Adaptive hypergraph learning reveals high-order functional network alterations in mild cognitive impairment.

Mild cognitive impairment (MCI) is an important prodromal stage of Alzheimer's disease, and its early identification is critical for risk assessment and timely intervention. Resting-state functional magnetic resonance imaging (rs-fMRI) can noninvasively characterize brain functional activity and connectivity. However, most existing methods rely on static second-order functional connectivity, limiting their ability to capture dynamic coordination and high-order interactions among multiple brain regions. To address these limitations, we propose a brain functional network analysis framework integrating multi-scale temporal feature fusion with adaptive high-order hypergraph convolution. Specifically, the multi-scale temporal fusion module captures functional dynamics at different temporal resolutions, while the adaptive hypergraph convolutional network learns high-order associations among brain regions in an end-to-end manner. Spatiotemporal feature encoding is further incorporated to facilitate automatic MCI identification. The proposed framework was evaluated using rs-fMRI data from the ADNI dataset, with regional BOLD time series extracted based on the AAL atlas. Under five-fold cross-validation, the proposed method achieved an accuracy of 82.66%, a sensitivity of 82.07%, and a specificity of 83.07% on ADNI-2, while maintaining an accuracy of 81.50% on ADNI-3, outperforming existing methods on both datasets. Further analysis identified abnormal high-order coordination patterns involving the default mode network, limbic system, and hippocampus-related memory circuits in patients with MCI. These findings provide methodological support for the early identification of MCI and the discovery of potential neuroimaging biomarkers.