Frequency-Aware Interpretable Deep Learning Framework for Alzheimer's Disease Classification Using rs-fMRI

Gaining insight into the spectral and temporal alterations in brain connectivity associated with Alzheimer's disease (AD) may offer pathways toward more informative biomarkers and a deeper understanding of disease mechanisms. We propose FINE (Frequency-aware Interpretable Neural Encoder), a novel deep learning model designed to capture multi-scale temporal and frequency-specific patterns in dynamic functional network connectivity (dFNC) derived from resting-state fMRI. FINE integrates multiple expert branches, including convolutional layers, learnable wavelet layers, transformers, and static encoders, enabling the joint modeling of temporal evolution and spectral content of brain networks in an end-to-end framework. Beyond classification, FINE supports frequency-wise interpretability by aligning gradient-based saliency maps with statistical group differences, revealing potential robust, biologically meaningful biomarkers of AD. Evaluated on the large OASIS-3 dataset (856 subjects), FINE achieves AD classification performance (ROC-AUC 0.769) and provides insights into frequency-specific connectivity disruptions, particularly within subcortical, sensorimotor, and cerebellar networks. Our results demonstrate that incorporating frequency-aware modeling and interpretable architectures can advance both disease classification and underlying functional disruption of AD-related brain dynamics.