Attention-Based Framework Achieves 88.95% Accuracy in Alzheimer's Detection via rs-fMRI

A team led by Harshiddhi Pathak and Gowtham Reddy N from IIT Kharagpur has introduced a novel attention-based deep learning framework for Alzheimer's disease (AD) classification using resting-state functional MRI (rs-fMRI). The model bypasses traditional handcrafted connectivity features by directly operating on functional connectivity matrices, treating each brain region as a token and employing a Transformer-inspired self-attention mechanism. This allows it to capture long-range, global functional dependencies across distributed brain networks—a critical improvement over conventional machine learning approaches that struggle with the high dimensionality and noise inherent in rs-fMRI data.

Evaluated on a longitudinal cohort from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the framework achieved 88.95% accuracy and a 0.90 ROC-AUC for binary AD versus cognitively normal classification. The team used subject-wise evaluation to prevent data leakage across visits and incorporated class-weighted optimization to handle mild class imbalance. These results demonstrate that self-attention-driven functional connectivity modeling can serve as a robust, interpretable tool for early Alzheimer's detection, potentially enabling non-invasive screening from routine resting-state scans. The paper has been accepted at the 48th IEEE EMBC 2026.