A Novel Redundancy-aware Spatio-temporally updated Hypergraph-Based Brain Network for Early Diagnosis of Alzheimer’s Disease

Early detection, particularly during the mild cognitive impairment (MCI) stages, provides an important opportunity for intervention.Recently, deep learning models have leveraged these Region of Interest (ROIs) from functional MRI (fMRI) to construct and represent the Brain network.However, recent deep learning models primarily capture local brain regions, limiting their ability to represent distant and global features.Constructing effective brain connectivity graphs that balance spatial correlations, and temporal information remains a challenge for Alzheimer's Disease (AD) diagnosis.To address these limitations, we propose the Hierarchical-Spatio-Temporal Graph Convolutional Network (Hi-STGCN), a novel deep learning framework designed to enhance AD classification.Hi-STGCN follows a two-tier hierarchical structure: first, it constructs Masked Spatio-Temporally Updated Brain networks (MSTUB) to refine features at the subject level by leveraging region-wise spatiotemporal dependencies.Then, a Population Network integrates these representations across multiple subjects, capturing global connectivity patterns that improve classification robustness.With classification accuracies of 95.95% (AD vs. CN), 93.65% (AD vs. MCI), and 92.16% (MCI vs. CN), Hi-STGCN outperforms existing models, demonstrating its potential in advancing early diagnosis and personalized treatment strategies for AD.