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

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Parkinson's disease (PD) diagnosis relies on accurate identification for timely intervention.
• Resting-state functional magnetic resonance imaging (rs-fMRI) dynamic functional connectivity (DFC) offers insights into brain network changes.
• Traditional DFC methods and machine learning models face challenges with temporal information, inter-subject variability, generalizability, and interpretability.

Purpose of the Study:

• To develop an advanced DFC analysis framework for improved PD identification.
• To leverage multi-view graph convolutional networks (GCNs) for integrating temporal dynamics and inter-subject differences.
• To enhance model interpretability using Grad-CAM for understanding PD-related brain mechanisms.

Main Methods:

• A multi-view GCN framework was proposed, constructing inter-subject similarity networks for each time window.
• Shared-weight GCNs extracted node embeddings from each window, followed by fusion for classification.
• Gradient-weighted class activation mapping (Grad-CAM) was integrated for model interpretability.

Main Results:

• The proposed DFC-GCN framework significantly outperformed traditional static functional connectivity and clustering-based DFC methods in classification accuracy.
• The model achieved superior classification performance compared to other benchmark models.
• Grad-CAM analysis identified the frontal and parietal lobes as key regions contributing to PD classification.

Conclusions:

• The multi-view GCN framework offers a robust and interpretable approach for PD diagnosis using rs-fMRI DFC.
• This method effectively captures dynamic brain network alterations in PD, outperforming existing techniques.
• The findings provide computational evidence supporting the role of frontal and parietal lobe network dynamics in PD pathogenesis.