Abstract
BACKGROUND: Alzheimer's disease (AD) requires early intervention at preclinical stages like subjective memory complaints (SMC). Traditional static brain network analyses lack sensitivity to detect early functional disruptions in SMC. This study aimed to improve preclinical AD stratification by integrating dynamic gray-white matter functional connectivity (DFC) and machine learning. METHODS: Using multi-cohort ADNI data [N = 1,415 participants across cognitive normal[CN], SMC, and cognitive impairment [CI]groups],dynamic functional networks were constructed via sliding-window analysis (20-50 TR windows, 98% overlap) of 200 gray matter (Schaefer atlas) and 128 data-driven white matter nodes. DFC metrics (standard deviation of Fisher z-transformed correlations) were used to identify group differences and classify AD spectrum stages. Support vector machine (SVM) models were trained to differentiate CN/SMC/CI, with subgroup analyses in Aβ + and APOE E4 + populations. RESULTS: DFC with short sliding windows (20-50 TRs, 98% overlap) demonstrated greater sensitivity than SFC in detecting early functional disruptions in gray-white matter networks, identifying 34 CN-SMC [p < 0.05, e.g., ventral attention network (VAN)-white matter 2 (WM2) via Gau20-DFC], 44 CN-CI (p < 0.001), and 49 SMC-CI (p < 0.01) differential connections. Key early abnormalities were identified in the anterior cingulate network (WM4) and sensorimotor network (WM5), with WM4-WM5 disconnections in Aβ + subgroups strongly correlated with Aβ deposition and APOE ε4 genotype. Dynamic graph theory models using SVM achieved superior AD spectrum classification (ADNI2/3 AUCs: 0.85-0.92 vs. static 0.77-0.87), particularly in Aβ + subgroups (ΔAUC = 0.15 for SMC+/CI + discrimination, p < 0.001), with the VAN-WM2 feature in short-window DFC strongly correlating with cognitive scales (MMSE: r = 0.40, p < 10(-11); CDR-SB: r = -0.41, p < 10(-12)). Window function type (e.g., Gau20 for early changes, Ham50 for late stability) and data sampling points influenced sensitivity, with short windows optimizing early detection and long windows capturing late-stage network degeneration. These findings establish dynamic gray-white matter connectivity, particularly WM4-WM5 disruptions and VAN-WM2/DMN-WM8 features, as sensitive preclinical AD biomarkers enabled by machine learning for early SMC stratification. CONCLUSION: This study confirms that dynamic gray-white matter connectivity serves as a sensitive biomarker for preclinical Alzheimer's disease. The WM4-WM5 disruption hub and machine learning framework provide effective tools for early stratification of SMC, facilitating timely intervention within the disease's therapeutic window.