BGCN: An EEG-based Graphical Classification Method for Parkinson's Disease Diagnosis with Heuristic Functional Connectivity Speculation

As the population ages, the prevalence of Parkinson's Disease (PD), a neurodegeneration disorder that deeply hinders one's daily intellectual and physical activities, has increased rapidly over the past years. However, finding an effective modifiable treatment for PD remains stagnant to date, elevating the significance of the accurate diagnosis. Meanwhile, studies on functional connectivity could provide insights into the neurophysiological mechanisms underlying PD. Hence, this study intends to provide a unified framework, Brain Graph Convolutional Networks (BGCN), incorporating the non-Euclidean heuristic-based brain functional connectivity into a graph-based deep learning model (GCN) for PD diagnosis. The graph representation of Electroencephalography (EEG) data priors in retaining the spatial interdependence among the EEG channels and facilitating the formulation of the functional connectivity construction problem. With the GCN, we modeled neural information exchange with convolutions between nodes along functional connectivity. In this work, functional connectivity was attained by solving a Minimum Spanning Tree (MST) problem with a heuristic search algorithm. As a result, the obtained functional connectivity corresponded to existing MRI studies in terms of the affected regions and hub shift. To evaluate the efficacy of the proposed framework, we compared the heuristic functional connectivity speculation with random/uniform connectivity generated by K-nearest neighbors(k-NN). The proposed framework has achieved excellent precision (95.59%) and learning robustness.