Investigating the Amyloid-Tau-Neurodegeneration Framework in Alzheimer's Disease Using Semi-Supervised Multimodal Imaging Data Fusion.

INTRODUCTION: Alzheimer's disease (AD) is heterogeneous, complicating diagnosis and prognosis. Uncovering patterns that link abnormalities across amyloid-tau-neurodegeneration (A-T-N) framework may improve prediction of clinical diagnosis. METHODS: We applied SuperBigFLICA (SBF), a semi-supervised multimodal data fusion method, to maps of gray matter density, cortical thickness, pial surface area, amyloid PET, and tau PET in 274 ADNI-3 participants. The model was trained to derive 50 latent components most predictive of a continuous measure of cognitive decline. Latent components' subject loadings were subsequently used to predict diagnosis (cognitively normal, mild cognitive impairment, dementia) and APOE4 status using LASSO logistic regression, and were compared against demographic, single-modality, and naïve fusion baselines. RESULTS: While SuperBigFLICA modestly predicted cognitive decline (r = 0.21), SBF loadings-based models outperformed baselines (AUROC = 0.80 for diagnosis; 0.83 for APOE4). Amyloid alterations in sensory areas along the sensory-association axis best separated dementia, while a multimodal A-T-N pattern was related to early cognitive decline. Subject loadings on these two patterns were associated with cerebrospinal fluid (CSF) biomarkers, highlighting how CSF AD biomarkers relate to spatial patterns of brain A-T-N burden. DISCUSSION: Semi-supervised multimodal fusion improves prediction and reveals interpretable imaging patterns that predict APOE4 and clinical diagnoses better than traditional approaches.