Artificial intelligence (AI) and machine learning (ML) are beginning to make headway into medical diagnosis, and its prospects are particularly evident among those complex health conditions that have no obvious single cause and for which conclusive medical tests remain elusive.
A condition that showcases the potential usefulness of AI and ML in diagnosis, as well as their versatility in data analytics, is autism spectrum disorder (ASD). This is an area of great need, according to Dennis Wall, Ph.D., professor of pediatrics and biomedical data science at Stanford University in California (Figure 1). “ASD diagnostics is crying out for a solution that’s faster, more equitable, and more quantitative. This is where biomedical data science can really help,” he said.
Figure 1. Dennis Wall, Ph.D., professor of pediatrics and biomedical data science at Stanford University in California, and founder/chief officer of science and innovation at Cognoa Inc. of Palo Alto, California, which is developing and marketing Canvas Dx, an AI- and ML-driven system to assist with autism diagnostics. (Photo courtesy of Cognoa.)
Biomedical data is already showing its utility. Researchers are now using AI and ML capabilities to sift through numerous information sources ranging from magnetic resonance imaging (MRI) scans to the patient’s motor, communication, and social skills, and developing quantitative measures doctors can use to help diagnose autism in young children. In some cases, those measures have the potential to identify where a child lies on the autism spectrum, which could then assist with treatment personalization, and to differentiate between autism and similar-manifesting conditions such as attention deficit hyperactivity disorder (ADHD), anxiety disorders, and dyspraxia or developmental coordination disorder (DCD).
Wall’s focus is on Canvas Dx, an AI- and ML-driven system that uses behavioral data as quantitative measures of ASD. It is the first and, so far, the only device authorized by the U.S. Food and Drug Administration as a diagnostic aid for ASD [1]. He pursued that work by founding and now serving as chief officer of science and innovation at Cognoa Inc., of Palo Alto, California, which has developed the technology and is now marketing Canvas Dx.
Other research groups are using AI/ML algorithms to analyze different types of data as possible diagnostic tools. That includes a team at the University of Louisville (UofL) in Kentucky, which is investigating the data contained within structural and functional MRI scans as predictors of ASD. Another at the University of Southern California (USC) is developing a coloring book activity that gathers data to identify motor-skill-related symptoms of autism.
All of these projects are designed to address issues with today’s status quo approach to ASD diagnosis. One of those issues is that autism specialists are far outnumbered by the children who need assessment, which can lead to a delay in a child’s diagnosis and therefore treatment, Wall said. “By the time a child is diagnosed with autism and getting services, they’re often already past the window of brain development where those services can be most impactful.” The data bear that out. At present in the United States, the average age of diagnosis is 4–6 years [2], while the optimal age to initiate treatment is 2–3 years [3].
Another issue is that current diagnoses rely on subjective assessments of behavioral observations, which may also delay a correct diagnosis and treatment, Wall said. In addition, ASD diagnoses have cultural and gender biases, he noted. In the United States, that translates to a disparate number of Hispanic and black American children, and females, who are missed [4], [5].
Through data analytics and quantitative assessments, researchers hope to speed diagnoses for all children. “With earlier diagnosis, you can start interventions when the brain is still developing, and possibly take a child who would otherwise live as a completely dependent person who cannot live alone, and help that child become an independent person,” remarked Ayman El-Baz, Ph.D., who co-invented UofL’s autism-diagnosis technology, and is a professor in the UofL Department of Bioengineering. “Parents are under stress for a long time waiting to have their children diagnosed, so if we can detect ASD early and advise the family about what they should do, it will be a big help.”
Coloring between the lines
The breadth of ASD symptoms vary in both their appearance and their severity from patient to patient. This presents a challenge to a clinician trying to assess a child, but it opens the door to numerous data avenues for AI and ML algorithms to search for patterns indicative of autism.
