Closing the Gap in Autism Screening and Care Through Interdisciplinary Innovation

“The average age of diagnosis of autism in the United States is close to five years,” says Geraldine Dawson, founding director of the Duke Center for Autism and Brain Development and the William Cleland Distinguished Professor of Psychiatry and Behavioral Sciences.

Yet the earliest behavioral signs of autism can appear much sooner — between six and 12 months of age — and routine screening typically doesn’t occur until 18 to 24 months. 

“There’s a huge disparity between when we can recognize the early signs and when a diagnosis is actually being made,” Dawson says. “It’s very important to identify children early so that they can have access to early interventions, because they’ve been shown to really have a substantial impact on long-term outcomes.”

One challenge is that the standard screening relies heavily on a questionnaire parents fill out. 

“It particularly misses girls, and it also misses children whose parents may not have the knowledge of child development to be able to respond to the questions,” says Dawson. 

For nearly a decade, Dawson and collaborators across Duke have been working to close those gaps. 

Interdisciplinary Connection and Seed Funding

Image

Shortly after Dawson arrived at Duke, an interdisciplinary connection helped reshape that effort. At a faculty event, she met Guillermo Sapiro, a world‑renowned expert in computer vision and Duke engineering professor who has since moved to Princeton University and Apple, Inc.

Their conversation led to a Bass Connections project that brought together students and faculty from across campus to explore whether computer vision analysis of behavioral signs could detect autism in young infants and toddlers.  

“This project has been about 10 years in the making, and it started with a vision of having a more accurate, objective screening tool for autism,” Dawson says. “This turned out to be very much an interdisciplinary task that required the combination of people with clinical expertise and developmental expertise, like myself, with people in engineering that could help develop these tools that could automatically quantify these behavioral signs.”

Duke seed funding, including from research grant program ABC Thrive, laid the groundwork for what would eventually become a digital screening technology.

SenseToKnow: A New Way to Screen for Autism

Image

The work has grown into a multiyear research program supported by major grants from the National Institutes of Health as part of an NIH Autism Center of Excellence award. 

The team developed SenseToKnow, an app parents can use on a smartphone or tablet. Toddlers watch short videos while the device camera records their responses using the camera in the device. 

“The child’s responses can show early signs of autism,” Dawson says. “For example, we know that toddlers who will go on to have a diagnosis of autism tend to pay more attention to non-social information than social information. So we designed a video where the social information was on one side of the screen and the non-social on the other side.”

In the video, a person blowing bubbles appears on one side of the screen, and the container and the bubbles appear on the other.

As the child watches the videos, the device records their facial expressions, eye gaze and movements, and then uploads the recordings to a secure server at Duke, where AI-based computer vision tools automatically analyze the data. 

Validation and Results

Using the app and AI analysis, Dawson says the team has identified 23 different digital phenotypes (specific behaviors technology can measure) including novel biomarkers (newly observed behavioral signs) that may indicate early signs of autism. 

“That was what was so fascinating for me, having studied autism for so long and thought I knew everything,” Dawson says. “I learned a lot because the computer is able to detect very subtle variations in behavior. The resolution and the precision are just so much greater than the human eye.”

One newly identified biomarker involves how children respond to social information. “We found that toddlers who would later be diagnosed with autism later would not suppress their blink rate when they were looking at social information as compared to a neurotypical child who would,” Dawson said. Humans spontaneously suppress their blink rate when they are interested what they are seeing.

The team also observed that the autistic toddlers made subtle back and forth repetitive movements of their heads. 

Over several years, supported by NIH funding, the researchers tested the app in Duke primary pediatric care clinics and in people’s homes. 

In one study conducted in primary care settings, the team used AI to combine the 23 digital phenotypes captured by the app into an algorithm capable of predicting whether a child would later receive an autism diagnosis. 

“We were showing for the first time that we could use this as a tool for screening and showed very good accuracy,” Dawson says. “One of the really gratifying things about this study is that we did not see the discrepancy in accuracy for girls versus boys or across the different racial and ethnic minority populations.” The results were published in Nature Medicine.

A second study examined whether parents could administer the screening themselves at home using their own iPhone or iPad. Published in The New England Journal of Medicine AI, they showed the approach could work reliably across devices and settings, and again the accuracy level was consistent across boys and girls and different ethnic and racial backgrounds.  

In 2023, the NIH recognized the digital autism screening app as a major human health advance. 

Next: FDA Approval

Now the research is moving toward regulatory review. 

With their latest NIH grant, Dawson and team are seeking clearance from the U.S. Food and Drug Administration. With guidance from the FDA, they are conducting a study looking at 200 children without an autism diagnosis and 150 with an autism diagnosis, who are recruited participants through Duke Primary Care. 

“Parents download and administer the app, and then every child in the study is given a gold standard diagnostic assessment. So we’re comparing the accuracy of the app to an expert clinical diagnosis of autism on every child.”

The team is also conducting three additional studies recommended by the FDA: a repeatability study, a reproducibility study and a human factors study. 

“We’ll analyze all those data and then this summer take those data to the FDA and hopefully get FDA clearance,” Dawson notes.

If cleared, SenseToKnow would be used through healthcare providers rather than directly by consumers. 

“It is envisioned as a prescription,” says Dawson. “The pediatrician or other healthcare provider would prescribe it. It’s good to get that kind of information from a physician who can then provide guidance.”

A Larger Duke Effort 

The screening app is part of a broader effort at Duke to improve early detection of autism using multiple data sources. Other faculty at the Duke Center for Autism & Brain Development are studying electronic health records to see how patterns of healthcare utilization detected in electronic health records can signal increased likelihood of autism even in infancy. Other projects focus on ethical issues related to the use of AI in autism screening.

Together, these approaches aim to give pediatricians a more complete picture for when to make a referral for an evaluation of a child’s development.

“The SenseToKnow screening app is really one piece of a larger vision that we have for improving early access to care,” Dawson concludes.

Main image: Geraldine Dawson smiles in her office. (Photo: Les Todd)