AI System Predicts Aggressive Behavior in Children With Autism
Among the biggest challenges facing caregivers of autistic children are outbursts in which they harm themselves or others. Such incidents can occur at unpredictable times and prevent autistic children from staying in school or going out in other community settings. In their most severe form, the outbursts can overwhelm caregivers and lead to long term hospital stays.
Now a team of researchers from the Institute of Experiential AI at Northeastern University has shown that a machine learning model analyzing physiological data from a wearable can accurately predict aggressive behaviors in children before they occur. In the study, which took place in psychiatric in-patient hospitals and involved 70 children with autism, the model was able to predict aggressive behaviors three minutes in advance of the outburst.
The approach holds promise for helping healthcare staff and caregivers prevent self-injury and mitigate other consequences of aggressive behaviors. If further validated and developed, the technology could also help families keep their children out of long-term care settings.
“We ultimately want to take this methodology and enable it to run in real-time,” says author Matthew Goodwin, a member of the Faculty Leadership Committee at the Institute. “That is, you could get the data from a sensor, send it to a phone, annotations of observed behavior are synchronized from the phone, both go to the cloud, our classifier runs, and then it sends a real-time push notification. We need to do a lot to get there, but that’s where we want this to go eventually so families can preemptively intervene and better manage the situation.”
Goodwin, a professor who holds a joint appointment in Northeastern University's Khoury College of Computer Sciences and Bouvé College of Health Sciences, co-authored the paper with Institute affiliate faculty member Deniz Erdogmus, senior research scientist Tales Imbiriba, and Northeastern University PhD candidates Ashutosh Singh and Ahmet Demirkaya.
The work is aimed at a truly massive societal issue. The Centers for Disease Control and Prevention estimates one in 36 children in the U.S. have autism. Of those children, roughly 27 percent have profound autism, meaning they are minimally verbal, have an IQ of less than 50, and require help with everyday tasks like eating and showering. Many people with severe autism engage in self-injury or other aggressive behaviors, forcing families to move them into long-term hospital settings.
“Because these kids can’t tell you what they’re experiencing, oftentimes these aggressive behaviors come out of the blue,” Goodwin says. “Caregivers can’t plan for them, they don’t know when they’re coming, and it’s that fear that the child is going to hurt themselves or someone else that keeps them secluded.”
Goodwin first had the idea of using biosensors to predict behavior in people with severe autism 20 years ago. But advances in sensor technology and AI only recently made studies like this possible (learn more about sensing technology in our webinar).
The researchers collected data for the study over 12 months across four hospitals. It used a commercially available sensor that measured physiological signals including cardiovascular activity, electrodermal activity, and motion. The researchers evaluated statistical techniques including logistic regression, support vector machines, neural networks, and domain adaptation. They found that logistic regression, a type of supervised machine learning, was the most successful approach, accurately predicting aggressive behavior 3 minutes before it occurred, though with more data,the researchers believe other approaches could work even better.
The researchers are already working on another project with an outpatient clinic to gather more data and further improve their system. They’d like to incorporate data without annotations, incorporate different sensors, and begin work developing the real-time alert system.
“We need to develop new techniques so we can, hopefully one day, really start impacting lives,” Imbiriba says.
The researchers plan to make the anonymized dataset they collected available to other machine learning scientists in an effort to push the field forward more quickly.
“As far as I’m aware, this is the first demonstration of this concept in a sizeable sample across different settings, where we could get a large enough dataset that we could subject it to machine learning,” Goodwin says.
In the future, they plan to test if the approach can also be used to predict aggressive behavior in other groups of people.
“There’s nothing specific about autism that makes me think physiology can’t be a proxy precursor for aggression elsewhere, and autism is not the only population with aggression,” Goodwin says. “People have reached out asking if this will work to predict aggression associated with PTSD, borderline personality disorder, self-harm in drug relapse, domestic violence, and people with aggressive histories in emergency room settings. This work may transcend and generalize not only in method but in population.”
The Institute for Experiential AI’s researchers are working on the forefront of AI innovation to build solutions to real-world problems. Read more about our research here and contact us to find out how we can help with your AI or data project.
