An Attention-Based Framework for Alzheimer’s Disease Classification Using Resting-State fMRI

This paper presents an attention-based deep learning framework for Alzheimer’s disease classification from resting-state fMRI functional connectivity matrices. Brain regions are treated as tokens, and a Transformer-inspired self-attention mechanism models long-range dependencies across distributed functional networks. Using a longitudinal Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort, the study applies subject-wise evaluation to prevent information leakage between visits and uses class-weighted optimization for mild class imbalance. For binary Alzheimer’s disease versus cognitively normal classification, the reported accuracy is 88.95% and ROC-AUC is 0.90, with a favorable precision-recall balance.