The Role of AI in Autism: Dr. Zachary Warren’s Perspective by Gabriele Barrocas
In the rapidly evolving landscape of autism research and intervention, Dr. Zachary Warren, a clinical psychologist and leading expert at Vanderbilt University, has dedicated nearly two decades to developing systems that promote early detection and support for individuals on the autism spectrum.
In this interview, Dr. Warren shares his insights on the transformative potential of artificial intelligence (AI) in autism diagnosis and intervention, the challenges posed by the heterogeneity of autism presentations, and the promising implications of integrating computational techniques with clinical expertise.
What role do you see AI playing in the future of autism diagnosis and intervention?
Dr. Warren: I’m a clinical psychologist by background in training. I’ve been here at Vanderbilt for almost 18 years and during the majority of that time, I’ve tried to develop better systems of care for early action and response. Our systems have historically been somewhat limited, meaning there aren’t enough professionals and also in the context of early markers. The points that we talk about regarding early detection oftentimes tend to be later than we think the actual neurological differences are unfolding.
There’s been a lot of attention in terms of asking how we can make different systems. How could we design better tools and processes for thinking about early detection and recognition. Many of my colleagues have been investigating whether complex data modeling techniques could be inserted into platforms that are trying to understand how infants understand sensory stimuli to see if we can start to dissect and understand meaningful ways to be supportive. It gets more difficult for those models to specifically identify an individual child and say with certainty that there are differences that may warrant attention or support. It also gets harder the earlier and earlier you go. Even though sometimes you are able to see group level differences, as you spoke to, there is such heterogeneity to presentations which is a challenge.
Historically, older systems have relied on downstream observation of behavioral differences at two or three years of age saying, well, this child may not be utilizing much verbal language at this point, or they’re having difficulties, you know, consistently responding to their name or other prompts. Those observations become clearer in children much later, but there is so much development that’s happening early on that we’d love to think about paradigms that could go beyond those human observations.
It gets tricky with that heterogeneity and because there’s so many different sorts of response trajectories that you may see. You may see early differences that improve dramatically without a ton of support or intervention. You may see early similarity but then later on see some regression or plateau.
I think hopefully AI will be able to help us think about ways in which we can separate off those trajectories and find where we may be able to fold in lots of different sources too. There are ways where such advanced computational techniques may be able to fold in differences in terms of speech sounds, motion, as well as response to certain paradigms or maybe even include biological assays like EEG or others that can be folded together to create risk models that are meaningful.
As it stands now, we rely on folks like me who at 18 or 24 months can say that these behaviors objectively look different and need support. There are ways that such models can definitely help us in terms of an early action paradigm and also ways where AI can help us understand what to do when there are early differences. I think that our field has become so much more in tune with the idea of speaking intentionally to the term difference because sometimes a difference can be powerfully impairing to an individual or family, but sometimes difference is difference and may not convey the same level of needs or support. That ability to meaningfully dissect the heterogeneity that exists is hopefully what AI and advanced math can help us with.
I think it’s really confusing to have this overarching spectrum of conditions which encompasses so many individuals in so many domains and trajectories. That heterogeneity actually can also become increasingly complex over time. One of the benefits of looking early is we think there are certain points early on where meaningful differences might look very similar even if they may unfold in different ways. Later, it becomes so much more challenging to understand biological, social, and behavioral differences.
On top of that, I’ve had the luxury of being able to collaborate with folks in the school of engineering for a couple decades who have been thinking about ways that you could use aI to come up with individual support strategies within computer assisted intervention paradigms.
We have colleagues who’ve designed driving simulators that aren’t just responding to whether you’re crashing into a wall or not, but are actually able to understand where you’re looking, how you’re scanning the environment to offer feedback that might be more optimal and supporting a sort of skill acquisition. I think there’s great hope for what I would call intelligent use of intelligent technology and that’s where AI I think holds promise. I also think it comes through the fusion of clinical expertise with that computational expertise.
Gabi: I think it’s very interesting. As I was listening, I was curious. If you do decide to use AI, wouldn’t you have to have huge data sets of people because of such heterogeneity?
