Multiscale multimodal graph convolutional networks for identifying essential tremor and dystonic tremor
Hongyu Wang1, Li Tao1, Xiaole Zhao1
1Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China.
Neurobiology of Disease
|May 1, 2026
Summary
A new multimodal graph convolutional network (MM-GCN) accurately distinguishes Essential Tremor (ET) and Dystonic Tremor (DT) using brain imaging. This approach reveals key brain circuit involvement in these overlapping movement disorders.
Area of Science:
- Neuroimaging
- Machine Learning
- Neurology
Background:
- Essential Tremor (ET) and Dystonic Tremor (DT) present similar motor symptoms, often leading to misdiagnosis.
- Advanced neuroimaging techniques and machine learning models are crucial for accurate diagnosis and understanding tremor pathogenesis.
Purpose of the Study:
- Develop a multiscale multimodal Graph Convolutional Network (MM-GCN) integrating structural and functional brain connectivity.
- Identify novel neuroimaging biomarkers for differentiating ET and DT.
Main Methods:
- Collected resting-state fMRI, DTI, and sMRI data from subjects.
- Constructed eight inter-regional similarity matrices and input them into a GCN with multimodal attention fusion.
- Performed binary classification tasks (ET vs. HC, DT vs. HC, ET vs. DT) and used Grad-CAM for interpretability.
Main Results:
- MM-GCN models achieved high classification accuracies: 95.24% (DT vs. HC), 85.45% (ET vs. HC), and 97.27% (ET vs. DT).
- Identified the thalamus, basal ganglia, and cerebellar networks as key discriminative regions.
- Nodal efficiency in these regions correlated significantly with clinical characteristics.
Conclusions:
- The MM-GCN model shows significant diagnostic potential for distinguishing ET and DT.
- Findings highlight the involvement of cerebello-thalamo-cortical circuits in both ET and DT, offering new insights into their pathophysiology.
