{"title":"一种用于帕金森病检测的全局-局部动态有向图神经网络。","authors":"Xiaotian Wang, Guanhai Zhou, Zhifu Zhao, Xiaoyi Zhang, Fu Li, Fei Qi","doi":"10.1109/TNSRE.2025.3614430","DOIUrl":null,"url":null,"abstract":"<p><p>Graph Neural Networks (GNNs) for gait analysis utilizing Vertical Ground Reaction Force (VGRF) signals have demonstrated significant potential in Parkinson's Disease (PD) diagnosis and rehabilitation fields. However, existing GNN-based methods in this area normally model the VGRF signals as static topological structures, and ignore the dynamic variations of the VGRF graph structures during walking. To address this issue, a Global-Local Dynamic Directed Graph Neural Network (GLD<sup>2</sup>-GNN) is proposed to represent the dynamic spatio-temporal features of VGRF signals. The core component of the proposed model is the DyDGNN block, which is composed of a Dynamic Graph Learning (DGL) unit, a Dynamic Directed Graph Network (DyDGN) unit, and a Temporal Convolutional Network (TCN) unit. First, the DGL unit is proposed to learn dynamic topological relationships of VGRF signals. Based on learned graph structures, the DyDGN unit is constructed to extract the spatial patterns and capture topological dynamic features from VGRF signals. Subsequently, local temporal patterns of VGRF signals are extracted by the TCN unit. The proposed method is evaluated through k-fold cross-validation and cross-dataset validation on three datasets Ga, Ju and Si. Compared with existing methods, such as RFdGAD, Transformer, and AST-DGNN, GLD<sup>2</sup>-GNN demonstrates superior performance in the validation experiments. Notably, our method achieves an average improvement of 4.45% in accuracy, 2.93% in F1 score, and 2.88% in geometric mean across cross-dataset validation. Extensive experiments have demonstrated that GLD<sup>2</sup>-GNN exhibits both the representational ability for complex gait patterns and the generalization ability across various datasets by capturing dynamics of VGRF topological structures and spatio-temporal features from VGRF signals. For future work, we plan to combine our method with multi-modal methods and integrate our framework into a complete gait analysis system.</p>","PeriodicalId":13419,"journal":{"name":"IEEE Transactions on Neural Systems and Rehabilitation Engineering","volume":"PP ","pages":""},"PeriodicalIF":5.2000,"publicationDate":"2025-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"A Global-Local Dynamic Directed Graph Neural Network for Parkinson's Disease Detection.\",\"authors\":\"Xiaotian Wang, Guanhai Zhou, Zhifu Zhao, Xiaoyi Zhang, Fu Li, Fei Qi\",\"doi\":\"10.1109/TNSRE.2025.3614430\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p><p>Graph Neural Networks (GNNs) for gait analysis utilizing Vertical Ground Reaction Force (VGRF) signals have demonstrated significant potential in Parkinson's Disease (PD) diagnosis and rehabilitation fields. However, existing GNN-based methods in this area normally model the VGRF signals as static topological structures, and ignore the dynamic variations of the VGRF graph structures during walking. To address this issue, a Global-Local Dynamic Directed Graph Neural Network (GLD<sup>2</sup>-GNN) is proposed to represent the dynamic spatio-temporal features of VGRF signals. The core component of the proposed model is the DyDGNN block, which is composed of a Dynamic Graph Learning (DGL) unit, a Dynamic Directed Graph Network (DyDGN) unit, and a Temporal Convolutional Network (TCN) unit. First, the DGL unit is proposed to learn dynamic topological relationships of VGRF signals. Based on learned graph structures, the DyDGN unit is constructed to extract the spatial patterns and capture topological dynamic features from VGRF signals. Subsequently, local temporal patterns of VGRF signals are extracted by the TCN unit. The proposed method is evaluated through k-fold cross-validation and cross-dataset validation on three datasets Ga, Ju and Si. Compared with existing methods, such as RFdGAD, Transformer, and AST-DGNN, GLD<sup>2</sup>-GNN demonstrates superior performance in the validation experiments. Notably, our method achieves an average improvement of 4.45% in accuracy, 2.93% in F1 score, and 2.88% in geometric mean across cross-dataset validation. Extensive experiments have demonstrated that GLD<sup>2</sup>-GNN exhibits both the representational ability for complex gait patterns and the generalization ability across various datasets by capturing dynamics of VGRF topological structures and spatio-temporal features from VGRF signals. For future work, we plan to combine our method with multi-modal methods and integrate our framework into a complete gait analysis system.</p>\",\"PeriodicalId\":13419,\"journal\":{\"name\":\"IEEE Transactions on Neural Systems and Rehabilitation Engineering\",\"volume\":\"PP \",\"pages\":\"\"},\"PeriodicalIF\":5.2000,\"publicationDate\":\"2025-09-25\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"IEEE Transactions on Neural Systems and Rehabilitation Engineering\",\"FirstCategoryId\":\"5\",\"ListUrlMain\":\"https://doi.org/10.1109/TNSRE.2025.3614430\",\"RegionNum\":2,\"RegionCategory\":\"医学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"ENGINEERING, BIOMEDICAL\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Transactions on Neural Systems and Rehabilitation Engineering","FirstCategoryId":"5","ListUrlMain":"https://doi.org/10.1109/TNSRE.2025.3614430","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ENGINEERING, BIOMEDICAL","Score":null,"Total":0}
A Global-Local Dynamic Directed Graph Neural Network for Parkinson's Disease Detection.
Graph Neural Networks (GNNs) for gait analysis utilizing Vertical Ground Reaction Force (VGRF) signals have demonstrated significant potential in Parkinson's Disease (PD) diagnosis and rehabilitation fields. However, existing GNN-based methods in this area normally model the VGRF signals as static topological structures, and ignore the dynamic variations of the VGRF graph structures during walking. To address this issue, a Global-Local Dynamic Directed Graph Neural Network (GLD2-GNN) is proposed to represent the dynamic spatio-temporal features of VGRF signals. The core component of the proposed model is the DyDGNN block, which is composed of a Dynamic Graph Learning (DGL) unit, a Dynamic Directed Graph Network (DyDGN) unit, and a Temporal Convolutional Network (TCN) unit. First, the DGL unit is proposed to learn dynamic topological relationships of VGRF signals. Based on learned graph structures, the DyDGN unit is constructed to extract the spatial patterns and capture topological dynamic features from VGRF signals. Subsequently, local temporal patterns of VGRF signals are extracted by the TCN unit. The proposed method is evaluated through k-fold cross-validation and cross-dataset validation on three datasets Ga, Ju and Si. Compared with existing methods, such as RFdGAD, Transformer, and AST-DGNN, GLD2-GNN demonstrates superior performance in the validation experiments. Notably, our method achieves an average improvement of 4.45% in accuracy, 2.93% in F1 score, and 2.88% in geometric mean across cross-dataset validation. Extensive experiments have demonstrated that GLD2-GNN exhibits both the representational ability for complex gait patterns and the generalization ability across various datasets by capturing dynamics of VGRF topological structures and spatio-temporal features from VGRF signals. For future work, we plan to combine our method with multi-modal methods and integrate our framework into a complete gait analysis system.
期刊介绍:
Rehabilitative and neural aspects of biomedical engineering, including functional electrical stimulation, acoustic dynamics, human performance measurement and analysis, nerve stimulation, electromyography, motor control and stimulation; and hardware and software applications for rehabilitation engineering and assistive devices.