One such avenue is motor sensory variation, or differences in motor skills as a response to sensory information. This drew the attention of Lisa Aziz-Zadeh, Ph.D., professor in the USC Brain and Creativity Institute and Mrs. T. H. Chan Division of Occupational Science and Occupational Therapy (Figure 2). She was intrigued by the work of researchers at the University of Strathclyde in Glasgow, Scotland, who had developed smart-tablet gameplay technology to distinguish typically developing children from those with autism [6]. Aziz-Zadeh set out to see if her research group could expand that approach to differentiate children with autism from those with dyspraxia, a condition that has movement and coordination symptoms like those seen in autism.
Figure 2. Canvas Dx uses data from several sources, including parents. Using an app, they complete a questionnaire about their child’s behavior and send in two 90-second videos of the child interacting with others, communicating, and playing. (Photo courtesy of Cognoa.)
To do it, Aziz-Zadeh’s group conducted a study using the Scotland group’s game and AI/ML system. They collected touchscreen kinematic data while each of 54 children played the coloring game for 5 minutes. Data included finger pressure, smoothness of finger movements, changes in movement velocity, and “hundreds of other parameters,” she said. The 54 children ranged in age from 8- to 17-years old. Of the 54, 20 were developing typically, 18 were diagnosed with autism, and 16 were diagnosed with dyspraxia. Results of the study showed the AI/ML could correctly identify typical versus ASD with 76% accuracy, typical versus dyspraxia with 78% accuracy, and ASD versus dyspraxia with 71% accuracy [7].
Additionally, Aziz-Zadeh reported good results for identifying females with autism. Because females in general tend to be more social than males, autism symptoms can be less obvious in females than males. This can cause screeners to overlook those social symptoms and cause underdiagnosis among females, she explained. “What we found in this study is that motor differences may be better at picking up autism in females.”
The USC researchers now hope to test the coloring book activity with children as young as three years old, and to extend the game to incorporate social aspects and possibly data from parent observations to see whether they can improve accuracy even further, she said.
“The idea is to come up with the least amount of tests necessary to get the best screening tool for autism and dyspraxia—and to differentiate them—and perhaps to do the same with ADHD and anxiety versus ASD and dyspraxia,” she said.
Thinking forward, she envisions pediatricians’ offices adding the coloring game to the tablets parents commonly use to fill out forms, so children can play the game for a few minutes in the waiting room. “If the game gives a red flag for possible autism, the pediatrician could then follow up and send the child for additional diagnosis or tests, as needed. It would be ideal if we could get the technology to that point.”
Putting numbers on behaviors
Wall’s interest in ASD originally centered on finding genetic signatures for autism, but that hinged on a more exacting clinical diagnosis for autism than was currently available. “The status quo for diagnosis has an outcome that is only 70% reliable, so 70% of the time doctors will agree, and the remainder they don’t agree,” he said. He redirected his group toward creating a digital/quantitative phenotype that could be used to identify not only ASD, but where a child falls on the spectrum.
“That led to some of the original thinking and framework for starting Cognoa in 2013,” he said. “We ultimately figured out how we could look at behavioral measures in a 90-second video of a child in their natural environment, and through the lens of AI, produce an objective score that would be repeatable, reliable, and consistent, and that would perform equitably across demographics.”
The full system, now called Canvas Dx, takes into account data from several sources: A child’s health care provider completes and submits responses to a questionnaire describing the child’s developmental behaviors; parents submit responses to a separate questionnaire about their child’s behavior, along with two 90-second videos of the child interacting with others, communicating, and playing; and trained reviewers analyze the parent-provided videos and upload their findings of specific behavioral features and the points in the video where they occur (Figure 3).
Figure 3. Lisa Aziz-Zadeh, Ph.D., professor in the USC Brain and Creativity Institute and Mrs. T. H. Chan Division of Occupational Science and Occupational Therapy. (Photo courtesy of Sarah Gimbel.)
For the latter, reviewers look for a wide range of features, such as various kinds of eye contact or social smiles, or particular hand mannerisms or head movements. “We are using all of that to create a training library of features,” Wall said. “So if you think of the videos as 90 seconds in length and with 30 frames per second, the image frames where our reviewers are identifying specific features go into a library that—once it is big enough—will enable the development of a neural network model that will be proficient for identifying signatures of autism.”
