AI Tool Diagnoses ADHD and Autism in 15 Minutes

Attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorder (ASD) in children and adolescents can be diagnosed in as quickly as 15 minutes with up to 70% accuracy using a sensor-based tool, a new study showed.

“[Diagnosis] takes a very long time. This thing that we did in the initial diagnosis takes 15 minutes. It’s just incredible. We were in shock,” principal investigator Jorge V. José, PhD, told Medscape Medical News.

The researchers hypothesized that people’s random movements imperceptible to the naked eye contain important cognitive information. These movements can be picked up by high-definition sensors at roughly 220 snapshots per second.

“When you look at the millisecond time scale and you look at all these fluctuations, we can tell you if you are low-functioning, mid-functioning, or high-functioning and the whole spectrum of functionality and cognitive abilities,” said José, professor of physics and adjunct professor of anatomy at the Indiana University School of Medicine in Indianapolis.

The study was published online on July 8 in Nature Scientific Reports.

Long Diagnostic Wait Times

As rates of diagnoses continue to rise, the system is coming under increasing strain, and investigators believe this new tool could significantly cut long wait times for psychiatric assessment.

A 2023 CMS survey showed that 28% of US autism centers had diagnostic wait times in excess of 7 months, with some centers so overwhelmed they stopped accepting new referrals.

A 2018 study conducted by the same investigators identified key movement differences between individuals with and without ASD. This new study builds on these findings by expanding the number of kinematic variables measured with higher-definition Bluetooth sensors and by including patients with ADHD and those with comorbid ASD and ADHD.

The researchers used deep learning models to track subtle movement biomarkers that differentiate individuals with neurodevelopmental disorders (NDD) from neurotypical (NT) individuals. The models also accurately determined both the diagnosis and severity of the conditions. The study included 92 participants aged 5 years or older, with a mean age of 23.96 years and 35.9% identifying as females. Of these, roughly a third were NT.

To measure movement, participants wore a glove with sensors as they reached and retracted their hand toward a target on a touchscreen about 100 times for 15 minutes.

The deep learning model classified participants’ conditions based on kinematic variables, including roll, pitch, and yaw (RPY); linear acceleration; and angular velocity.

To assess severity, the data were filtered to remove high-frequency sensor noise, and individuals’ random kinematic fluctuations were measured using biometric Fano Factor and Shannon Entropy.

Artificial Intelligence (AI) ‘Sees’ What Physicians Can’t

The predictive ability of the deep learning model increased with multiple kinematic variables (mean test accuracy, 71.48%), but the angular velocity gave limited benefit to this combination. RPY had the largest predictive capability (test accuracy, 67.83%), relative to linear acceleration (test accuracy, 44.44%) and angular velocity (test accuracy, 32.17%).

“The area under the receiver operator characteristics curve (AUC) suggests that we can predict, with high accuracy, NDD participant’s conditions,” the investigators wrote.

The AUC on the validation sets was between 0.50 and 0.92, which increased slightly when they ignored angular velocity. The model performed consistently well at differentiating NDD patients from NT patients but less so at identifying comorbid ADHD and ASD.

One potential benefit of using this tool could be its ability to give objective data on a patient’s condition rather than purely relying on qualitative behavioral observations, the researchers noted.

José said his team plan to further test it in a range of settings, including schools and clinics, and conduct longitudinal studies. He remains excited about its potential to streamline the time-intensive and complex process of diagnosing and determining the severity of NDDs.

Promising but Still Preliminary

Commenting on the research for Medscape Medical News, Anna Van Meter, clinical psychologist and associate professor at the Department of Child and Adolescent Psychiatry at NYU Grossman School of Medicine, New York City, said that there’s “definitely great potential” for AI to help clinicians better use the data already collected from patients — and, as demonstrated in this study, to integrate “different types of data” that historically haven’t been part of psychiatric evaluations.

However, she noted that the research — although interesting — is still at “a very preliminary stage.”

One of the study’s limitations, said Van Meter, was its small sample size.

“Related to that it seemed like maybe about 20% of the people that they recruited, they weren’t actually able to use their data, which just raised some questions for me about the feasibility of this approach clinically,” she said.

She also pointed out that study participants were generally much older than the typical diagnostic age range, which is under 5 years for autism and between 5 and 12 years for ADHD.

This research was partially funded by the National Science Foundation grant 1640909 and grant 1735095 from the Interdisciplinary Training in Complex Networks (CM). José and Van Meter reported having no conflicts of interest.