A Global-Local Dynamic Directed Graph Neural Network for Parkinson's Disease Detection

Area of Science:

• Biomedical Engineering
• Machine Learning
• Gait Analysis

Background:

• Graph Neural Networks (GNNs) show promise for Parkinson's Disease (PD) diagnosis via gait analysis using Vertical Ground Reaction Force (VGRF) signals.
• Current GNN methods often overlook the dynamic nature of VGRF signal structures during walking, treating them as static.

Purpose of the Study:

• To propose a novel Global-Local Dynamic Directed Graph Neural Network (GLD2-GNN) to represent dynamic spatio-temporal features of VGRF signals.
• To address the limitations of static graph modeling in existing GNN-based gait analysis for PD.

Main Methods:

• Introduced the DyDGNN block, comprising Dynamic Graph Learning (DGL), Dynamic Directed Graph Network (DyDGN), and Temporal Convolutional Network (TCN) units.
• DGL learns dynamic VGRF signal topology; DyDGN extracts spatial patterns and dynamic topological features; TCN captures local temporal patterns.
• Evaluated using k-fold and cross-dataset validation on three datasets (Ga, Ju, Si).

Main Results:

• GLD2-GNN demonstrated superior performance compared to RFdGAD, Transformer, and AST-DGNN.
• Achieved an average improvement of 4.45% in accuracy, 2.93% in F1 score, and 2.88% in geometric mean across cross-dataset validation.
• Showcased strong representational ability for complex gait patterns and generalization across datasets.

Conclusions:

• GLD2-GNN effectively captures dynamic VGRF topological structures and spatio-temporal features for improved gait analysis.
• The model exhibits significant potential for PD diagnosis and rehabilitation by enhancing gait pattern recognition and cross-dataset generalization.
• Future work includes integrating GLD2-GNN with multi-modal approaches and developing a comprehensive gait analysis system.