Dr. Warren: Right. So, if you just think about the genetics of autism spectrum conditions, there are about 150 to 200 different known genetic contributions that only explain 10 to 20% of presentations of autism, which means there are likely thousands of different genes in reality. So, to be able to meaningfully utilize AI to dissect those trajectories, you would need some pretty complicated sort of data sets. But that also speaks to the idea of why you might need some of those advanced computational sort of paradigms too or how are we going to sort our way through that to find the signals of meaning. I think many folks have been trying to get to the point of powering those types of collections but when there are so many differences is when we may need to figure out where our team’s findings are split off again.
In terms of speaking to autism as this broad construct that is supposed to meaningfully encompass individuals that have totally different presentations and totally different genetic vulnerabilities, maybe they aren’t the same thing. Maybe it’s a word that we’ve used in an attempt to bring meaning there, but maybe there are ways in which the word oversimplifies the complexity of humans and of the experience of humans that are autistic too.
It’s really complicated. I think there tends to be conversations around lumping and splitting in a lot of different domains in psychology and other areas of behavioral science. I think this idea of collapsing our previous sort of diagnostic criteria from old terms into a modern version was really smart and well-motivated but has had this unintended consequence of really pulling together in a way where we may need to be separating off.
In ways to provide better answers from an early intervention sort of perspective to parents and families from a later perspective to autistic adults, I think there’s a lot of work that we need to do in those areas.
Gabi: I think it’s interesting because some biomarkers during early stages of life could have biological underpinnings or maybe have genetic causes but then during development, you have all of these environmental and social factors that change all of those manifestations.
Dr. Warren: Yeah. Development is this unbelievably complicated and dynamic sort of process. It’s not a singular point in time. It’s unfolding. It’s a process of biology and the environment.
From a research perspective, what insights did the pandemic-era studies provide that could shape future practices in ASD intervention strategies?
Dr. Warren: Most of my career has been focused on trying to build better systems. I do a lot of work around early action and screening support but a lot of that is built on this idea that we’re doing it so we can provide support and help and so we can optimally support developmental trajectories over time. It’s not just detection, but it’s support service too.
Working here at the Children’s Hospital, we are mandated with serving not only kids that are in our backyard here at Vanderbilt, but also kids across the state and in bordering states. There have been huge disparities in terms of care based on geography over time. One of the things that we are doing here at Vanderbilt prior to the pandemic was trying to utilize tele support strategies for identification, tele support strategies for parent coaching and interventionist capacity building and similar strategies for working with schools, teachers, and administrators.
We have been doing that prior to COVID, and there were a lot of folks who thought we were a little bit nuts for doing that. They’re like, well we should all be doing this in person, etc. and then what happened when the world shut down, we saw this great need for tools, processes, and systems for delivering things through Zoom and other similar platforms that we are on right now. We saw this mega-expansion in terms of thinking about new ways of delivering service.
I think if there was a good that came from that, it was really just highlighting and puncturing sort of previous systems and identifying if those systems weren’t working well before. There are folks, however, that are reluctant in trying these new things.
I think it just highlighted that tele capacity and service perspective. I think through COVID, we learned that you could use different tools, different tele support strategies, and think differently about what it means to provide early support and action.
Viewing everybody on the spectrum through one pipeline wasn’t working previously. We now have opportunities to think about how we can design lots of different care strategies and systems for the myriad of trajectories that are part of the spectrum. That’s my perspective on how the COVID shutdowns allowed us to do things differently and also forced folks who had more limited willingness to do it to also do things differently as well.
Gabi: I think it was a blessing in disguise for this field.
Dr. Warren: Yeah. I mean, not during it, but after. I think that sometimes there are a lot of unintended consequences of horrible things. I think there were some positives from that experience but it also highlighted for me how stressful things could be for parents, families, and autistic individuals when they weren’t receiving appropriate support. When we had cases of individuals who had great needs and weren’t getting those needs met, we saw how hard that was. It provided additional motivation and shined a spotlight on how much need and support there still is in the autism community.
