Insights into gait performance in Parkinson's disease via latent features of deep graph neural networks

Jiecheng Wu^1,2Ning Su³Xinjin Li^1,2Chao Yao⁴Jipeng Zhang⁵Xucheng Zhang^1*Wei Sun^1*

• ¹Institute of Software, Chinese Academy of Sciences (CAS), Beijing, China
• ²School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing, China
• ³Chinese Academy of Medical Sciences and Peking Union Medical College, Dongcheng, Beijing Municipality, China
• ⁴University of Science and Technology Beijing, Beijing, Beijing Municipality, China
• ⁵Health Service Department of the Guard Bureau of the Joint Staff Department, Beijing, China

Introdcution: Parkinson's Disease (PD) is a progressive neurodegenerative disorder that primarily impacts motor function and is prevalent among older adults worldwide. Gait performance (such as speed, stride, step, and so on) has been shown to play a significant role in diagnosis, treatment, and rehabilitation. Fortunately, advancements in computer science have provided serial ways to calculate gait-related parameters, offering a more accurate alternative to the complex and often imprecise assessments traditionally relied upon by trained professionals. However, most of the current methods depend on data preprocessing and feature engineering, often require domain knowledge and laborious human involvement, and require additional manual adjustments when dealing with new tasks.To reduce the model's reliance on data preprocessing, feature engineering, and traversal rules, we employed the Spatial-Temporal Graph Convolutional Networks (ST-GCN) model. We also defined five distinct states within a complete gait cycle: standstill (S), left swing (L), double support (D), right swing (R), and turnaround (T). Using ST-GCN, we extracted spatial and temporal patterns from these five states directly from the data, thereby enhancing the accuracy of gait parameter calculation. Furthermore, to improve the interpretability of the ST-GCN model and increase its clinical relevance, we trained the model on data from both healthy individuals and PD patients. This allowed us to explore how the model's parameters (different ST-GCN Layers) could assist clinicians in understanding.The dataset used to evaluate the model in this paper includes motion data from 65 PD participants and 77 healthy control participants. Regarding the classification results from the 5 classifiers, ST-GCN achieved an average precision, recall, and F1-score of 93.48%, 93.21%, 1 Wu et al.and 93.32%, outperforming both the Transformer-based and LSTM-based methods. Displaying the joints and edge weights from various layers of the ST-GCN, particularly when comparing data from healthy individuals and PD patients, enhances the model's feasibility and offers greater interpretability. This approach is more informative than relying on a purely black-box model.This study demonstrated that the ST-GCN approach can effectively support accurate gait parameter assessment, assisting medical professionals in making diagnoses and reasonable rehabilitation plans for patients with PD.