New study: AI can identify autism from tiny hand motion patterns

Subtle differences in how people grasp everyday objects may help identify autism, according to a new study published in Autism Research. By analyzing the fine motor movements of young adults during a simple grasping task, researchers were able to distinguish autistic from non-autistic individuals with more than 84% accuracy using machine learning techniques. These findings suggest that small variations in motor control may serve as reliable indicators of autism, potentially paving the way for non-invasive diagnostic tools based on natural behavior.

Autism is a neurodevelopmental condition marked by social and communication challenges as well as repetitive behaviors. But movement differences—often referred to as motor abnormalities—are also common and may appear in infancy, long before core social symptoms become apparent. Previous studies have shown that many autistic individuals exhibit altered coordination, clumsiness, or inconsistent movement patterns. These motor issues are now recognized as more than just secondary symptoms; they may reflect core differences in how autistic brains process sensory and motor information.

Because of this, researchers have begun investigating whether specific motor patterns could aid in identifying autism. In particular, reach-to-grasp movements—which are among the most basic and frequent actions people perform—offer a window into motor planning and control.

Earlier studies have reported that autistic individuals show distinctive patterns during these actions, such as slower movement speeds or delayed hand shaping. Still, most studies used artificial tasks or measured a limited set of movement features. This new research sought to test whether a naturalistic, real-world grasping task, using minimal equipment, could provide enough information to reliably classify participants based on autism diagnosis.

“My research centers on how the visual system extracts information to help us recognize and interact with objects in the world,” said study author Erez Freud, an associate professor and York Research Chair in Visual Cognitive Neuroscience at the Centre for Vision Research at York University.

“In recent years, my lab has become increasingly focused on how these abilities—perception and action—emerge throughout development and how they may be altered in cases of atypical development, such as in autism. We believe that by better understanding the specific perceptual and motor challenges autistic individuals face, we may be able to design early, objective diagnostic tools. These tools could significantly improve the support and interventions available to autistic individuals and their families.”

The researchers recruited 59 young adults—31 autistic and 28 non-autistic—matched in age and IQ. Each participant was asked to grasp a series of rectangular objects of varying lengths using just their thumb and index finger. The movements were recorded with a motion tracking system that monitored the three-dimensional positions of the fingers during each grasp. Only two passive markers—one on the thumb and one on the index finger—were used. This minimal setup allowed for a more practical and accessible approach than earlier studies that used full-body motion capture or complex equipment.

Each participant completed 120 grasping trials. From the recorded data, the researchers extracted over a dozen kinematic features, such as the maximum distance between fingers during a grasp, how fast the fingers moved, how long the movement lasted, and the trajectory of the fingers in space. These features captured key aspects of how the brain controls the hand during object interaction.

To explore whether these features could predict autism, the researchers trained five types of machine learning models, including logistic regression, support vector machines, and decision tree ensembles. Importantly, the models were trained and tested using a “leave-one-subject-out” approach. This means that the algorithm learned from all but one participant and then tried to predict whether that left-out participant was autistic. This process was repeated until every participant had been left out once. This method ensures that the results are not overfitted to any individual’s data and can generalize across people.

Across all five models, the accuracy of classifying autistic versus non-autistic participants exceeded 84%, with some models reaching 89%. These results remained strong even when evaluating performance on individual trials, not just averages across participants. In addition, the models achieved area under the curve (AUC) scores of over 0.95 at the subject level, meaning they had excellent discrimination ability.

“The key takeaway is that how we move—especially how we coordinate vision and action—can offer important insights into how the brain works,” Freud told PsyPost. “In autistic individuals, we find distinctive visual-motor patterns that could be leveraged as diagnostic markers, particularly when analyzed using computational methods. These subtle behavioral differences are often overlooked, yet they might hold critical information about underlying brain function.”

Interestingly, the researchers also tested whether fewer features could still yield good results. They created smaller models using just eight kinematic variables that were not highly correlated with each other. Even with this reduced feature set, classification accuracy remained high—above 82%. These features spanned different domains, including timing (like when the fingers reached maximum separation), velocity (how fast the hand moved), and spatial position (where the fingers traveled in space). This finding suggests that autism-related motor differences are spread across multiple aspects of grasping behavior, and no single feature explains everything.

In contrast, when the researchers selected features that were strongly correlated with each other—often ones commonly used in motor control research—the models performed worse. This highlights the importance of choosing features that capture distinct dimensions of movement when using machine learning for classification.

The study also replicated some known differences between autistic and non-autistic individuals at the group level. Autistic participants showed longer movement times, meaning they took more time to grasp the object. They also exhibited slightly different timing in how their fingers opened and closed during the movement, especially when object size varied. However, these group-level comparisons, while informative, were not as reliable for distinguishing individual participants as the machine learning models.

“Previous studies have documented differences in how autistic individuals move and grasp objects,” Freud said. “Our findings extend this work by showing that these differences are not just present, but measurable in a consistent and structured way—even during naturalistic, real-world tasks.”

The researchers emphasize that the strength of their approach lies in its simplicity. By using only two markers and a straightforward grasping task, they demonstrated that autism-related motor differences can be detected in a natural, unobtrusive setting. Unlike studies that rely on brain scans, complex diagnostic interviews, or artificial tasks, this method could be easily adapted for use in clinics or schools. It may also help identify autistic individuals who do not meet traditional diagnostic criteria but still experience challenges in motor control.

But the study has some limitations. The sample consisted only of young adults with normal IQ levels. It’s unclear whether the same methods would work as well in children, who are a primary target for early diagnosis. In addition, while the models performed well in distinguishing autistic from non-autistic participants, they did not attempt to classify subtypes of autism or predict symptom severity. Future studies could expand this approach to younger populations or explore whether grasping patterns correlate with other features, such as sensory sensitivities or social behaviors.

“This study focused exclusively on adults,” Freud noted. “We chose this approach to rule out the possibility that any observed differences were simply due to delayed development. However, to truly assess the diagnostic potential of these findings, future work needs to examine younger populations. Additionally, we are investigating whether these visuomotor signatures are specific to autism or whether they can also help distinguish between different neurodevelopmental conditions.”

“Our long-term goal is to understand how visuomotor behaviors unfold across the lifespan in both typical and atypical development. By integrating computational tools with behavioral data, we hope to uncover robust markers of brain function and dysfunction. Ultimately, this could lead to the creation of simple, accessible diagnostic tools based on everyday movements.”

“This study was the product of a collaborative effort,” Freud added. “It wouldn’t have been possible without the dedication and insight of my PhD student Zoha Ahmad, as well as my colleagues Professor Bat-Sheva Hadad from Haifa University and Professor Eitan Shelef from the University of Pittsburgh. Scientific progress is deeply collaborative, and this project is a great example of that.”

The study, “Effective Autism Classification Through Grasping Kinematics,” was authored by Erez Freud, Zoha Ahmad, Eitan Shelef, Bat Sheva Hadad.