Gabi: It’s just interesting because you never know the kind of support is needed until a huge event or obstacle like these forces you to know.
How do you address potential cultural differences in the way families interpret or respond to an ASD diagnosis?
Dr. Warren: I think everybody has a different experience with autism. When I’m providing feedback, I’ve given up trying to adequately capture the thoughts or feelings that go along with that process. An individual’s experiences are rooted in culture and a broader world which can vary so tremendously from individual to individual and from group to group. I think it’s really important to have an awareness that those effects can be profound. And there are. I think we’ve seen, at least in my couple decades of working with autistic individuals and families at Vanderbilt, seen some great shifts in terms of acceptance and embracing difference. However, that’s not necessarily universal within our culture or across other cultures too.
I think there are many cultures where there is still a lot of stigma and stereotypes that are part of the cultural landscape that individuals with autism and their families are often confronted with. It’s really important to be aware of that. It’s important to be aware that many of our tools and systems of care were quite literally built for white boys. There’s been a prevalence difference assumed in autism of almost five to one.
In global contexts, what are some challenges you’ve encountered in applying telemedicine or other technologies to ASD diagnosis in countries with fewer resources?
Dr. Warren: We took an approach in terms of using tele support strategies where it was meant to be ultimately embedded within a system of care. So, it was never really constructed to be this idea of a zoom where a clinician would come and give you information and then just send you on your way. We’re trying to build systems where you can zoom in from an early intervention or preschool setting or medical setting, or doctors office, and we can loop in care professionals that would be providing support over time. Not only would we provide information about what’s going on but we would also be saying here is the specific individual who would be helping out. In those situations where it is harder to intersect with what’s going to happen next, that can be really challenging.
With that being said, there can be power in just having that information and empowering individuals to seek out resources and support where there may not be some. In trying to kind of think about deploying tele strategies to lower resource sort of environments, it’s really been about how do we make that identification process supportive and meaningful, not only in terms of the process itself, but what’s going to happen next.
That’s one of the challenges of this sort of early work is that as you try and build capacity to screen and detect, oftentimes that advances at a speed that’s faster than your ability to build up the ability to help and support and provide intervention. You can’t build over time, but over decades you can. I think it’s one of those things of really trying to design things for that context. We’ve realized that we not only have to be working on detection tools, but also support tools that can be linked hand in hand to what’s going on.
Gabi: Yeah. It’s almost like you have to develop support systems that work for now, but then at the same time, you also have to look into the future because priorities are constantly changing.
Dr. Warren: Yeah. I mean, a lot of the times when we’re working on sort of, you know, we have had paradigms where we’re trying to teach folks who may not necessarily have been as comfortable with early action and detection, like pediatricians or nurse practitioners, or speech and language folks who really historically have wanted to kind of refer to an expert to talk about that. We’ve kind of had to work on building up skills for those individuals because there aren’t just enough tertiary sort of experts and it takes so long to be able to access them.
You have to build to that specific context and those people in that environment. I think one of the lessons learned is just making sure that you’re designing two contexts and that you’re not just designing for, but you’re designing with. If you can actually partner with the true end user shareholders and individuals with autism, you’re going to do a much better job in terms of delivering products or service systems of meaning.
Looking back, how have your views on autism diagnosis and intervention evolved over the years, especially as new research and technologies have emerged, and what do you hope to see?
Dr. Warren: Things have changed a whole lot. The words that we use, the language that we use, our approaches, you know. Tennessee isn’t really number one in a lot of public health outcomes. Certainly, it has historically struggled in terms of thinking about service for autistic individuals and those with developmental disabilities. For a long time, we were fighting to be able to get access to intervention services.
For a long time, we were trying to access intensive behavior analytic services for young children as a first line intervention strategy. It took Tennessee a long time to catch up to that compared to other states. It was hard even then because our science has been really poor in trying to predict individual response. It’s been poor in terms of saying what kids need what and at what time and at what intensity. Do we really need to have kids involved in 20-40 hours of intervention? Or is it actually better to get services that are delivered through the parents? Is it enough? Our science will hopefully be able to start identifying individuals with some of that specificity over time. We’ve done some trials here recently that have been examining what types of interventions should be used. Should it be more of teaching or play-based? What intensity? What style? We’ve also done some more recent studies that have compared traditional practice to thinking about modular interventions that are defined based on building core communication skills. We’ve really seen the pendulum swinging in different directions over time.
