A multi-view graph neural network framework for Parkinson's disease identification based on dynamic functional connectivity.

Parkinson's disease (PD) is a common neurodegenerative disorder, and accurate diagnosis is crucial for timely intervention. Dynamic functional connectivity (DFC) analysis of resting-state functional magnetic resonance imaging (rs-fMRI) can capture the time-varying characteristics of brain networks, offering a new perspective for PD identification. However, traditional clustering-based DFC methods suffer from issues such as the loss of continuous temporal information. Existing machine learning approaches struggle to effectively integrate the temporal dynamics of DFC with inter-subject differences in functional connectivity patterns and are further limited by the poor generalizability and opaque decision-making processes inherent to the small sample sizes typical of neuroimaging data. Here, we propose a DFC analysis framework based on a multi-view graph convolutional network (GCN). This method independently constructs inter-subject similarity networks for each sliding time window, forming multi-view graph-structured inputs. It employs a shared-weight GCN to extract node embeddings from each window, which are then fused for classification. Furthermore, the gradient-weighted class activation mapping (Grad-CAM) algorithm is incorporated to provide visual interpretability of the model's decisions. Our method outperforms traditional static functional connectivity and clustering-based approaches in classification accuracy. Concurrently, it achieves superior classification performance compared to other benchmark models. Grad-CAM-based interpretability analysis further reveals that the frontal and parietal lobes contribute most significantly to the classification decisions, providing computational evidence for understanding the brain network mechanisms underlying PD.