Introduction: Alzheimer's disease (AD) heterogeneity complicates diagnosis and prognosis. Uncovering amyloid-tau-neurodegeneration (A-T-N) patterns may improve diagnostic prediction.
Methods: We applied SuperBigFLICA (SBF), a semi-supervised multimodal fusion method, to gray matter density, cortical thickness (CT), pial surface area, amyloid and tau positron emission tomography maps from 274 Alzheimer's Disease Neuroimaging Initiative 3 participants to derive 50 latent components predictive of cognitive decline. Subject loadings were then used to predict diagnosis (cognitively normal, mild cognitive impairment, dementia) and apolipoprotein E (APOE) ε4 status via least absolute shrinkage and selection operator logistic regression, compared to demographic, single-modality, and naïve fusion comparator models.
Results: SBF modestly predicted out-of-sample concurrent clinical severity (Clinical Dementia Rating Sum of Boxes; r = 0.21), yet models using SBF-derived loadings were among the strongest comparator models (area under the receiver operating characteristic curve; = 0.80 for diagnosis; 0.83 for APOE ε4). Amyloid alterations in sensory areas best separated dementia, while a tri-modal tau-neurodegeneration pattern related to disease progression. Loadings were validated through cerebrospinal fluid correlations.
Discussion: SBF improves prediction and reveals interpretable patterns that better classify clinical diagnoses and APOE ε4 than traditional approaches.
Keywords: Alzheimer's disease; amyloid–tau–neurodegeneration framework; latent components; multimodal data fusion; multimodal neuroimaging; semi‐supervised learning.
© 2026 The Author(s). Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring published by Wiley Periodicals LLC on behalf of Alzheimer's Association.