Objectives: Alzheimer's disease (AD) begins with subtle symptoms, making early detection of mild cognitive impairment (MCI) crucial for timely intervention. Current deep learning models for brain imaging often process different modalities separately, failing to account for demographic variations and the need for integrated analysis of structural and functional data from the same brain regions. This approach can result in reduced sensitivity to early cognitive decline due to oversmoothing effects in Graph Convolutional Networks (GCNs). Methods: This study proposes a hybrid multi-channel transformer graph attention network (HMT-GAT) for MCI identification. First, a structurally constrained fused brain network is constructed by incorporating DTI-derived anatomical information into rs-fMRI-based functional connectivity estimation. A locally weighted clustering coefficient (LWCC) is then used to extract multi-scale local topological features from the fused network. Demographic and acquisition-related variables, including acquisition site, age, and sex, are further integrated into a sparsely connected population graph to model inter-subject relationships. Results: Under the same evaluation protocol, HMT-GAT achieved competitive and generally superior performance compared with implemented baseline models. For NC vs. EMCI classification, HMT-GAT obtained an ACC of 87.97 %, SEN of 80.46%, and SPE of 91.45%. For NC vs. LMCI classification, it achieved an ACC of 87.63%, SEN of 95.13%, and SPE of 94.46%, indicating balanced classification performance for MCI-related identification tasks. Discussion:Interpretability analysis identified disease-related regions, including the inferior temporal gyrus and amygdala, suggesting that HMT-GAT may provide biologically meaningful evidence for MCI-related brain network alterations within the AD continuum.

alzheimer's disease, brain imaging patterns, graph convolutional network, LWCCl, HMT-GAT