Vision-Based Freezing of Gait Detection With Anatomic Directed Graph Representation

Area of Science:

• Biomedical Engineering
• Computer Vision
• Neurology

Background:

• Parkinson's disease affects millions globally, with freezing of gait (FoG) being a common and debilitating symptom.
• Current FoG assessment methods are time-consuming and subjective, necessitating objective and efficient alternatives.
• Computer-aided detection can significantly enhance the assessment of FoG, improving patient care and research.

Purpose of the Study:

• To develop an automatic, vision-based method for detecting freezing of gait (FoG) in Parkinson's disease patients.
• To create a computer-aided tool that provides objective and time-efficient FoG assessment.
• To leverage graph convolution neural networks for improved characterization of FoG patterns from video data.

Main Methods:

• Proposed a novel graph convolution neural network architecture to represent videos as directed graphs, with candidate regions as vertices.
• Implemented a weakly-supervised learning strategy and a weighted adjacency matrix estimation layer to reduce reliance on extensive data annotation.
• Incorporated global context from clinical videos and investigated fusion strategies to enhance detection accuracy.

Main Results:

• The proposed vision-based method achieved promising performance in automatic FoG detection.
• Demonstrated improved FoG detection by identifying key contributing regions and reducing interference from irrelevant visual information.
• Achieved an Area Under the Curve (AUC) of 0.887 on a dataset of over 100 videos from 45 patients.

Conclusions:

• The developed graph convolution neural network method offers an effective solution for automatic FoG detection in Parkinson's disease.
• This computer-aided approach provides an objective and efficient alternative to traditional clinical assessments.
• The findings suggest significant potential for improving the management and understanding of freezing of gait in Parkinson's disease.