A graph deep learning method for diagnosis of Parkinson's disease using brain functional connectivity features

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Early Parkinson's disease (PD) identification is critical for effective treatment.
• Resting-state functional magnetic resonance imaging (rs-fMRI) reveals brain functional connectivity (FC) differences in PD.
• Current FC analysis methods often overlook dynamic fluctuations and lack interpretability, especially with small datasets.

Purpose of the Study:

• To develop an interpretable Graph Convolutional Network (GCN) framework for PD identification.
• To integrate static and dynamic FC information for improved classification accuracy.
• To enhance model generalization and interpretability in PD detection.

Main Methods:

• Utilized rs-fMRI data to compute both static and dynamic functional connectivity (FC).
• Developed an interpretable GCN framework incorporating inter-subject similarity graphs.
• Integrated interpretability analysis techniques to identify critical brain regions for classification.

Main Results:

• The proposed GCN model demonstrated superior performance in classifying Parkinson's disease.
• The model exhibited good generalization capabilities, even with limited sample sizes.
• Interpretability analysis successfully identified key brain regions crucial for PD discrimination.

Conclusions:

• The developed computational framework offers an effective approach for PD identification using rs-fMRI.
• The study provides valuable insights into the pathological mechanisms of Parkinson's disease through interpretable AI.
• This work highlights the potential of integrating dynamic FC and GCNs for neurological disorder research.