New research published in Scientific Report We have introduced promising diagnostic tools that can dramatically reduce the long waiting times that many families face when seeking assessments of autism and attention-related conditions. Using artificial intelligence, the researchers analyzed subtle patterns of how people move their hands during simple tasks, and identified surprisingly accurately whether someone could have autism, attention deficit traits, or both. This method, which relies on wearable motion sensors and deep learning, could one day serve as a quick and objective screening tool to help clinicians triage their children for further evaluation.
Autism and attention deficit disorder are both classified as neurodevelopmental conditions. This means it affects brain growth and function. Spectral conditions for autism are usually identified by social interaction, communication differences, and repetitive behavioral challenges. Attention deficit disorders, which may include hyperactivity, are characterized by difficulties with focus, impulsivity, and persistent attention.
These are clear diagnosis, but in many cases they coexist in the same individual. In fact, up to 70% of people diagnosed with autism also show signs of attention-related difficulties. Despite their prevalence, diagnosing these conditions often remains a long and complex process involving interviews, surveys, and behavioral observations that can take months or years to schedule.
Researchers led by Jorge Jose have sought to develop new ways to detect differences in neurodevelopment using objective data. They focused on very basic behaviors. It's a movement to reach for something. While such movements may seem simple, research shows that subtle patterns of how someone moves their bodies can reveal a lot about how their brains process information.
Previous studies suggest that children with autism and attentional problems often exhibit differences in motor planning and coordination, even during early development. With these insights, researchers asked whether high-resolution motion tracking combined with machine learning could detect patterns specific to different neural developmental conditions.
“I was trained as a theoretical physicist. But for the past 20 years I have been working on quantitative neuroscience. I now have a lab. We started by studying the visual cortex of non-human primates, at Indiana University.
“A 2013 autism paper found that by looking at the natural movement on a millisecond time scale away from naked eye detection, each participant's movement was completely different and that autism was more random than neurotypes.
“In the current paper, we expanded the participant pool to include those that include ASD, ADHD, co-existing ADHD+ASD, and neural controls,” Jose said. “Importantly, we have developed a deep learning technique that allows for the diagnosis of new participants in just a few minutes with high accuracy. Looking at movements on such fine time scales also allows us to quantitatively assess the degree of severity.”
For their new study, the researchers recruited 92 participants, including autistic individuals, attention deficit characteristics, both conditions, and a comparison group without diagnosis. Participants were of age, ranging from children to young adults. Seventeen additional individuals were excluded from the final analysis due to task completion problems, motor disorders unrelated to conditions under study, or technical issues with the sensor.
All participants were able to perform a basic reach task that required them to touch the target on the screen, pull out their hands, and repeat the motions about 100 times. The movement of the dominant hand was tracked using a high-resolution sensor placed on the glove, which recorded the data at millisecond resolution.
In this study, we used two complementary approaches to analyse the data. First, the researchers applied deep learning techniques to raw motion data. Trained in thousands of exercise tests, artificial neural networks have learned to classify each participant as either autistic, attention deficit characteristics, or either. Network architectures relied on a type of model suitable for time-based sequences called long short-term memory cells that could capture both short-term and long-range patterns of data.
Researchers used cross-validation to carefully train and validate the models to test performance of data they had never seen before. When multiple types of motion data were combined, such as hand angle, velocity, and acceleration, the model achieved diagnostic accuracy to approximately 70%.
Jose was surprised by the fact that deep learning can detect inherent cognitive information properties of humans with high accuracy.
In addition to classification accuracy, researchers looked at another standard machine learning measure known as the area under the receiver operating characteristic curve (AUC). This metric evaluates how well a model can distinguish between categories. This model worked particularly well in identifying neuronal individuals with an AUC score of 0.95. Distinguishing the characteristics of autism and attention deficits has proven more difficult, especially for individuals with both conditions. This is known difficulties even in clinical settings.
To better understand the differences in underlying motion, the team also conducted a second analysis focusing on the statistical properties of motion data. After excluding electronic noise from the sensor signal, they measured how much randomness or variation existed in each person's movement.
Two important statistical tools were used. Fano factors that assess variation compared to the mean, and Shannon entropy that measures signal unpredictability. These measures provided a numerical fingerprint for each participant's style of movement, and as assessed by the clinician, higher levels of randomness tended to be adjusted along the severity of symptoms.
The researchers found that individuals diagnosed with both autism and attention deficit characteristics often exhibit intermediate levels of motor variability, which overlap with both neurotypic participants and only one diagnosis. Those with more severe autism symptoms had the most distinctive exercise patterns that were in good agreement with previous studies linking exercise differences to their condition.
Interestingly, biometrics were stable throughout the session. This means that 30-60 trials can be enough to get reliable measurements. This suggests that this method may be feasible in actual screening without the need for extensive testing.
New research provides evidence that “when the inherently random nature of human movements is considered to be visible to the human eye, it contains important information about their cognitive abilities.”
Although the findings of the study are promising, researchers acknowledge some limitations. The sample size was larger than many previous studies of this type, but still relatively small. In particular, groups that were only attention deficit characteristics were not very well represented, which could affect the extent to which the model generalizes to the wider population.
“These are preliminary findings that need to be further examined in a much larger group of nervous participants,” Jose noted.
This study did not control whether participants were taking medications such as meth and other treatments that could affect exercise, such as meth and other treatments. Furthermore, deep learning models can be difficult to interpret, and researchers have taken steps to visualize the importance of various input functions, but the algorithmic decision-making process remains a black box.
Despite these limitations, this study provides promising proof of concepts for new orientations in mental health diagnosis. By using affordable wearable sensors and machine learning, clinicians may one day be able to get early signs of differences in neurodevelopment without waiting for a full psychiatric evaluation.
Tools like this are particularly useful in rural and local areas where access to professionals is limited. Researchers hope that larger datasets can improve and expand the system to track changes over time, such as how motion patterns respond to treatment and development.
Going forward, the team plans to expand its approach to study how these motion-based metrics evolve with age and drug use. The ultimate goal is not to replace traditional diagnosis, but to provide clinicians with additional tools to support early detection and personalized care.
“We hope that our protocol will become another tool that providers can use to assess early on the status of Neurodivergent participants, which is much needed today,” Jose said.
The study, “Deep Learning Diagnosis and Kinematic Severity Assessment of Neuroproliferative Disorders,” was written by Khoshrav P. Doctor, Chaundy McKeever, Di Wu, Aditya Phadnis, Martin H. Plawecki, John I. Nurnberger Jr., and Jorge V. José.
