Abstract
The progression of Alzheimer's disease (AD) involves complex changes in brain structure and function that are driven by their interaction, making structure-function coupling (SFC) a valuable indicator for early detection of AD. Static SFC refers to the overall structure-function interaction, whereas dynamic SFC refers to transient coupling variations. In this study, we aimed to assess the potential of combining static and dynamic SFC with machine learning (ML) for the early detection of AD. We analyzed a discovery cohort and an external validation cohort, including AD, mild cognitive impairment (MCI), and healthy control (HC) groups. Then, we quantified differences between static SFC and dynamic SFC at different stages of AD progression. Feature selection was performed using ElasticNet. A Gaussian naive Bayes (GNB) classifier was used to test the ability of SFC to classify AD stages. We also analyzed the correlations between SFC features and early AD physiological biomarkers. Static SFC increased with AD progression, whereas dynamic SFC showed greater variability and decreased stability. Using SFC features selected by ElasticNet, the GNB classifier achieved high performance in differentiating between the HC and MCI stages (area under the curve [AUC] = 91.1%) and between the MCI and AD stages (AUC = 89.03%). Significant correlations were found between SFC features and physiological biomarkers. The combined use of SFC features and ML has strong potential value for the accurate classification of AD stages and significant potential value for the early detection of AD. This study demonstrates that combining static and dynamic SFC with ML provides a novel perspective for understanding the mechanisms of AD and contributes to improving its early detection.