How AI could help screen children for autism – USC Viterbi


A new adaptive computer test, powered by machine learning, could help doctors make more effective diagnoses of social communication disorders. Photo / iStock.

For children with autism spectrum disorders (ASD), receiving an early diagnosis can make a huge difference in improving behavior, skills, and language development. But although it is one of the most common developmental disabilities, affecting 1 in 54 children in the United States, it is not that easy to diagnose.

There is no lab test and no genetic cause identified. Instead, clinicians examine the child’s behavior and conduct structured interviews with the child’s caregivers based on questionnaires. But these questionnaires are large, complicated and not foolproof.

“When trying to discern and stratify a complex condition such as autism spectrum disorder, knowing what questions to ask and in what order becomes difficult,” said the USC University professor. Shrikanth Narayanan, Niki and Max Nikias Chair in Engineering and Professor of Electrical and Computer Engineering, Computer Science, Linguistics, Psychology, Pediatrics and Otolaryngology.

“As such, this system is difficult to administer and can produce false positives, or confuse ASD with other comorbid conditions, such as Attention Deficit Hyperactivity Disorder (ADHD).”

As a result, many children are not getting the treatment they need at a critical time.

PhD in computer science student and main author Victor Ardulov.

PhD in computer science student and main author Victor Ardulov.

An interdisciplinary team led by computer scientists at USC, working with clinical experts and autism researchers, hope to improve this by creating a faster, more reliable, and more accessible system for screening children with ASD. The AI-based method takes the form of an adaptive computer test, powered by machine learning, that helps clinical practitioners decide which questions to ask next in real time based on previous responses from caregivers.

“We wanted to maximize the diagnostic power of the interview by priming the clinician with an algorithm that can be more curious if necessary, but which will also try not to ask more questions than necessary,” said said the study’s lead author, Victor Ardulov. , a doctorate in computer science. student advised by Narayanan. “By training the algorithm in this way, you are optimizing it to be as efficient as possible with the information collected so far. “

Besides Narayanan and Ardulov, co-authors of the study published in Scientific reports of nature research are Victor Martinez and Krishna Somandepalli, both recent PhDs from USC. graduates; autism researchers Shuting Zheng, Emma Salzman, and Somer Bishop of the University of California at San Francisco; and Catherine Lord of the University of California at Los Angeles.

A game of 20 questions

In the study, the research team of computer scientists and clinical psychologists specifically examined the differentiation between ASD and ADHD in school-aged children. Both ASD and ADHD are neurodevelopmental disorders, which are often misdiagnosed for each other.

As such, children may be flagged as being at risk for diseases they may not have, potentially delaying correct assessment, diagnosis, and intervention. In fact, autism can be overdiagnosed in up to 9% of children, according to a study from the Centers for Disease Control and Prevention and the University of Washington.

The current process of diagnosing autism, from article "Robust diagnostic classification via Q-learning" courtesy of Ardulov et al.

The current autism diagnostic process, taken from the article “Robust diagnostic classification via Q-learning” courtesy of Ardulov et al.

To help make a diagnosis, the practitioner assesses the child’s communication skills and social behaviors by gathering their medical history and asking open-ended questions of caregivers. The questions cover, for example, repetitive behaviors or specific rituals, which could be characteristics of autism.

At the end of the process, an algorithm helps the practitioner to calculate a score, which is used as part of the diagnosis. But the questions asked do not change according to the respondent’s answers, which can lead to information overlaps and redundancies.

“This idea that we have all of this data, and we calculate all the numbers at the end, it’s not really a good diagnostic process,” Ardulov said. “The diagnosis is more like a game of 20 questions: What’s the next thing I can ask that helps me make the diagnosis more effectively? “

Maximize diagnostic accuracy

Instead, the researchers’ new method acts like a smart flowchart, adapting based on the respondent’s previous responses and recommending which item to ask next as more data about the child becomes available.

For example, if the child is able to carry on a conversation, it can be assumed that he has verbal communication skills. “So our model may suggest asking questions about speaking first and then deciding whether to ask questions on conversational skills based on the response. This effectively balances minimizing queries while maximizing the information gathered, ”Ardulov said.

They used Q-learning – a reinforcement learning method based on rewarding desired behaviors and punishing unwanted behaviors – to suggest steps to follow to differentiate disorders and make an accurate diagnosis.

“The diagnosis is more like a game of 20 questions: What’s the next thing I can ask that helps me make the diagnosis more effectively? »Victor Ardulov.

“Instead of just crunching the answers at the end, we said, here’s the next best question to ask during the process,” Ardulov said. “As a result, our models are better at making predictions when presented with less information.”

The test is not meant to replace a diagnosis by a trained clinician, the researchers said, but to help them make the diagnosis faster and more accurately.

“This research has the potential to enable clinicians to go through the diagnostic process more effectively, either faster or by alleviating some of the cognitive strain, which reduces the effect of burnout,” Ardulov said.

“It could also help doctors triage patients more efficiently and reach more people by acting as an app-based home screening method.”

While there is still work to be done before this technology is ready for clinical use, Narayanan said it is a promising proof of concept for adaptive interfaces in the diagnosis of communication disorders. social, and maybe more.

“Such an approach is really important because of its applicability not only within ASD,” said Narayanan. “It could also help diagnose many mental and behavioral health issues throughout life and around the world, including anxiety disorders, depression, substance abuse, and dementia all of which rely on similar procedures to understand them. and process them. “

Posted on November 9, 2021

Last updated on November 9, 2021