A Riemannian Hypergraph Spectral-Temporal Network with Directed Effective Connectivity for Parkinson’s Disease Recognition from Resting-State EEG

Abstract

The electrophysiological changes that accompany Parkinson’s disease are expressed not only in the power of individual rhythms but also in the direction and geometry of the coupling between cortical regions, information that undirected and Euclidean descriptors only partially retain. This work introduces a Riemannian Hypergraph Spectral-Temporal Network, termed RHST-Net, that combines three complementary views of resting-state electroencephalography. Directed effective connectivity is first estimated for five rhythmic bands with phase transfer entropy, which captures the asymmetric flow of information between electrodes. In parallel, band-wise spatial covariance matrices are treated as points on the manifold of symmetric positive definite matrices and are projected to a Euclidean tangent space through a log-Euclidean mapping, producing geometry-aware descriptors. The electrode array is then modelled as a hypergraph whose hyperedges couple functionally related sensor groups, and a Chebyshev spectral convolution propagates information over this higher-order structure. A temporal convolutional network with dilated causal filters models the non-stationary evolution of connectivity, and a squeeze-and-excitation module recalibrates the five band streams before classification. On the public University of California San Diego resting-state dataset the network reaches 97.85 percent accuracy, a 97.80 percent F1-score and an area under the curve of 0.998 under a stratified protocol, and it preserves 90.36 percent accuracy under a subject-independent protocol. Ablation and interpretability analyses confirm that directed beta-band coupling over sensorimotor cortex drives the decision..