After several years and hundreds of fussy toddlers at the Duke Clinic, Duke professors were able to create a method of neurodivergence screening that can be downloaded on tablets and phones.
An article published by Nature Medicine on Oct. 2 presented the results of SenseToKnow, an app developed by an interdisciplinary team of researchers to screen 18- to 24-month-old children for autism.
The two senior authors of the SenseToKnow project are of different backgrounds: Geraldine Dawson, the William Cleland Professor of Psychiatry and Behavioral Science and director of the Center of Autism and Brain Development, and Guillermo Sapiro, a James B. Duke Distinguished Professor of Electrical and Computer Engineering.
“It was really an interdisciplinary, creative effort that brought together concepts in the world of autism research in developmental psychology and computer science and engineering,” Dawson said. “That, of course, was part of the fun and creativity of doing this kind of work.”
Due to the subjective nature of parent questionnaires that are currently relied on to diagnose autism, the accuracy of screening is often limited by literacy, parent understanding, accessibility of local health resources and cultural and family values.
Current autism screening struggles to include children of color and girls, according to Juan Matias Di Martino, Assistant Research Professor of Electrical and Computer Engineering and co-author of the foundational article. By screening autism through objective factors like gaze and behavior, he says that the app's tools are "built to be invariant to certain demographic properties."
To screen for autism among children, SenseToKnow presents visual and aural stimuli that potentially elicit different behaviors among autistic and nonautistic children, with a wide variety of different stimuli to account for the complex nature of autism spectrum disorder.
“We found that when you're very interested in something, you unconsciously will suppress the rate at which you're blinking,” Dawson said. “So that’s one of our biomarkers: what is the blink rate when an autistic child sees a social stimulus like a woman singing nursery rhymes, versus seeing a really cool object like a top that's spinning?”
While Dawson decided on stimuli that could distinguish a neurotypical child from one with autism, Sapiro and many others in the Electrical and Computer Engineering department used facial recognition and tracking technologies to quantitatively observe child reactions.
“We also have a stimuli where two people are taking turns in a conversation,” Di Martino, who studied computer vision and faces with Sapiro, said. “And so we measure the correlation between the gaze and who is talking. Are you following the conversation? Are you paying attention to the conversation back and forth?”
After collecting child reactions as quantified data, SenseToKnow processes a wide variety of different behaviors and presents a likelihood of autism, as well as the specific behavior patterns that led to this conclusion.
For several years, the SenseToKnow team has collected their own data for the app to compare and process its user data, due to the relatively novel approach of using machine learning to screen for neurodivergence.
“We grabbed kids that visited the Duke clinics at 18 months, and they used the app. We also got their clinical information,” Di Martino said. “In the sample that we received in primary care, we had everything that you can encounter with kids: we had kids that were neurotypical, we had kids with developmental delay, language delay and kids with autism.”
Through this method of sampling, SenseToKnow was able to achieve "classification of autism versus neurotypical development with sensitivity 87.8% and specificity 80.8%," with the capability of differentiating between behaviors associated with autism and developmental delay, which may be mistaken for autism.
“We're also looking at how certain co-occurring conditions might influence how the app performs,” Dawson said. “We have one study where one group of kids has autism only and another has autism and ADHD. And we're comparing to see how ADHD affects the biomarkers.”
The SenseToKnow team is also testing whether the app maintains accuracy at home, with no involvement from researchers or care workers, to ensure accessible autism screening for all demographics. Another future improvement planned for SenseToKnow is to increase the range of ages targeted by SenseToKnow, which is currently only 17 to 36 months old.
Hoping to conclude the SenseToKnow study this year, the team’s goal is for SenseToKnow to become a downloadable and accessible app that can be used alongside standard autism screening questionnaires. The team also hopes that their years of research will lead to precise screening of autism that is equally accessible and accurate among diverse demographics.
“[The paper] brought all of those biomarkers together in that sort of screening algorithm for the first time,” Dawson said. “And that's why it was a capstone paper for us that brought together all the work we've been doing over the last several years.”
Get The Chronicle straight to your inbox
Signup for our weekly newsletter. Cancel at any time.
Winston Qian is a Pratt first-year and a staff reporter for the news department.