Wearable Biosensors Predicts Aggressive Behaviors in Autism, Study Finds

Published on: 

After participants with autism wore biosensors tracking physiological signals, a logistic regression predicted aggressive behaviors 3 minutes before it happened.

Wearing biosensors and machine learning analyses predicted imminent aggressive behaviors in youth with autism, according to a new study.1

Aggressive behaviors, including self-harm, tantrums, meltdowns, property destruction, and aggression toward others, may serve as a challenge for individuals with autism. Approximately 70% of children and adolescents with autism demonstrate such aggressive behaviors.

In a previous study conducted by the same investigators, 20 youths with autism wore biosensors which recorded peripheral physiological and motion signals. Then, the investigators used logistic regression to evaluate if they could predict aggressive behaviors. Because individuals with autism struggle with self-awareness and expressing their emotions, the sensor and the machine learning analyses allow people to know when to expect an aggressive behavior.

“If we could give caregivers advance notice, it would prevent them from getting caught off guard and potentially allow them to relax the individual and make sure everyone in the environment is safe,” Matthew Goodwin, investigator of both the study with 20 participants the newer replication study, told Northeastern Global News.2

A new study, led by Tales Imbiriba, PhD, and Ahmet Demirkaya, MS, both from Northeastern University, replicated the earlier study with a larger population sample. Again, they sought to find out if a wearable biosensor and machine learning could predict aggressive behaviors.1

The non-interventional prognostic study included 70 participants with autism who exhibited self-harm behavior, emotion dysregulation, or aggression toward others. Participants had a mean age of 11.9 years old, and majority were male (88.6%, n = 62). Moreover, most of the participants were white (90%, n = 63), while there were 1 Asian (1.4%), 5 Black (7.1%), 1 Native Hawaiian or Other Pacific Islander (1.4%), 5 Hispanic (7.5%). Then, 62 participants reported being non-Hispanic (92.5%).

Many participants were minimally verbal (45.7%, n = 32), and 42.8% of participants had an intellectual disability (n = 30). Participants had an inpatient hospital stay anywhere from 8 – 201 days, with a mean stay of 27.28 days.

Data was collected from March 2019 – March 2020 from 4 primary care psychiatric inpatient hospitals, and the data was analyzed from March 2020 – October 2023.

The wearable biosensor recorded peripheral physiological signals, such as heart rate, sweat production, skin surface temperature, and arm movements.2 The investigators then analyzed the data using logistic regression, support vector machines, neural networks, and domain adaptations.1

During the study, 429 naturalistic observational coding sessions were recorded, a total of 497 hours. The investigators observed 6665 documented aggressive behaviors, with 3983 behaviors of self-harm (59.8%), 2063 behaviors of emotion dysregulation (31%), and 619 behaviors of aggression toward others (9.3%).

The logistic regression performed the best, predicting aggressive behaviors 3 minutes before they happened with a mean area under the receiver operating characteristic curve (AUROC) of 0.80 (95% CI, 0.79 – 0.81). Thus, wearable biosensors and machine learning analyses can predict imminent aggressive behaviors in youths with autism.

The neural networks performed the second best (AUROC = 0.81; 95% CI, 0.79 – 0.83). Moreover, the support vector machines performed well in predicting the farthest ahead (τf = 180 seconds; AUROC = 0.85; 95%CI, 0.84-0.86), with the logistic regression shortly following.

“While [support vector machines] produced the best single AUROC result, it did not perform as consistently well across all experiments as [logistic regression],” the investigators wrote. “Hence, we considered [logistic regression] our best performing classifier.”

The investigators noted that, when they put the biosensor data into a learning machine, they were able to differentiate between the different types and intensities of aggressive behavior. They wrote how self-injurious behavior was the most predictable aggressive behavior. However, they reasoned this could possibly be a power issue since the self-injurious behavior episodes were 1.93 times more frequent than emotion dysregulation episodes and 6.43 times more frequent than episodes of aggression toward others.

“Our findings may lay the groundwork for developing just-in-time adaptive intervention mobile health systems that may enable new opportunities for preemptive intervention,” the investigators wrote. “By focusing on reducing the unpredictability of aggressive behavior, we anticipate that this ongoing research program may enable inpatient youths with autism to more fully participate in their homes, schools, and communities.”