Title: Harnessing Multi-Scale Hypergraphs for High-Order Brain Connectivity Analysis

Abstract:

Accurately characterizing the intricate interplay among various brain regions is essential for the early identification of neurodegenerative conditions, including Alzheimer’s Disease (AD) and Parkinson’s Disease (PD). Although graph-based models are commonly employed to investigate brain networks, they predominantly concentrate on pairwise interactions between directly linked nodes. This focus inherently restricts their ability to capture higher-order dependencies that span multiple regions. While hypergraph-based methods have emerged to address higher-order relationships, numerous existing techniques depend on pre-defined hyperedges or limit learning to hyperedge weights. Such constraints reduce flexibility and hinder the capacity to detect multi-resolution structural patterns. To address these limitations, we propose MuHL, an adaptive multi-scale hyperedge learning framework. MuHL builds hierarchical node features and dynamically acquires high-order interactions by continuously constructing hyperedges across multi-resolution graph signals. Comprehensive experiments conducted on various brain network benchmarks reveal that MuHL consistently enhances disease classification accuracy across different stages. Furthermore, the framework successfully identifies key regions of interest (ROIs) and their group-wise interactions derived from learned hyperedges linked to disease progression, underscoring its potential as a robust tool for analyzing brain networks in neurodegenerative disorders.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC