Resting-state EEG offers a low-cost, non-invasive biomarker for Alzheimer's and Parkinson's diseases, yet most deep learning models fail to generalize to new patients due to subject-dependent evaluation protocols, isolated frequency-band processing, and the neglect of cross-frequency interactions. We propose CFGT-Net, a Cross-Frequency Graph-Transformer Network. For each EEG epoch, five canonical bands are processed by a graph attention encoder on a learnable phase-lag index adjacency to produce spatial embeddings. A cross-frequency coupling (CFC) attention models band interactions, a temporal transformer tracks their evolution across epochs, and a correlation-alignment loss enforces subject... More >