End-To-End Hyperbolic Graph Neural Networks for Brain Age Prediction with MEG Data

Alzheimer's disease (AD) is a neurodegenerative disease that affects a large population. This disease drives a devastating rapid decline of brain functionality in the human brain. Identifying AD early is essential to effective treatment. Magneticoencephalography (MEG) is an effective tool for identifying changes in normal brain aging trajectories. In this study, we created an end-to-end Hyperbolic Graph Convolutional Neural Network (HGCN) that is able to learn the subtle hierarchical representations from training data to achieve high accuracy on unseen data. We found that the HGCN model outperformed the regular GCN model and other classic classification models in all classification metrics, and in both binary and multiclass classification tasks. We also found that our MEG data were best represented with a relatively large negative hyperbolic curvature. In our embedding visualizations, we motivated an intuition explaining why hyperbolic space is able to better fit hierarchical data. This model can be used for both future research and clinical applications to help predict and prevent neurogenerative disorders before they progress. ¹¹The source code is publicly available at https://github.com/astarryknight/hgcn_e2e