TL;DR
Researchers have developed the Computerised Assessment of Motor Imitation (CAMI), a one-minute, videogame-like task that uses computer vision to objectively measure a child's ability to copy movements. The study demonstrates that CAMI can specifically identify autism-related imitation difficulties, successfully distinguishing children with Autism Spectrum Disorder (ASD) from both neurotypical children and those with ADHD—a common diagnostic "look-alike."
What Problem Does This Paper Address?
Diagnosing Autism Spectrum Disorder (ASD) is often complicated by the high rate of co-occurrence with Attention Deficit/Hyperactivity Disorder (ADHD); roughly 50-70% of children with autism also meet the criteria for ADHD. Because these conditions share certain behavioral overlaps, clinical pictures can become "clouded," leading to delayed or missed diagnoses.
While motor imitation (the ability to copy someone else's physical actions) has long been recognized as a core deficit in ASD, traditional assessments rely on human observers. These manual methods are time-consuming, subjective, and difficult to scale. The authors sought to determine if an automated, objective, and rapid tool could provide a "biomarker" for ASD that remains accurate even when ADHD symptoms are present.
Methods
The core of the study is CAMI (Computerised Assessment of Motor Imitation).
- The Task: Children stand between two sensorless Microsoft Kinect cameras and are asked to copy the "dance-like" movements of a video avatar for one minute.
- The Algorithm: The system uses computer vision to track joint movements in 3D space. It automatically identifies the most critical joints for each move and calculates an imitation score ranging from 0 (no imitation) to 1 (perfect imitation, based on a researcher's gold-standard performance).
- Validation: To ensure CAMI was measuring what it claimed to, the researchers compared results against "gold standard" clinical tools: the ADOS-2 (observational autism measure), SRS-2 (parent-report autism traits), Conners (ADHD traits), and PANESS (general motor coordination).
- Analysis: The team used linear regression to find associations and employed Machine Learning—specifically Support Vector Machines (SVM) and Receiver Operating Curve (ROC) analysis—to test how well the CAMI score alone could categorize a child's diagnosis.
Datasets
The study involved a cross-sectional sample of 183 children aged 7–13 years, divided into four groups:
- ADHD only (n=35)
- ASD + ADHD (n=63)
- ASD only (n=20)
- Neurotypical (NT) (n=65)
All participants had a full-scale IQ of ≥ 70 to ensure they could follow the game's instructions.
Key Insights and Ideas
- Specificity is Key: The most striking finding is that imitation difficulties are specific to the autism phenotype. Children with ADHD (without ASD) performed nearly identically to neurotypical children. This suggests that poor imitation isn't just a general sign of neurodevelopmental struggle, but a specific marker for ASD.
- Independence from ADHD and General Motor Skill: In children with ASD, poor CAMI performance correlated strongly with core autism traits (like social affect and repetitive behaviors) but did not correlate with ADHD symptoms or general motor coordination (PANESS scores).
- Divergent Pathways: Interestingly, for neurotypical children, better imitation was associated with better attention and general motor skills. For children with ASD, these factors didn't matter; their imitation performance was tied strictly to their autism-specific social-communicative profile.
- Scalability: Because the task takes only one minute and requires no wearable sensors or expert coding, it represents a massive leap in potential clinical efficiency.
Key Results
- Classification Accuracy: CAMI distinguished ASD children from neurotypical children with an 80% true positive rate (AUC = 0.80) and from ADHD children with a 70% true positive rate (AUC = 0.70).
- Statistical Significance: ASD children (regardless of whether they had co-occurring ADHD) showed significantly poorer performance than NT children (p < .0001).
- Machine Learning Validation: Using a 3-fold cross-validated SVM, the CAMI scores alone achieved 71.6% accuracy in separating ASD from NT and 66.4% accuracy in separating ASD from ADHD.
- The ADHD/NT Overlap: The classification rate for ADHD vs. NT was only 59.2%, which was not significantly better than chance, further proving that ADHD does not fundamentally impair this specific type of motor imitation.
Limitations and Future Work
The authors acknowledge several areas for improvement:
- Sample Size: The "ASD-only" group was relatively small (n=20), which may have limited the statistical power to find subtle differences between them and the ADHD group.
- Intellectual Range: The study only included children with an IQ > 70. Future iterations need to simplify instructions to include children with intellectual disabilities to represent the full autism spectrum.
- Age Range: While 7–13 is a critical window, the average age of ASD diagnosis is currently around 10 years old. The researchers aim to adapt CAMI for younger children to enable earlier intervention.
- Directionality: It remains unclear if impaired social communication leads to imitation trouble, or if early imitation deficits cascade into social-communicative issues. Longitudinal studies are needed to untangle this "chicken or egg" problem.
Summary updated