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

Accurate diagnosis of Alzheimer's disease (AD) remains a significant clinical challenge. While resting-state functional MRI (rs-fMRI) provides valuable insights into brain network disruptions, traditional machine learning methods often rely on handcrafted features that fail to capture the complex, high-dimensional, and globally distributed nature of these functional changes. This paper investigates whether a Transformer-inspired self-attention mechanism can better model these inter-regional dependencies to improve AD classification.

The researchers propose an attention-based deep learning framework that operates directly on functional connectivity (FC) matrices. By treating individual brain regions (ROIs) as tokens, the model uses a self-attention mechanism to learn how different regions interact across the entire brain. This approach eliminates the need for manual feature engineering. The model was trained and evaluated using longitudinal data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort, consisting of 60 subjects (30 AD and 30 cognitively normal). To ensure the model's reliability, the authors implemented a subject-wise data splitting protocol to prevent information leakage between different visits of the same patient and used class-weighted optimization to handle potential imbalances.

The proposed model achieved an accuracy of 88.95% and a ROC-AUC of 0.90 in distinguishing between AD patients and cognitively normal controls. The results demonstrate that self-attention is effective at capturing the long-range functional interactions that are characteristic of neurodegeneration. The authors highlight that this framework provides a more robust and interpretable alternative to traditional convolutional neural networks, which are often constrained by local receptive fields that may miss global connectivity patterns.