Computerized Assessment of Motor Imitation for Distinguishing Autism in Video (CAMI-2DNet).

Area of Science:

• Neuroscience
• Computer Science
• Developmental Psychology

Background:

• Motor imitation impairments are common in autism spectrum conditions (ASCs), presenting a potential phenotype for understanding autism heterogeneity.
• Traditional motor imitation assessments are subjective and labor-intensive, while current computerized methods (CAMI-3D, CAMI-2D) require extensive data preprocessing and human annotation.

Purpose of the Study:

• To introduce CAMI-2DNet, a scalable and interpretable deep learning approach for automated motor imitation assessment using video data.
• To eliminate the need for ad hoc normalization, data cleaning, and human annotations in motor imitation analysis.

Main Methods:

• Developed an encoder-decoder deep learning architecture (CAMI-2DNet) to create motion representations disentangled from nuisance factors.
• Utilized a combination of synthetic data (motion retargeting) and real participant data for training the model.
• Computed similarity scores between motion encodings to discriminate between individuals with ASCs and neurotypical (NT) individuals.

Main Results:

• CAMI-2DNet demonstrated a strong correlation with human scoring of motor imitation.
• The method outperformed existing CAMI-2D in discriminating between children with ASCs and NT children.
• CAMI-2DNet achieved performance comparable to CAMI-3D but with greater practicality, directly using video data without extensive preprocessing.

Conclusions:

• CAMI-2DNet offers a practical, scalable, and automated solution for motor imitation assessment in videos.
• This deep learning approach has the potential to advance research into autism heterogeneity and improve diagnostic tools.
• The method's ability to work directly with video data significantly reduces the labor and subjectivity associated with traditional and current computerized methods.