Motion sensor biomarkers in neurodivergence

A novel artificial intelligence (AI) tool using motion-tracking data may one day help diagnose autism and attention-deficit/hyperactivity disorder (ADHD) in children with greater speed and precision, according to a study published in Scientific Reports.¹

The study was led by Jorge V. José, PhD, James H. Rudy professor of physics and adjunct professor of anatomy, cell biology, and physiology at Indiana University, who said this new tool has the potential to change how neurodivergent conditions are assessed.² His team demonstrated that deep learning models trained on high-resolution kinematic data could accurately distinguish between patients who are neurotypical and those with autism, ADHD, or both, based on subtle movement biomarkers invisible to the naked eye.³

Motion as a Diagnostic Window

Children with neurodivergent disorders often face long delays before receiving a formal diagnosis—up to 18 months in some areas, including Indiana, according to the news release.² Current diagnosis methods largely rely on behavioral observations and surveys, which can be time-consuming and subjective. This study proposes an objective and scalable alternative: using movement data captured during a simple reaching task to screen for neurodivergent traits.

“By studying the statistics of the motion fluctuations, invisible to the naked eye, we can assess the severity of a disorder in terms of a new set of biometrics,” José said. “No psychiatrist can currently tell you how serious a condition is.”

To test this, participants wore wireless motion sensors while performing touchscreen-based reaching tasks.¹ The sensors recorded linear acceleration, angular velocity, and roll-pitch-yaw (RPY) orientation at millisecond resolution. These data streams were then analyzed using a long short-term memory deep learning model, trained to classify participants into 1 of 4 categories based on whether they were neurotypical, had autism, had ADHD, or had comorbid autism and ADHD.

Strong Accuracy in Classification

The deep learning models achieved a mean test accuracy of 71.48% when using all 3 kinematic signal types, but classification performance varied by input type. RPY data alone yielded the highest individual signal accuracy at 67.83%, compared with just 44.44% with linear acceleration data and 32.17% with angular velocity.

Looking at combinations, RPY and linear acceleration data together yielded a 71.79% accuracy in categorizing participants, performing better as a pair than with angular velocity and with all 3 combined.

Notably, classification was most accurate for distinguishing patients who are neurotypical from those with neurodivergence. The tool was less reliable for identifying children who have both autism and ADHD, echoing clinical challenges with comorbid diagnoses.

“After training on a larger and more comprehensive dataset, the Deep Learning approach could play an important role as an early screening tool for participants suspected of having a neurodivergent disorder, not only in the clinic but also in schools and other non-medical settings,” the study team wrote. “With rapid improvements in sensor technology, MEM [micro-electromechanical] sensors are becoming more affordable, reliable, and ubiquitous (such as in smartphones and smartwatches) making the study of kinematic data for applications such as this increasingly relevant.”

Quantifying Severity With Biometrics

Beyond simply identifying neurodivergent conditions, the study also looked at novel biomarkers, specifically the Fano Factor and Shannon Entropy, based on the statistical patterns in participants’ micromovements. These metrics quantified randomness in movement, which the researchers linked to symptom severity.

Children with more severe autism or ADHD tended to have higher entropy and distinct fluctuation patterns in their acceleration data. For example, participants with low-functioning autism exhibited much greater variability in their hand motions than those with milder forms of the condition.

While the technology is not intended to replace physicians or a clinical diagnosis, the authors envision it as a triage or screening tool that could be deployed in primary care offices, schools, or telehealth settings, especially in underserved or high-wait regions.² José’s team estimates that a 15-minute session could be sufficient for data collection, making it suitable for early screening interventions.

“Some patients will need a significant number of services and specialized treatments,” José said in the news release. “If, however, the severity of a patient’s disorder is in the middle of the spectrum, their treatments can be more minutely adjusted, will be less demanding and often can be carried out at home, making their care more affordable and easier to carry out.”

References

1. Doctor KP, McKeever C, Wu D, et al. Deep learning diagnosis plus kinematic severity assessments of neurodivergent disorders. Sci Rep. Published online July 8, 2025. doi:10.1038/s41598-025-04294-9
2. Artificial intelligence used to improve speed and accuracy of autism and ADHD diagnoses. News release. EurekAlert. July 8, 2025. Accessed July 8, 2025. https://www.eurekalert.org/news-releases/1090448
3. Wu D, José JV, Nurnberger JI, Torres EB. A biomarker characterizing neurodevelopment with applications in autism. Sci Rep. 2018;8(1):614. doi:10.1038/s41598-017-18902-w