New research demonstrated that analyzing simple finger movements during grasping the potential for an accessible diagnostic approach. Motor abnormalities are well-documented features of autism that often present as clumsiness, atypical movement patterns, and difficulties in motor planning and execution.
Researchers achieved classification accuracy above 84% and area under the curve (AUC) values exceeding 0.95 by using machine learning algorithms to analyze data from two motion-tracking markers placed on participants' thumbs and index fingers during natural reaching and grasping movements.
A recent meta-analysis revealed that 92.5% of reviewed studies reported significant motor impairments in 50% to 88% of children with autism, based on standardized tasks or questionnaires, consistent with developmental coordination disorder. These motor symptoms often manifest in early childhood—before classical social communication deficits appear—making them potential targets for early diagnosis.
Study Design and Methodology
The researchers, led by Erez Freud, of York University, analyzed reach-to-grasp movements in 31 patients with autism and 28 without. All had normal IQ levels. Using an Optitrack system with six cameras, the researchers tracked the three-dimensional trajectories of the participants' index fingers, thumbs, and wrists at 100 Hz as they grasped and lifted objects of different sizes.
Key Findings
Overall accuracy across the evaluated classifiers was consistently high, ranging from approximately 84% to 89% when the full training data set was employed. Logistic Regression and XGBoost achieved the highest overall accuracies at about 89%, whereas k-NN and Random Forest were numerically the lowest at about 84%.
Sensitivity ranged from 85% to 89%, whereas specificity varied between 84% and 90%.
The researchers also conducted receiver operating characteristic (ROC) analysis. At the subject level, the classifiers achieved AUC values ranging from 0.96 to 0.98, indicating that for any randomly selected pair of participants, the classifier had a 96% to 98% probability of correctly ranking the individuals.
Even at the individual trial level, ROC analysis yielded AUC values in the range of 0.85 to 0.88, meaning that on any given pair of trials, there was an 85% to 88% likelihood that the classifier would correctly distinguish between the autistic and nonautistic instances.
Notably, the researchers found that comparable accuracy to the full model could be achieved with as few as eight features, provided these features were relatively less correlated.
Clinical Implications
"The current study provides strong evidence that grasping movements are strongly diagnostic of autism, and that [machine learning] techniques can be utilized to enhance the robustness of such diagnosis," the study authors wrote. "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 feature."
Research Context and Limitations
This study built on previous research showing distinctive kinematic properties in grasping movements among autistic individuals. For instance, the researchers found that participants with autism exhibited longer movement times compared with those without autism, consistent with previous findings.
The researchers acknowledged limitations. The study sample consisted of young adults with normal IQ levels, limiting the generalizability of these results to younger children or individuals with cognitive impairments. They recommended that future research should investigate whether similar kinematic features can be observed in younger populations, particularly early in childhood when the visuomotor system is still developing.
The research team also suggested exploring the integration of other sensory-motor measures such as eye-tracking to provide a more comprehensive assessment of motor alterations in autism.
The study was supported by multiple funding sources, including the Natural Sciences and Engineering Research Council of Canada, the Vision Science to Applications program, the Israel Science Foundation, and the United States–Israel Binational Science Foundation.
The authors declared having no competing interests.
Source: Autism Research