The neural network model is still in development, but even without it, Canvas Dx and its AI/ML algorithms are working exceptionally well at identifying ASD, Wall said. In the fall of 2023, Cognoa presented results of a 124-patient study showing Canvas Dx correctly determined a negative result (no autism) with an accuracy of 95.2%; and a positive result with an accuracy of 94.4% [8]. “Those were really good numbers, and they’re even better now,” Wall said. “We’ve done an additional real-world analysis on 255 patients—not published yet, but it’s in the works—and the metrics are even better: in the high 90s for both positive and negative predictive values.”
Moving forward, Cognoa is working closely with the FDA to update Canvas Dx with new metrics and ideally with new indications, Wall said. “It is a tight partnership, and the FDA works faster now than they ever have before, especially with AI and ML in medicine.” In particular, he noted, the FDA’s predetermined change control plan (PCCP) now allows applicants to request permission to adapt that AI/ML technology as they gain more real-world data [9].
“For others in the field who are thinking of software as a medical device, this PCCP is an exciting opportunity,” he remarked. It is also one that fits well into Cognoa’s plans. “In the future, we do anticipate making authorized changes to Canvas Dx to make it better,” he said. “And thankfully, because this is a positive-feedback loop through which we can gather information as we go, it behooves us to check for cultural adaptability and demographic adaptability, so it can work well for everyone.”
Bringing autism into view
At UofL, the data taking center stage on the autism front comes from MRI scans. El-Baz got involved in this area in 2005 during a meeting with UofL psychiatry professor Manuel Casanova, M.D., who lamented the subjective nature of autism diagnosis and suggested they work together to find an objective alternative.
Casanova (currently the SmartState Endowed Chair in Translational Childhood Neurotherapeutics for the University of South Carolina School of Medicine and the Greenville Health System) had helped found a brain tissue bank that compared typically developing vs. autistic patients. “Some of his microscopic images showed differences in the vertical arrangement of neurons, which were a little bit narrower in autistic subjects compared to control subjects, and we also saw a problem with connectivity,” El-Baz recalled. “That started our work using structural MRI images to develop a noninvasive and objective AI method that could diagnose autism in young children.”
The project got an added boost in 2012 when the National Institutes of Health launched its Autism Brain Imaging Data Exchange (ABIDE), which began amassing structural MRI, functional MRI, diffusion tensor imaging (information about neuronal connections in the brain), and genomic data from individuals aged 5–64. In the ensuing years, El-Baz and his group have collaborated with Gregory Barnes, M.D., Ph.D. of the UofL School of Medicine, to design an AI/ML system that uses information in that diverse database to find patterns suggesting autism diagnoses (Figure 4). “Our AI-based model achieved an average accuracy of 96%,” El-Baz said, noting that it also was able to identify the brain regions associated with ASD, and use them for a quick and accurate prediction not only of the occurrence of ASD, but of its severity [10], [11]. “So, first we could tell autism versus not autism, then we could identify abnormal neuroreceptors associated with autism, and then we could generate an autism assessment report, such as the commonly used Autism Diagnostic Observation Schedule (ADOS) or Social Responsiveness Scale (SRS),” he summarized.
Figure 4. University of Louisville bioengineer Ayman El-Baz, Ph.D. (left), and his research group worked with Gregory Barnes, M.D., Ph.D. (right), of the UofL School of Medicine, to develop AI/ML technology that uses structural MRI, functional MRI, diffusion tensor imaging, and genomic data to predict autism in children as young as two years old. (Photo courtesy of Mohamed Salah.)
