Researchers from York University and the University of Haifa have demonstrated that machine learning analysis of naturalistic hand movements during simple grasping tasks can classify autism with approximately 85 percent accuracy. This breakthrough suggests that subtle motor differences could provide valuable diagnostic markers for autism spectrum disorder, potentially enabling earlier intervention and improved outcomes.
New diagnostic approach leverages motor markers in autism
A groundbreaking study published in the journal Autism Research on May 5, 2025, has identified a potentially simpler approach to diagnosing autism spectrum disorder through the analysis of hand movements during everyday grasping tasks.
The international research team, led by Associate Professor Erez Freud from York University’s Department of Psychology and the Centre for Vision Research, used machine learning to analyse naturalistic finger movements during grasping in autistic and non-autistic young adults.
“Our models were able to classify autism with approximately 85 per cent accuracy, suggesting this approach could potentially offer simpler, scalable tools for diagnosis,” says Freud.
Motor abnormalities provide early diagnostic opportunity
Autism spectrum disorder affects approximately one in 50 Canadian children and is typically diagnosed through behavioural assessments that focus on social communication challenges and repetitive behaviours. However, these characteristic markers often appear relatively late in development.
The research team notes that motor abnormalities, which are widely documented in autism, frequently manifest in early childhood and could potentially provide earlier diagnostic signals – something not yet widely leveraged in clinical practice.
“The main behaviours markers for diagnosis are focused on those with relatively late onset and the motor markers that can be captured very early in childhood may thus lower age of diagnosis,” explains Professor Batsheva Hadad of the University of Haifa, a key collaborator in the study.
Study methodology focused on natural movements
The researchers recruited 31 autistic and 28 non-autistic young adults with normal IQ scores. Participants were asked to perform a simple grasping task – using their thumbs and index fingers to grasp, lift, and replace blocks of varying sizes while tracking markers attached to their fingers captured precise movement data.
By focusing on young adults rather than children, the researchers ensured that any differences observed could not be attributed to developmental delays but instead reflected fundamental differences in motor control.
The research team used five different machine learning classifiers to analyse the data, achieving consistent classification accuracy above 84% across all models. When examining the area under the curve (AUC) – a measure of classification reliability – they achieved scores above 0.95 at the subject-wise analysis and above 0.85 at the trial-wise analysis.
Classification possible with minimal features
One particularly promising finding was that the classifiers maintained high accuracy even when using a reduced set of features. With just eight carefully selected, minimally-correlated features spanning multiple domains – including experimental condition, timing information, velocity data, and location information – the classifiers achieved 82-86% accuracy.
“These findings suggest that subtle motor control differences can be effectively captured, offering a promising approach for developing accessible and reliable diagnostic tools for autism,” note the authors in their paper.
Implications for earlier diagnosis and intervention
The authors highlight the potential clinical significance of their work in the paper’s conclusion: “The current study provides strong evidence that grasping movements are strongly diagnostic of autism, and that ML techniques can be utilized to enhance the robustness of such diagnosis. By focusing on naturalistic tasks and minimal data inputs, our approach offers a promising avenue for developing objective, accessible, simple, and reliable diagnostic tools for autism based on motor control features.”
This approach could complement existing diagnostic methods and potentially enable earlier intervention, which is known to improve outcomes for autistic individuals. The researchers suggest that further studies should explore whether similar kinematic markers can be observed in younger populations, particularly in early childhood when the visuomotor system is still developing.
Reference:
Freud, E., Ahmad, Z., Shelef, E., et. al. (2025). Effective autism classification through grasping kinematics. Autism Research, 0(0), 1-12. https://doi.org/10.1002/aur.70049
