An Edge Feature Inclusive Variational Graph Autoencoder for Pet-Driven Alzheimer's Diagnosis

Early and accurate diagnosis of Alzheimer's disease (AD) is critical for effective intervention. We propose GINEVGAE(Modified Graph Isomorphism Network with Variational Graph Autoencoder, a novel variational graph autoencoder that leverages GINEConv(Modified Graph Isomorphism Network operator) layers to integrate both node features and continuous edge weights derived from 18F-FDG PET(Fludeoxyglucose F18 positron emission tomography) scans. We construct subject-specific graphs using anatomical ROIs(Region of interest) and NBS-pruned(Network Based Statistic) connectivity, learn latent embeddings via GINE-VGAE, and classify with a support vector machine. On ADNI PET data, GINE-VGAE + SVM(Support Vector Machine) achieves 93.8% accuracy and 0.937 F1-score, outperforming ROI-based, voxel based baselines and other graph embedding based methods.