Keypoints-Based Multi-Cue Feature Fusion Network (MF-Net) for Action Recognition of ADHD Children in TOVA Assessment

Area of Science:

• Computer Science
• Neuroscience
• Biomedical Engineering

Background:

• Attention Deficit Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder in children and adolescents.
• Behavioral analysis is vital for ADHD diagnosis, but existing methods struggle with ADHD-specific movements.
• Current algorithms often overlook the unique behavioral patterns associated with ADHD.

Purpose of the Study:

• To develop a novel keypoints-based system for recognizing ADHD-specific behaviors in children.
• To accurately assess ADHD symptoms using human body and facial keypoint data.
• To improve the objective quantification of hyperactivity and impulsivity in ADHD diagnosis.

Main Methods:

• Proposed the Multi-cue Feature Fusion Network (MF-Net), a keypoints-based system for ADHD behavior recognition.
• Utilized a Multi-scale Features and Frame-Attention Adaptive Graph Convolutional Network (MSF-AGCN) for body keypoint analysis.
• Employed MobileVitv2 for facial keypoint analysis after image transformation, integrating both body and facial features.

Main Results:

• The MF-Net system achieved 90.6% top-1 accuracy and 97.6% top-2 accuracy on 3801 video samples of children with ADHD.
• The system demonstrated robust performance across six action categories relevant to ADHD.
• Validation on public datasets (NW-UCLA, NTU-2D, AFEW-VA) confirmed the network's effectiveness.

Conclusions:

• The developed MF-Net system offers a promising approach for objective ADHD behavioral assessment.
• The fusion of body and facial keypoint features enhances the accuracy of recognizing ADHD-specific actions.
• This technology has the potential to significantly aid in the diagnosis and monitoring of ADHD in children.