one of their– – One of the main deficits with
which children with autism, I think unanimously struggle is with making social overtures, understanding solar overtures,
engaging in social scenarios, making eye contact and understanding what faces are telling them. – This is The Future Of
Everything, I’m Russ Altman, I’m speaking with Denis wall
and he was just telling us about the traditional ways to treat which is basically flashcards. And so we wanted to really just which are affected by the way? – But children can progress
off the autism spectrum if this kind of behavioral
training is done extensively. – So I should say flashcard with respect? – Exactly. This is quite an effective system. However its administration is spotty and access to it is difficult. And so and this is not
scalable generally speaking. And it can be not general, it has this issue of non generalizability. If the angry face is bearded, the child may maladaptively
associate beardedness with angriness. So there’s a lack of
natural elements to it, it can’t capture the variation in the human expression sort of landscape. You just can’t do it on flash shots. – We’re seeing a lot of faces and a lot of things are going on. – Exactly not everyone has
the same mechanism for smiling and so what we propose to do is to use computer vision
artificial intelligence to move what is an effective
system out of the clinic again onto a form factor
that’s augmented reality not virtual or mixed reality but augmented reality Google glass so that we can let children experience their social worlds better in
their natural environments. And so what it does is it provides them with a cue to look to faces and when they find a
face they get a prompt that says good job you found a face. – You are saying that on the glasses, you’re kinda projecting… They’re looking at you or me but they’re also getting
some other information in their visual field from the glass? – In the peripheral visual field which really does become
peripheral and almost momentarily. It’s quite effective. It just sits off like Jiminy
Cricket on your shoulders. – And what do they see? – They see a green box in the prism which is the heads-up display which indicates you found a face. And immediately thereafter
they get an emoji or emoticon that tells them what kind of face it is. So happy, sad, surprised. – You have figured out. So this is where artificial
intelligence technology has been used to pre-process the face and say Russ is looking angry. And then you cue that to the child. – And so that while it
sounds relatively simplistic, it is incredibly effective. So much so in fact after
a randomized control trial which we just completed, children are shown to get
quite quite a bit better on a standard outcome measure. This paper will publish in
JAMA Pediatrics on March 25th, documenting really the first ever that I know of anyways
demonstration of efficacy of a digital therapy for
this particular condition. – So this is a little
bit of a side comment but the economic future of Google glass is my understanding is was very unclear but it just sounds like you just found that a remarkably useful application– – Absolutely. – So are you in stuck
because you need Google glass to exist or are there gonna be other ways to deliver the same
kind of thing throughout – We were very deliberate in
making this system software that’s agnostic to the the
form factor essentially. The augmented reality form factor while we believe we looked at a bunch of different off-the-shelf offerings. There are many by the way, this is becoming a more
ubiquitous technology as we speak, sport goggles
and things like that that snowboarders like to use
to see how fast they’re going when they’re going down the slopes, is an augmented form factor. – So that helps because
maybe you can commandeer one of these? – And so that’s also another
thing that was interesting is we learned right away that the the Google glass
itself wasn’t powered enough to actually run the AI that we built and so we ported to a phone, built an app and became totally independent
of the glass form factor. It’s now a vehicle for message passing. It’s watching the world,
it’s getting images, passing that to the phone, the
phone’s doing the processing and sending back some
feedback in real time. So we know we can port to other augmented reality form factors and it’s exactly what we’re doing now. – So how do we deliver this this sounds fantastic and
it sounds like you even have a clinical trial, this is not pretend, the kids are responding well to it. Can we expect to see
these kinds of treatments out in the world in the
next couple of years? – Yes absolutely. In fact as augmented reality
becomes more ubiquitous which will happen. Apple’s developing things, other companies are developing
things on the market. These will become really commodity items that are in the homes of everyone just like phones in many ways and we can pour it to all of those phones as for as a medical device. – This is The Future Of
Everything I’m Russ Altman. More with Dr. Dennis Wall about
autism diagnosis, treatment and the future next on
SiriusXM insight 121. – [Narrator] What are you gonna do? You’re gonna go to Google