El-Baz’s group now plans to broaden its system’s capabilities by tapping into MRI and developmental-assessment data on younger children, as gathered by the Infant Brain Imaging Study (IBIS) Network. “IBIS data is collected from 6-, 12- and 24-month-old month children who didn’t receive an autism diagnosis until they had reached three or four years old, and were from families that have an autism history,” he described. His group has already conducted a study of 50 IBIS subjects, and found the system was able to make an autism diagnosis with an accuracy of more than 90%. A large and inclusive dataset is critical to the tool’s across-the-board accuracy, and he anticipates conducting more encompassing studies as the dataset grows. “We need good data from different regions of the world and different lifestyles, so we can get a diagnostic tool that will work in any place and on anyone,” he said.
Eventually, El-Baz hopes to determine the genes or gene sequences behind the abnormal neuroreceptors. When added to ADOS and SRS reports, he said, such quantitative information will make autism diagnosis even more robust.
The UofL has licensed some of its AI/ML autism-predictive software to Autism Diagnostic Technology, Inc. of Bedminster, New Jersey, and El-Baz is looking forward to getting the technology into clinic where it will help children and their families. “I was speaking at a local school recently, and parents of autistic children told me how long they had to wait to have their children diagnosed, and how very stressful that is,” he said. “It really touched my heart to know we were doing something that will help so many people.”
References
- U.S. Food and Drug Administration. FDA Authorizes Marketing of Diagnostic Aid for Autism Spectrum Disorder. [Online]. Available: https://www.fda.gov/news-events/press-announcements/fda-authorizes-marketing-diagnostic-aid-autism-spectrum-disorder
- U.S. National Institute of Child Health and Human Development. Early Intervention for Autism. Accessed: May 28, 2024. [Online]. Available: https://www.nichd.nih.gov/health/topics/autism/conditioninfo/treatments/early-intervention
- M. van ’t Hof et al., “Age at autism spectrum disorder diagnosis: A systematic review and meta-analysis from 2012 to 2019,” Autism, vol. 25, no. 4, pp. 862–873, May 2021.
- B. S. Aylward, D. E. Gal-Szabo, and S. Taraman, “Racial, ethnic, and sociodemographic disparities in diagnosis of children with autism spectrum disorder,” J. Develop. Behav. Pediatrics, vol. 42, no. 8, pp. 682–689, Oct. 2021.
- R. Brickhill et al., “Autism, thy name is man: Exploring implicit and explicit gender bias in autism perceptions,” PLoS One, vol. 18, no. 8, Aug. 2023, Art. no. e0284013.
- A. Anzulewicz, K. Sobota, and J. T. Delafield-Butt, “Toward the autism motor signature: Gesture patterns during smart tablet gameplay identify children with autism,” Sci. Rep., vol. 6, no. 1, p. 31107, Aug. 2016.
- C. Butera et al., “Motor signature differences between autism spectrum disorder and developmental coordination disorder, and their neural mechanisms,” J. Autism Develop. Disorders, Dec. 7, 2023, doi: 10.1007/s10803-023-06171-8.
- S. Taraman et al., “6.27 exploring the real-world performance of an artificial intelligence-based diagnostic device for ASD: An aggregate analysis of early canvas dx prescription and output data,” J. Amer. Acad. Child Adolescent Psychiatry, vol. 62, no. 10, p. 294, Oct. 2023. Accessed: Jun. 10, 2024. [Online]. Available: https://www.jaacap.org/article/S0890-8567(23)01918-4/fulltext
- U.S. Food and Drug Administration. (Oct. 24, 2023). Predetermined Change Control Plans for Machine Learning-enabled Medical Devices: Guiding Principles. Accessed: Jun. 10, 2024. [Online]. Available: https://www.fda.gov/medical-devices/software-medical-device-samd/predetermined-change-control-plans-machine-learning-enabled-medical-devices-guiding-principles
- M. T. Ali et al., “A personalized classification of behavioral severity of autism spectrum disorder using a comprehensive machine learning framework,” Sci. Rep., vol. 13, no. 1, p. 17048, Oct. 2023, doi: 10.1038/s41598-023-43478-z.
- E. Helmy et al., “Role of artificial intelligence for autism diagnosis using DTI and fMRI: A survey,” Biomedicines, vol. 11, no. 7, p. 1858, Jun. 2023, doi: 10.3390/biomedicines11071858.