I still am in clinics a day awake where I’ll see two to four autistic children coming in and parents wanting to know what they should be doing. I wish we had a science that is able to say this is exactly what you should be doing for this specific child based on this profile and trajectory. What we have right now are core systems.
It’s really tricky right. I wish that we had that because when parents come to me too, they are looking for not only an answer for what’s going on, but honestly most folks are looking for an answer on what this means over time. Folks really want to know where this will go over time. They want to know if it matters what they are doing now in terms of looking into the future. It’s really complicated ethically in terms of how we can do this for 3% of the population in a way that makes sense. Things have definitely changed. Things have shifted. Underneath all of it, however, I think there is still this core theme that there are individuals whose neurodevelopmental underpinnings are related to some differences that can get in their way. We just want to make sure that we are doing everything we can to make sure that those differences don’t get in the way. We want individuals to embrace an autonomous trajectory for themselves. That’s easier said than done of course in terms of what we can apply systems to.
Gabi: Along those lines, do you think we’ll ever get to a point where autism has become almost too heterogeneous where it almost encompasses too many people and becomes too broad in its definition?
Dr. Warren: At a certain point, I wonder about the term versus the benefit of exploding the term over time with this idea of is there a way of doing that to allow the individual to still fully embrace their differences in identity which can be healing for some while getting us away from the idea that we are trying to cure this one thing or that there is a single cause for this one thing.
Whenever I give a lecture or a talk about autism, I will say explicitly that autism is a word. It is a word that we use in an attempt to try to capture meaningful difference over time. One point that we see coming to a head now is individuals who are trying to carve out this idea of a profound autism, but there are others that say how is my autism less profound just because I’m verbal, or I have cognitive skills that aren’t apparent in others. Everybody’s truth can be heard but it may need different words for being able to understand it. It’s also tricky because we rely on core sets of behaviors that we observe early on. However, we see those behaviors go in so many different trajectories over time. Were those behaviors always the same thing? Or were those behaviors different and they just had this period where they looked similar? I am a fan of eventually being able to come up with much better terms. Autistic Disorder. Aspergers Disorder. These were terms that were created by physicians in the 40s. We’ve come a long way since then, but we are still relying on these terms.
Gabi: Yea. I think that’s why I’m so interested in researching autism because there’s never one correct perspective. One perspective can be used to explain one thing, but that same perspective may not be sufficient to use in another case. It’s constantly changing. It’s never static. It’s very cool.
Dr. Warren: Yes! I think it’s very interesting. A couple of decades ago, it was this idea of identity-first over person-first language. Then, the field converted to this idea of person-first language because we wanted to see individuals on the spectrum as humans ultimately. Now, it has flung in the complete opposite direction where some say that their autism is an integral part of their human nature and qualities. It’s interesting to see how these things shift over time.
Hopefully, we’ll do a better job. I think we are doing more to recognize the limits of our historical systems, but I still think that people are scared about pushing too much in this category because they don’t want negative appraisal or backlash. We just need a better way of understanding all of the different individuals that we try to capture with one single word. When you find the right words, they can be so helpful. They can also be very limiting. There is a delicate balance in terms of how to define these things well.
This article was originally published in The Real Spectrum Blog.
About the Author:
Gabriele Barrocas is a neuroscience major at Vanderbilt University with a strong interest in autism research, epilepsy, and inclusive science communication. She is the founder of The Real Spectrum, a blog focused on autism awareness and acceptance, where she collaborates with experts to amplify neurodiverse voices and promote inclusivity. Gabriele also conducts research on SUDEP (Sudden Unexpected Death in Epilepsy) in the Nobis Lab at Vanderbilt University Medical Center, where her work focuses on understanding the neural mechanisms underlying seizure-related mortality in mouse models of Dravet Syndrome.
Leave a Response