Graph Sequence Recurrent Neural Network for Vision-based Freezing of Gait Detection

Area of Science:

• Neurology
• Biomedical Engineering
• Computer Science

Background:

• Freezing of gait (FoG) is a prevalent and debilitating symptom of Parkinson's disease (PD).
• Current FoG assessment relies on manual, time-intensive video analysis by experts.
• There is a critical need for automated FoG detection methods to improve patient care and treatment assessment.

Purpose of the Study:

• To develop an automated, vision-based algorithm for detecting freezing of gait (FoG) in Parkinson's disease (PD).
• To model FoG detection as a fine-grained graph sequence task using anatomical joint movements.
• To introduce a novel deep learning approach for analyzing dynamic graph structures in patient videos.

Main Methods:

• Formulated vision-based FoG detection as a fine-grained graph sequence modeling task.
• Developed a novel deep learning model, the graph sequence recurrent neural network (GS-RNN), utilizing graph recurrent cells.
• Proposed a data-driven adjacency estimation method for cases lacking prior edge annotations.

Main Results:

• The proposed GS-RNN demonstrated promising performance in FoG detection.
• Achieved an Area Under the Curve (AUC) value of 0.90 in experiments.
• Validated on over 150 videos from 45 Parkinson's disease patients.

Conclusions:

• The GS-RNN offers an effective deep learning solution for automated, vision-based freezing of gait detection.
• This study pioneers the use of deep neural networks for graph sequences of dynamic structures in FoG analysis.
• The findings support the potential of automated methods to enhance the assessment and management of Parkinson's disease symptoms.