Multi-level functional network-based PD identification via graph deep learning

Background Parkinson’s disease (PD) is a neurodegenerative disease characterized by degenerative changes in nigrostriatal dopaminergic neurons and Lewy body morphology. Functional magnetic resonance imaging (fMRI) has become an important tool for identifying biomarkers of PD by virtue of its sensitivity in detecting differences in functional connectivity (FC) of the brain. Most of the current FC-based PD diagnostic methods only consider the connectivity topology between brain regions, ignoring the differences and complementarities of FCs between patients, which have been proven to be critical in identifying PD. Methods In this article, a patient similarity network is first constructed to mine the complementarity of FC between patients, and a multi-level functional network structure is constructed, which consists of FCs between brain regions as well as the patient similarity network. Then, a graph convolutional network (GCN) model is established to extract the complex structural information of the multi-level network. Meanwhile, to avoid the overfitting problem that may be caused by the small sample of fMRI, the Laplacian regularization term is enforced in the GCN model. Results The results of the study show that the multi-level functional network structure based classification model performs well in PD identification with high levels of accuracy, precision, and recall of 76.2%, 72.2% and 75.3%, respectively. In addition, the role of different brain networks in the categorization task was deeply analyzed by the occlusion sensitivity analysis method, and it was found that the frontal lobe has an important role in recognizing PD. The work verified the significance of complementarity of FC between patients in identifying PD.