A Novel Deep Learning Approach for Recognizing Stereotypical Motor Movements within and across Subjects on the Autism...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Autism Spectrum Disorder (ASD) is a neurodevelopmental condition marked by difficulties in social interaction and communication, often accompanied by repetitive behaviors like stereotypical motor movements (SMM).
• Accurate identification and tracking of SMMs are crucial for understanding ASD progression and developing effective interventions.
• Existing methods for SMM detection face challenges with individual variability and the need for extensive labeled data.

Purpose of the Study:

• To develop a robust deep learning framework for recognizing stereotypical motor movements (SMM) in Autism Spectrum Disorder (ASD).
• To address both intrasubject (within-individual) and intersubject (across-individuals) variability in SMM detection.
• To create a generalizable and efficient SMM detection system that mitigates the need for subject-specific labeled data.

Main Methods:

• Utilized convolutional neural networks (CNNs) for SMM recognition, analyzing signals in both time and frequency domains.
• Developed a subject-specific CNN model optimized for detecting SMMs within individuals, outperforming existing classification methods.
• Implemented a lightweight framework combining knowledge transfer and Support Vector Machine (SVM) classification for cross-subject SMM detection, addressing data scarcity.

Main Results:

• The proposed CNN model demonstrated superior performance in intrasubject SMM detection compared to state-of-the-art techniques.
• The cross-subject framework effectively identified SMMs without requiring subject-specific training data, resolving a key clinical challenge.
• Transfer learning across different domains proved effective for generalizing SMM detection, reducing the general need for supervised data.

Conclusions:

• Deep learning, particularly CNNs, offers a powerful approach for accurate and robust SMM detection in ASD.
• The developed framework successfully addresses variability challenges in SMM recognition, enhancing diagnostic and monitoring capabilities.
• The study highlights the potential of transfer learning to overcome data limitations in clinical applications of AI for ASD.