and you’re gonna type in. – [Narrator] The Future Of
Everything with Russ Altman from the campus of Stanford University. – Welcome back to The
Future Of Everything, I’m Russ Altman, I’m
speaking with Dr. Dennis Wall about autism and the future of autism. It came to our attention
that people might be not familiar with Google glass and so let me just say Google
glass is a pair of glasses that you can wear but that can
project things on the glass so that while you’re looking
at the world around you, you’re also seeing at the periphery
computer-generated display and that is the basic
technology that was used in our previous discussion. But Dennis I wanted to ask about where your interest
in autism comes from. – From my family. My wife, sister Becky
is on the spectra sees of severe form of autism, I’ve known her since I was a teenager. She’s largely the inspiration
behind all of this. I’ve been working on
it for a bunch of years from different angles
but motivated by Becky. – So as a family member with
a loved one with autism, can you describe the kinds of challenges and perhaps the rewards of having someone in
the family with autism. – Well I think it teaches all of us an appreciation for differences, which is really important,
important for my children and it’s one of the things that we try to instill in them every day and Becky is a big source of that in that training and teaching. Becky’s severity is high so there are a lot of
comorbid kind of symptoms including self-mutilation and another thing she can be nonverbal, she can be destructive you know, things like that that
create challenges obviously for a number of things. Being able to go places, do things. So has impact of course Abby my wife or our family very very seriously and a lot of positive
and not so positive ways. – Now I don’t wanna
violate the health privacy of your sister-in-law but can I ask how early on was she diagnosed and that some of the
work that you’re doing, could it have changed her trajectory as you think about what you’re doing now and what was available
when she was diagnosed. – It’s a great question. She was diagnosed actually quite young, relative to the current
average age of diagnosis. And she was diagnosed by
professionals in Boston who were really at the time the only ones really well equipped to do it. The best in the field so to speak and so she was lucky in that regard. The severity of her autism
is such that it’s unlikely that she would have been
accessible to certain things like the Google glass. – The Google glass would
be hard just logistically– – Logistically to get it but you know, I say this having you know
she’s now age she’s 36 and it’s sort of too late, I hate to say it, for her now it’s about management. A lot of these other kids that’s what the inspiration is about. The kids like Becky who are four, we can get to those kids and if we give them a pair of glasses and maybe a social engagement game that we’ll talk about perhaps, then these children will learn a lot more about their social world
and engage in that world. And once they do that it’ll
unlock layers of the onion that enable them to
grow more on their own. – Fantastic. So yes I did wanna ask you about a game that you’ve kind of invented. It’s called Guess What. It’s an iPhone game and
I believe it’s related to Ellen DeGeneres’s game which
whose name I can’t remember but tell me the story
of how Ellen DeGeneres kind of leads to a interaction
therapeutic tool for autism. – Well yes absolutely. Her game at least popularized
by Ellen DeGeneres it’s called Heads Up and I love it, we played a lot when we were traveling. – Can you just describe it for us? – It’s a game that runs on phones in this case I think both Android and iOS but you put the phone on your… You choose a deck from a selection, let’s say African safari. And once you hit that it
starts the engagement, you put the phone on your forehead and your job is to act like– – So you can’t see it? – I can’t see it but you can. Your job is to act like a giraffe. And if I guess it– – But the charade is coming from an iPhone and you’re supposed to be guessing what your iPhone is showing? – Exactly and if I get it, if you can convince me correctly and you imitate the giraffe
well enough that I guess it. I tip the phone down
passes to the next prompt and suddenly you’re acting like a lion. – Yes I believe now that you say this. I think I’ve been asked to play this game. – What’s amazing about this is
its social inherently social it doesn’t deviate, it doesn’t distract the person
from making eye contact. In fact you need to make eye contact because you want to create a connection. – You’re killing me here, you are not giving me a good giraffe. – Yes exactly. And so there’s that facial engagement there’s eye contact and there’s imitation. All of these things are
fundamental components of the autism behavioral
training programs. – So Guess What is
different from Heads Up? – It is different from
Heads Up for reason… largely inspired by Heads Up, we didn’t want to use the name
Heads Up for trademark issues but would be happy if Ellen DeGeneres wants to give us a call and
sort of support the system. – You’re gonna hear Ellen. – Nevertheless what we decided
