Classification of Movement Disorders Using Video Recordings of Gait with Attention-based Graph Convolutional Networks

Wei Tang; Peter M.A. van Ooijen; Deborah A. Sival; Natasha Maurits

Abstract

Early Onset Ataxia (EOA) and Developmental Coordination Disorder (DCD) are two pediatric movement disorders characterized by similar phenotypic traits, often complicating clinical differentiation. This resemblance poses substantial challenges in clinical practice as accurate diagnosis, a critical factor in determining appropriate treatment strategies, becomes increasingly intricate due to the similarity of symptoms. Despite the recognized reliability of current clinical scales like the Scale for the Assessment and Rating of Ataxia (SARA), their dependence on specialist expertise, time-consuming nature, and inherent subjectivity can potentially limit their efficacy in assessing movement disorders, thereby underscoring the need for more objective, and efficient diagnostic methods. This study introduces a novel approach that utilizes 2D video recording in the coronal plane coupled with pose estimation to differentiate gait patterns in children with EOA, DCD, and healthy controls. An attention-based Graph Convolutional Network (A-GCN) was proposed for the classification process, achieving an f1-score of 76% at the group level. The model incorporates channel-wise attention to stress the semantic nuances of body joints, and temporal attention to highlight important sequences in gait patterns. These mechanisms enhance the model's ability to accurately classify EOA and DCD. The promising results demonstrate the potential of this method in contributing to improved diagnosis and understanding of these movement disorders, thereby paving the way for more targeted treatment strategies. The code is available at https://github.com/jiudaa/Attention-basedGCN-EOA.git.

Keywords: Early Onset Ataxia (EOA); Developmental Coordination Disorder (DCD); Graph Convolutional Network (GCN); deep learning

Links
[Full text PDF][Bibtex]