Identification of essential tremor and dystonic tremor using Graph Convolutional Networks with multiple connectivity patterns.

INTRODUCTION

As a deep learning algorithm, Graph convolutional network (GCN) can efficiently process graph-structured data to identify salient brain regions and brain connectivity patterns. We combine GCNs with a multi-connection pattern (MCGCN) to identify salient brain regions implicated in Essential Tremor (ET) and Dystonic Tremor (DT), aiming to explore the underlying neuropathological mechanisms of these conditions.

METHODS

Rs-fMRI data were collected from 55 ET patients, 51 DT patients, and 52 healthy controls (HCs). BOLD time series from each subject were extracted and functional connectivity (FC) matrices were constructed using three distinct connectivity modes. These matrices were then input to four GCN architectures for binary classification tasks (ET vs. HCs, DT vs. HCs, ET vs. DT). We utilized Grad-CAM to identify the more discriminative brain regions, and graph theory and correlation analyses were employed to validate the behavioral relevance of the discriminative regions identified by MCGCN, confirming the salient brain regions for ET and DT.

RESULTS

All GCN models demonstrated strong classification performance, with the highest mean accuracies of 91.36 % for DT vs. HCs, 85.91 % for ET vs. HCs, and 86.64 % for ET vs. DT. Discriminative brain regions were mainly localized in the basal ganglia, cerebello-thalamo-cortical motor circuitry, and non-motor cortical regions. Correlation analysis revealed that the nodal efficiency of the four salient brain regions was negatively correlated with clinical characteristics.

CONCLUSION

Our findings suggest the critical role of the classic tremor network in ET and DT pathogenesis, enhancing our comprehension of their FC-based pathophysiological mechanisms.