to do is turn the camera on. So we’re not only creating the opportunity for a social engagement and an opportunity to see the
child’s ability to see faces. – Now I gotta ask, is the child the one holding the phone or the one trying to act out the giraffe. – In these scenarios the
parents or care providers are holding the phone and the child is attempting
to act out what’s going on. And so that imitation tells us
not just about their ability to see, to understand the prompt that it’s an angry face for example but that they have some sort of inherent physiological
understanding of that emotion, so much so they convince the
person that they’re angry. And so it gives us a whole layer that’s very different from
what we can get from glass. But the data collection becomes an incredibly powerful
secondary benefit of the system. – Yes I distracted you but you had said that the mom’s phone is not
only showing the word giraffe but it’s taking movies of the child. – And it’s a structured movie. So unlike what we have done
previously with diagnostics, we’re using natural video with relatively low levels of structure. Here we’re getting egocentric
full facial image data of the child attempting
to do something social. And it’s incredibly
enriched for the features that we need for diagnostic purposes but it’s also enriched
for computer vision labels that we can use in the future
to build better models. So it becomes this action
to data feedback loop, a system through which we can deliver care while capturing data on which
we can build better models in the future. – So I’m playing this game, there’s a video of me that video may be useful
for diagnostic purposes, it might be used to monitor the progress of therapy I’m guessing this, as you get better at Guess What, that actually means your autism
symptoms are getting better. This is The Future Of
Everything I’m Russ Altman, I’m speaking with Dennis Wall about some of the amazing
uses of augmented, I don’t know, augmented
reality and technology– – Gamification technology. – Gamification, in order to identify and now treat autism. So the analytic challenge of those movies. You said that in your first
story that you told us, you had trained people
to look at the videos, the natural videos and diagnostic purposes and fill out a checkbox of
eye contact and whatnot. But now I think what you
just said a moment ago is that you’re now thinking
about having AI systems that can automatically analyze the video? – Absolutely yes. And so the reason why I
said I’ll give you one kind of straightforward example, the moment in time where
the gameplay corresponds with a sort of a rough label that you can generate on the image data of the child attempting to act out. The moment of time that
when they correspond is a two-point indication that the labeling of that is accurate. So let’s say it’s a happy
emoji that’s on the screen, child successfully acts happy a rough labeler, a noisy
labeler says probably happy. When those two things correspond, we can automatically deposit that labeled image data into a repository that represents examples
of children emoting happy. And with that we can
build emotion classifiers that are tailored to children because right now the
computer vision libraries that we have available to us are highly enriched for
actors who are adults. And so Google glass our system– – Beautiful people who are trained actors but now we’re gonna have a database of young children acting happy. – And you know with any system
and computer vision and AI, you’re only as good as the
examples you can give it. And a numbers of examples you can give it. This is the promise of deep learning. There is promise, quite a bit of promise. Look at Google inception,
this incredible system. But if you’re not able
to give good examples of the kinds of things you’re focusing on which in this case is emotions
in children for example and social abilities in children, then we’re not gonna get there. So the labeling… Moreover this
human-in-the-loop requirement is a challenge. – It’s very expensive. – It’s so expensive. – You now have a model
for getting this data– – Minimizing the manual labor required to generate a repository upon which we can build
a deep learning model. – So just popping up, do you believe that these technologies will have applications outside of autism? Are there other behavioral psychiatric neurological diagnosis that might be able to benefit
from this same technology and what are they? Paint a little picture
of the future for us. Is that what you’re all about? I think many of the deficits in the development development
of Pediatrics are addressable with this exact same paradigm. This action to data on ubiquitous
devices kind of paradigm. That includes ADHD,
developmental delays of speech of general development, of
things that are associated with reading deficits,
dyslexia in particular is in addressable area where we can not only detect
three systems like this but also intervene. – So this is just the beginning. Thank you for listening to
The Future Of Everything, I’m Russ Altman. If you missed any of this episode, listen anytime on demand
with the SiriusXM app.

Leave a Reply

Your email address will not be published. Required fields are marked *