Deep Learning Approaches for Classifying Children With and Without Autism Spectrum Disorder Using Inertial...

Area of Science:

• Neuroscience
• Machine Learning
• Developmental Pediatrics

Background:

• Autism spectrum disorder (ASD) diagnosis relies on behavioral assessments, lacking objective methods.
• Children with ASD often exhibit motor control differences, with 50-88% showing movement issues.
• Objective diagnostic tools are needed to improve ASD identification and intervention.

Purpose of the Study:

• To evaluate deep learning models for classifying ASD using hand-tracking data.
• To explore the potential of inertial measurement unit (IMU) data for objective ASD assessment.
• To identify movement differences indicative of ASD in children.

Main Methods:

• Collected IMU hand-tracking data from 41 children during a goal-directed arm movement task.
• Preprocessed IMU data using moving average and z-score normalization.
• Applied and validated various deep learning models, including convolutional autoencoders and LSTMs, using k-fold and patient-separated approaches.

Main Results:

• A convolutional autoencoder combined with LSTM layers achieved 90.21% accuracy and 90.02% F1-score.
• The best model, retrained on a patient-separated dataset, demonstrated generalization with 91.87% accuracy and 93.66% F1-score.
• Significant differences in physical movements between typically developing children and those with ASD were identified.

Conclusions:

• Deep learning models analyzing movement data show potential for facilitating ASD diagnosis.
• Hand-eye coordination analysis can identify movement differences associated with ASD.
• Small-scale models can achieve high accuracy and generalization for medical data classification, enabling future research.